Search

Moving Object Detection (MOD) is a critical vision task for successfully achieving safe autonomous driving. Despite plausible results of deep learning methods, most existing approaches are only frame-based and may fail to reach reasonable performance when dealing with dynamic traffic participants. Recent advances in sensor technologies, especially the Event camera, can naturally complement the conventional camera approach to better model moving objects. However, event-based works often adopt a pre-defined time window for event representation, and simply integrate it to estimate image intensities from events, neglecting much of the rich temporal information from the available asynchronous events. Therefore, from a new perspective, we propose RENet, a novel RGB-Event fusion Network, that jointly exploits the two complementary modalities to achieve more robust MOD under challenging scenarios for autonomous driving. Specifically, we first design a temporal multi-scale aggregation module to fully leverage event frames from both the RGB exposure time and larger intervals. Then we introduce a bi-directional fusion module to attentively calibrate and fuse multi-modal features. To evaluate the performance of our network, we carefully select and annotate a sub-MOD dataset from the commonly used DSEC dataset. Extensive experiments demonstrate that our proposed method performs significantly better than the state-of-the-art RGB-Event fusion alternatives. The source code and dataset are publicly available at: https://github.com/ZZY-Zhou/RENet., ICRA'23

International audience; This work numerically investigates the dynamics of a Chua’s circuit model experiencing a truncated sinusoidal force and driven by an external perturbed excitation. We mainly study the impact of the system’s nonlinearity on the occurrence of Vibrational Resonance (VR) and Ghost-Vibrational Resonance (GVR) phenomena. When a truncated sinusoidal nonlinearity is used, the system requires relatively smaller perturbation amplitude to attain its maximum response better than the one achieved with a sawtooth nonlinearity which requires a larger perturbation amplitude. Therefore, the system with a truncated sinusoidal nonlinearity outperforms the one with a sawtooth nonlinearity. Exciting the system with two low frequency inputs and an additive high frequency perturbation, we identify different ranges of the perturbation amplitude in which the occurrence of VR and GVR phenomena are maximized. We show that depending on the perturbation amplitude, the system can synchronize its response with the ghost frequency or one of the two input low frequencies.

This paper presents the design of a low-power and low-noise CMOS photo-transduction circuit. We propose to use the new technique of composite transistors for noise reduction of photoreceptor in the subthreshold by exploiting the small size effects of CMOS transistors. Several power and noise optimizations, design requirements, and performance limitations relating to the CMOS photoreceptor are presented. This new structure with composite transistors ensures low noise and low power consumption. The CMOS photoreceptor, implemented in a 130 nm standard CMOS technology with a 1.2 V supply voltage, achieves a noise floor of 2μV/⎷Hz within the frequency range from 1 Hz to 10 kHz. The current consumption of the CMOS photoreceptor is 541 nA. This paper shows the need for the design of phototransduction circuit at low voltage, low noise and how these constraints are reflected in the design of CMOS vision sensor.

International audience; Reflectance Transformation Imaging (RTI) is a technique for estimating the surface local angular reflectance and characterizing the visual properties by varying lighting directions and capturing a set of stereo-photometric images. The proposed method, namely HD-RTI, is based on the coupling of RTI and HDR imaging techniques. The HD-RTI automatically optimizes the necessary exposure times for each angle of illumination by using the response of the scene. Our method is applied to industrial surfaces with micro-scratches from which we will estimate saliency information. Results show that coupling HDR and RTI enhance the characterization and therefore the discrimination on the surfaces visual saliency maps. It leads to an increase in robustness for visual quality assessment tasks.

International audience; Texture characterization from the metrological point of view is addressed in order to establish a physically relevant and directly interpretable feature. In this regard, a generic formulation is proposed to simultaneously capture the spectral and spatial complexity in hyperspectral images. The feature, named relative spectral difference occurrence matrix (RSDOM) is thus constructed in a multireference, multidirectional, and multiscale context. As validation, its performance is assessed in three versatile tasks. In texture classification on HyTexiLa, content-based image retrieval (CBIR) on ICONES-HSI, and land cover classification on Salinas, RSDOM registers 98.5% accuracy, 80.3% precision (for the top 10 retrieved images), and 96.0% accuracy (after post-processing) respectively, outcompeting GLCM, Gabor filter, LBP, SVM, CCF, CNN, and GCN. Analysis shows the advantage of RSDOM in terms of feature size (a mere 126, 30, and 20 scalars using GMM in order of the three tasks) as well as metrological validity in texture representation regardless of the spectral range, resolution, and number of bands.

In this paper, we first propose a brief overview of nonlinear resonance applications in the context of image processing. Next, we introduce a threshold detector based on these resonance properties to investigate the perception of subthreshold noisy images. By considering a random perturbation, we revisit the well-known stochastic resonance (SR) detector whose best performances are achieved when the noise intensity is tuned to an optimal value. We then introduce a vibrational resonance detector by replacing the noisy perturbation with a spatial high-frequency signal. To enhance the image perception through this detector, it is shown that the noise level of the input images must be lower than the optimal noise value of the SR-based detector. Under these conditions, considering the same noise level for both detectors, we establish that the vibrational resonance (VR)-based detector significantly outperforms the SR-based detector in terms of image perception. Moreover, we show that whatever the perturbation amplitude, the best perception through the VR detector is ensured when the perturbation frequency exceeds the image size. This article is part of the theme issue ‘Vibrational and stochastic resonance in driven nonlinear systems (part 2)’.

International audience; In this paper, we propose an automated adaptive focus pipeline for creating synthetic extended depth of field images using a reflectance transformation imaging (RTI) system. The pipeline proposed detects object regions at different depth levels relative to the camera’s depth of field and collects a most focused image for each. These images are then run through a focus stacking algorithm to create an image where the focus of each pixel has been maximized for the given camera parameters, lighting conditions, and glare. As RTI is used for many cultural heritage imaging projects, automating this process provides high quality data by removing the need for many separate images focused on different regions of interest on the object. It also lowers the skill floor for this image collection process by reducing the amount of manual adjustments that need to be made for focus. Furthermore, this can help to minimize the amount of time that a sensitive cultural heritage object is outside of its ideal preservation environment.

Respiration is an important parameter in critical and pediatric care since its monitoring allows medical staff to detect many life-threatening diseases. One of the existing monitoring methods is based on remote photoplethysmography (rPPG). This technique consists of extracting a signal related to blood volume variations using a camera. This signal carries useful physiological information such as cardiac and respiratory rates. However, the quality of the signal is lower than regular contact-based methods and represents a major weakness of the rPPG method. In this paper, we propose an algorithm to explicitly maximize the respiratory signal quality by maximizing the Signal-to-Noise Ratio (SNR). Instead of using the regular Fast-Fourier-based energy ratio for the signal-to-noise-ratio estimation, we propose to use the continuous wavelet transform to deal with non-stationarities of the respiratory signal. The method, named Wavelet Variance Maximization (WVM), is based on the Generalized Eigen Value Decomposition (GEVD) algorithm and estimates the optimal combination of the temporal color traces to obtain a high-quality rPPG signal. Our method was tested on 12 healthy adult volunteers and the results confirm that the estimated signal has better quality than existing methods, with approximately 20% reduction of error compared to the best tested state-of-the-art method.

International audience; Reflectance Transformation Imaging (RTI) is a technique for estimating surface local angular reflectance from a set of stereo-photometric images captured with variable lighting directions. The digitization of this information fully fits into the industry 4.0 approach and makes it possible to characterize the visual properties of a surface. The proposed method, namely HD-RTI, is based on the coupling of RTI and HDR imaging techniques. This coupling is carried out adaptively according to the response at each angle of illumination. The proposed method is applied to five industrial samples which have high local variations of reflectivity because of their heterogeneity of geometric texture and/or material. Results show that coupling HDR and RTI improves the relighting quality compared to RTI, and makes the proposed approach particularly relevant for glossy and heterogeneous surfaces. Moreover, HD-RTI enhances significantly the characterization of the local angular reflectance, which leads to more discriminating visual saliency maps, and more generally to an increase in robustness for visual quality assessment tasks.

Interpreting human actions requires understanding the spatial and temporal context of the scenes. State-of-the-art action detectors based on Convolutional Neural Network (CNN) have demonstrated remarkable results by adopting two-stream or 3D CNN architectures. However, these methods typically operate in a non-real-time, ofline fashion due to system complexity to reason spatio-temporal information. Consequently, their high computational cost is not compliant with emerging real-world scenarios such as service robots or public surveillance where detection needs to take place at resource-limited edge devices. In this paper, we propose ACDnet, a compact action detection network targeting real-time edge computing which addresses both efficiency and accuracy. It intelligently exploits the temporal coherence between successive video frames to approximate their CNN features rather than naively extracting them. It also integrates memory feature aggregation from past video frames to enhance current detection stability, implicitly modeling long temporal cues over time. Experiments conducted on the public benchmark datasets UCF-24 and JHMDB-21 demonstrate that ACDnet, when integrated with the SSD detector, can robustly achieve detection well above real-time (75 FPS). At the same time, it retains reasonable accuracy (70.92 and 49.53 frame mAP) compared to other top-performing methods using far heavier configurations. Codes will be available at https://github.com/dginhac/ACDnet., Comment: Accepted for publication in Pattern Recognition Letters

International audience; A gradient-based texture feature for hyperspectral image is formulated with straightforward application to grayscale and color images. Processed in full band, GHOST is expressed as a four-dimensional probability density distribution encompassing joint metrological assessment of spectral and spatial properties. Its performance is close to Opponent Band Local Binary Pattern (OBLBP) in HyTexiLa texture classification (91 %-99 % accuracy) with feature size 0.2 % of OBLBP's.

The respiratory rate is important information in the healthcare environment. Consequently, research is done to develop a device that could measure the respiratory rate continuously with non-contact devices. Various methods were tried, such as radio-based, thermal imaging or remote photoplethysmography (rPPG). The rPPG method uses a video recording of the skin in ambient light conditions. It measures the small variations of light reflection induced by the amount of blood in vessels. This method allows the extraction of physiological parameters such as the heart rate or respiratory rate without any contact with the skin. The main issue with the rPPG technique is the lower signal quality compared with contact-based methods. In this paper, we assess the performance of the respiratory rate estimation algorithms with rPPG signals. The tested algorithms were designed for contact-PPG signals input. The use of the algorithms designed for contact PPG on remote PPG signals can lead to respiratory rate estimations with a mean absolute error below 3 breaths-per-minute. We benchmark our results using this standard and some other metrics to interpret the quality of the assessment.

International audience; Remote photoplethysmography (rPPG) is a recent technique for estimating heart rate by analyzing subtle skin color variations using regular cameras. As multiple noise sources can pollute the estimated signal, post-processing techniques, such as bandpass filtering, are generally used. However, it is often possible to see alterations in the filtered signal that have not been suppressed, although an experienced eye can easily identify them. From this observation, we propose in this work to use an LSTM network to filter the rPPG signal. The network is able to learn the characteristic shape of the rPPG signal and especially its temporal structure, which is not possible with the usual signal processingbased filtering methods. The results of this study, obtained on a public database, have demonstrated that the proposed deep-learning-based filtering method outperforms the regular post-processing ones in terms of signal quality and accuracy of heart rate estimation.

International audience; The properties of nonlinear systems have attracted a considerable interest these past years since they can be used to develop bio-inspired applications ranging from signal detection to image processing. Among these nonlinear signatures is Vibrational Resonance (V.R.) which expresses the ability of a nonlinear system to take benefit of a high frequency perturbation in order to enhance its response to a weak low frequency excitation . Since its introduction in mechanical systems and electronic devices, this effect has been widely reported in various systems and for different applications. Indeed, V.R can be used to detect subthreshold signals or to enhance the perception of subthreshold images. Among the systems which exhibits V.R., we can cite bio-inspired circuits, networks of coupled electronic neurons, as well as inhomogeneous and space-dependent nonlinear damped systems.These last mechanical systems constitute the subject of our communication which adresses more specifically the conditions of existence of V.R. The system under numerical investigation models a particle of mass m submitted to a multistable potential and a space dependent nonlinear damping. After showing that multiple resonances can be controlled by the system parameters, we show that the particle mass can allow or hinder V.R.. Indeed, V.R. ceases to exist for particles exceeding a critical mass. Keywords: Vibrational Resonance, nonlinear dynamics.

Monitoring of breathing rate considerably improves healthcare quality by providing information about the patient’s state. Yet, the current methods used to assess the respiratory rate have some drawbacks that could be improved. Recent developments of the non-contact based measurement of the respiratory rate brought the remote Photoplethysmography (rPPG) technique. This technique allows physiological parameters monitoring with a camera and ambient light. It measures the light absorption variation by the blood to extract several physiological parameters. The main issue with the rPPG method is that its signal quality is lower compared with the contact-based methods. Several ways to improve this quality may be explored to obtain a robust rPPG pipeline. One of these is the combination of the multidimensional temporal traces generated by a camera into a single temporal trace. The literature methods aimed to extract a signal with the maximal information using smart estimation of the combination coefficients. These coefficients are computed using statistical or physiological/optical properties of the traces. In this paper, we developed an algorithm that estimates a better respiratory signal than with regular methods. The algorithm named Energy Variance Maximization (EVM) estimates the linear combination that maximizes the Signal Noise Ratio (SNR) of the output trace. We compare our contribution with three state-of-the-art method: CHROM, PBV and PVM. The results obtained with this method are better than the state of the art methods, with the Mean Absolute Error (MAE) being 0.87 rpm better than CHROM, the best of compared literature methods.

Programming in high abstraction level can facilitate the development of digital signal processing systems. In the recent 20 years, high-level synthesis (HLS) has made significantly progress. This technique greatly benefits the R&D productivity of the Field Programmable Gate Array (FPGA) developments and helps for adding to the maintainability of the products by automating the C-to-RTL (register transfer language) conversion. However, due to the high complexity and computational intensity, image processing algorithms usually necessitate a higher abstraction environment than C-synthesis, and the current HLS tools do not have the ability of this kind. This paper presents a conception of very high-level synthesis method which allows fast prototyping and verifying the FPGA-based image processing designs in the MATLAB environment. We build a heterogeneous development flow by using currently available tool kits for verifying the proposed approach and evaluated it within two real-life applications. Experiment results demonstrate that it can effectively reduce the complexity of the development by automatically synthesizing the algorithm behavior from the user level into the low register transfer level and give play to the advantages of FPGA related to the other devices.

International audience; Micro-Expression (ME) recognition is a hot topic in computer vision as it presents a gateway to capture and understand human's daily emotions. It is nonetheless a challenging problem due to the fact ME typically being transient (lasting less than 200 ms) and subtle. Recent advances in machine learning enable new and effective methods to be adopted for solving diverse computer vision tasks. In particular, the use of deep learning techniques on large datasets outperforms classical approaches based on classical machine learning which rely on hand-crafted features. Even though available datasets for spontaneous ME are scarce and much smaller, using off-the-shelf Convolutional Neural Networks (CNNs) still demonstrates satisfactory classification results. However, these networks are heavy in terms of memory consumption and computational resources. This poses great challenges when deploying CNN-based solutions in many applications such as driver's monitoring or comprehension recognition in virtual classrooms, which demand fast and accurate recognition. As these networks are initially designed for tasks of different domains, they are over-parameterized and need to be optimized for ME recognition. In this paper, we propose a new network based on the well-known ResNet18 which we optimize for ME classification in two ways. Firstly, we reduce the depth of the network by removing residual layers. Secondly, we introduce a more compact representation of optical flow used as input to the network. We present extensive experiments and demonstrate that the proposed network obtains accuracies comparable to the state-of-the-art methods while significantly reducing the necessary memory space. Our best classification accuracy reaches 60.17% on the challenging composite dataset containing 5 objectives classes. Our method takes only 24.6 ms for classifying a ME video clip (less than the occurrence time of the shortest ME which lasts 40 ms). Our CNN design is suitable for real-time embedded applications with limited memory and computing resources.

International audience; Calibration of a camera have various applications in computer vision. This process is particularly challenging with a thermal camera instead of a standard color one. A calibration target with two layers of different thermal emissivity is commonly used but reflects appear and degrade the corner extraction of the chessboard. While traditional and commonly used methods heat the target chessboard, we cool it down that inverts the distribution of emission of the two layers. The reflects on the reflective layer enhance the contrast instead of decreasing it. Experiments on various configurations and orientations demonstrate the superiority of the proposed method.

International audience; For several decades, the camera spatial resolution is gradually increasing with the CMOS technology evolution. The image sensors provide more and more pixels, generating new constraints for suitable optics. As an alternative, promising solutions propose super-resolution (SR) techniques to reconstruct high-resolution images or video without modifying the sensor architecture. However, most of the SR implementations are far from reaching real-time performance on a low-budget hardware platform. Moreover, convincing state-of-the-art studies reveal that artifacts can be observed in highly textured areas of the image. In this paper, we propose a local adaptive spatial super resolution (LASSR) method to fix this limitation. LASSR is a two-step SR method including a machine learning-based texture analysis and a fast interpolation method that performs a pixel-by-pixel SR. Multiple evaluations of our method are also provided using standard image metrics for quantitative evaluation and also a psycho-visual assessment for a perceptual evaluation. A first FPGA-based implementation of the proposed method is then presented. It enables high-quality 2-4 k super-resolution videos to be performed at 16 fps, using only 13% of the FPGA capacity, opening the way to reach more than 60 fps by executing several parallel instances of the LASSR code on the FPGA.

International audience; This paper presents a new hardware implementation of a supervision system used in robot manipulators with two degrees of freedom. In addition to the simulation results, the new System Generator tool of Xilinx® is used to ensure self-generation of HDL codes. This code is used to configure field programmable gate arrays (FPGA) devices in the loop, and the supervision system is used mainly to ensure real-time reconfiguration of robots. In the proposed system, we used a new fault detection (FD) method for a viscous friction fault in the supervised robot combined with a fault-tolerance control method. The first module, based on residual analysis, is used to FD and to properly estimate the necessary corrections of the second module. For data transmission between the supervisor and the supervised robots, we used an approach based on the transmission control protocol. The simulation results show that the proposed method adjusts the fault effect using information transferred from the remote supervisor robot. The hardware implementation generated using Xilinx® System Generator is used to validate the proposed contribution and to ensure real-time processing in the case of industrial robots. The simulation results and the response times of both proposed systems are compared and discussed.

International audience; Machine learning has known a tremendous growth within the last years, and lately, thanks to that, some computer vision algorithms started to access what is difficult or even impossible to perceive by the human eye. While deep learning based computer vision algorithms have made themselves more and more present in the recent years, more classical feature extraction methods, such as the ones based on Local Binary Patterns (LBP), still present a non negligible interest, especially when dealing with small datasets. Furthermore, this operator has proven to be quite useful for facial emotions and human gestures recognition in general. Micro-Expression (ME) classification is among the applications of computer vision that heavily relied on hand crafted features in the past years. LBP Three Orthogonal Planes (LBP_TOP) is one of the most used hand crafted features extractor in the scientific literature to tackle the problem of ME classification. In this paper we present a time unification method that provides better results than the classical LBP_TOP while also drastically reducing the calculations required for feature extraction.

International audience; The authors reveal the first evidence of subthreshold noisy image perception, which takes benefit of the vibrational resonance phenomenon. Indeed, an appropriate amount of a high-frequency perturbation allows revealing a grey level noisy image through a threshold detector.

International audience; In this paper we present a new 3D descriptor for human classification and a human detection method based on this descriptor. The proposed 3D descriptor allows classification of an object represented by a point cloud, as human or non-human. It is derived from the well-known Histogram of Oriented Gradient by employing surface normals instead of gradients. The process consists in an appropriate subdivision of the object point cloud into blocks. These blocks provide the spatial distribution modeling of the surface normal orientation into the different parts of the object. This distribution modelling is expressed as a histogram. In addition we have set up a multi-kinect acquisition system that provides us with Complete Point Clouds (CPC) (i.e. 360° view). Such CPCs enable a suitable processing, particularly in case of occlusions. Moreover they allow for the determination of the human frontal orientation. Based on the proposed 3D descriptor, we have developed a human detection method that is applied on CPCs. First, we evaluated the 3D descriptor over a set of CPC candidates by using the Support Vector Machine (SVM) classifier. The learning process was conducted with the original CPC database that we have built. The results are very promising. The descriptor can discriminate human from non-human candidates and provides the frontal direction of humans with high precision. In addition we demonstrated that using the CPCs improves significantly the classification results in comparison with Single Point Clouds (i.e. points clouds acquired with only one kinect). Second, we compared our detection method with two others, namely the HOG detector on RGB images and a 3D HOG-based detection method that is applied on RGB-depth data. The obtained results on different situations show that the proposed human detection method provides excellent performances that outperform the other two detection methods.

International audience

Nowadays, camera networks are part of our every-day life environments, consequently, they represent a massive source of information for monitoring human activities and to propose new services to the building users. To perform human activity monitoring, people must be detected and the analysis has to be done according to the information relative to the environment and the context. Available multi-camera datasets furnish videos with few (or none) information of the environment where the network was deployed. The proposed dataset provides multi-camera multi-space video sets along with the complete contextual information of the environment. The dataset regroups 11 video sets (composed of 62 single videos) recorded using 6 indoor cameras deployed on multiple spaces. The video sets represent more than 1 h of video footage, include 77 people tracks and captured different human actions such as walking around, standing/sitting, motionless, entering/leaving a space and group merging/splitting. Moreover, each video has been manually and automatically annotated to include people detection and tracking meta-information. The automatic people detection annotations were obtained by using different complexity and robustness detectors, from machine learning to state-of-art deep Convolutional Neural Network (CNN) models. Concerning the contextual information, the Industry Foundation Classes (IFC) file that represents the environment's Building Information Modeling (BIM) data is also provided. The BIM/IFC file describes the complete structure of the environment, it's topology and the elements contained in it. To our knowledge, the WiseNET dataset is the first to provide a set of videos along with the complete information of the environment. The WiseNET dataset is publicly available at https://doi.org/10.4121/uuid:c1fb5962-e939-4c51-bfd5-eac6f2935d44, as well as at the project's website http://wisenet.checksem.fr/#/dataset. Keywords: Indoor multi-camera multi-space dataset, People detection, People tracking, BIM, Contextual information

International audience; This paper aims to provide a comprehensive reference source on depth-based human motion descriptors. Motion description is a challenging problem which became popular with recent advances in 3D computer vision. Our purpose is twofold. First, we introduce the main trends in human 3D motion descriptor design and evaluation. Second, we present a review of recent methods belonging to three different application categories: action recognition, gesture recognition and gait assessment. Selected categories have different specifics, which allow us to highlight aspects of a motion descriptor construction. A comparison of different methods by their main characteristics is provided. Finally, possible directions and recommendations for future research in 3D motion description are outlined.

International audience

International audience

International audience

International audience; The SITIS conference is dedicated to research on the technologies used to represent, share and process information in various forms, ranging from signal, image, and multimedia data to traditional structured data and semi-structured data found in the web.SITIS spans two inter-related research domains that increasingly play a key role in connecting systems across network centric environments to allow distributed computing and information sharing. SITIS 2018 aims to provide a forum for high quality presentations on research activities centered on the following main tracks.The track titled "Signal Image & Vision Technologies" (SIVT) focuses on recent developments in digital signal processing and pays particular attention to evolutions in audiovisual signal processing, analysis, coding and authentication, and retrieval techniques.The track titled “Intelligent Web Computing and Applications” (I-WeCA) is on emerging and novel concepts, architectures and methodologies for information management, including the semantic web, big data, and cloud computing. Finally, social networks have grown rapidly in size and have witnessed a popular expansion to reach billions of users, providing a suitable space for sharing data, interacting with other users and forming communities with common interests. The focus of the Track on "Large, Heterogeneous Networks and their Applications" (LHNA) track is on emerging and novel concepts, architectures and methodologies for social networks analysis, tools and applications, including issues related to complex graphs, semantics, big data, and security.SITIS is also dedicated to bring together scientists from the world over, particularly from developing countries, to share and exchange knowledge, experience and research results. When the conference is hosted by research institutions in developing countries, a particular emphasis will be placed on the presentation of applied research, the organization of practical tutorials and workshops that may help address the information processing needs of these countries, and book give away sessions to allow conference attendees to donate books to the local institutions.

International audience; The contextual information in the built environment is highly heterogeneous, it goes from static information (e.g., information about the building structure) to dynamic information (e.g., user's space-time information, sensors detections and events that occurred). This paper proposes to semantically fuse the building's contextual information with extracted data from a smart camera network by using ontologies and semantic web technologies. The developed ontology allows interoperability between the different contextual data and enables, without human interaction, real-time event detections and system reconfiguration to be performed. The use of semantic knowledge in multi-camera monitoring systems guarantees the protection of the user's privacy by not sending nor saving any image, just extracting the knowledge from them. This paper presents a new approach to develop an "all-seeing" smart building, where the global system is the first step to attempt to provide Artificial Intelligence (AI) to a building.

The aim of this paper is to highlight differences between classical machine learning and transfer learning applied to low cost real-time face authentication. Furthermore, in an access control context, the size of biometric data should be minimized so it can be stored on a remote personal media. These constraints have led us to compare only lightest versions of these algorithms. Transfer learning applied on Mobilenet v1 raises to 85% of accuracy, for a 457Ko model, with 3680s and 1.43s for training and prediction tasks. In comparison, the fastest integrated method (Random Forest) shows accuracy up to 90% for a 7,9Ko model, with a fifth of a second to be trained and a hundred of microseconds for the prediction, enabling embedded real-time face authentication at 10 fps.

International audience; Experimental means developed in the Actor Prolog parallel object-oriented logic language for implementation of heterogeneous multichannel intelligent visual surveillance systems are considered. These means are examined by the instance of a logic program for permanent monitoring of people's body parts temperature in the area of visual surveillance. The logic program implements a fusion of heterogeneous data acquired by two devices: (1) 3D coordinates of human body are measured using a time-of-flight (ToF) camera; (2) 3D coordinates of human body skeleton are computed on the base of 3D coordinates of the body; (3) a thermal video is acquired using a thermal imaging camera. Special built-in classes are developed in the Actor Prolog language for the acquisition and analysis of 3D and 2D video data: the Kinect and KinectBuffer classes implement interaction between the logic program and the ToF camera of the Kinect-2 device; the BufferedScene class implements storing and transferring 3D images; the Canvas3D class implements 3D graphics based on the Java3D open source library. In the considered example, the thermal video is projected to the 3D surface of the human body; then the temperature of the human body is projected to the vertices and edges of the skeleton. A special logical agent (i.e., the logic program written in Actor Prolog) implements these operations in real-time and transfers the data to another logical agent. The latter agent implements a time average of the temperature of the human skeletons and displays colored 3D images of the skeletons; the average temperature of the vertices and edges of the skeletons is depicted by colors. A method of remote predicate calls is used for the interaction of the logical agents; this method was developed and implemented in Actor Prolog to support the agent logic programming paradigm. The logic programming means under consideration are developed for the purpose of the implementation of logical analysis of video scene semantics in the intelligent visual surveillance systems. © 2018 Radioelektronika, Nanosistemy, Informacionnye Tehnologii. All rights reserved.

Markerless vision-based 3D hand motion tracking is a key and popular component for interaction studies in many domains such as virtual reality and natural human-computer interfaces. While this research field has been well studied in the last decades, most approaches have considered the human hand in isolation and not in action or in interaction with the environment or the other articulated human body parts. Employing contextual information about the surrounding environment (e.g. the shape, the texture, and the posture of the object in the hand) can remarkably constrain the tracking problem. The goal of this survey is to develop an up-to-date taxonomy of existing vision-based hand tracking methods with a new classification scheme: hand-object constraints. This taxonomy allows us to examine the strengths and weaknesses of the current state of the art and to highlight future trends in the domain.

International audience; Smart camera, i.e. cameras that are able to acquire and process images in real-time, is a typical example of the new embedded computer vision systems. A key example of application is automatic fall detection, which can be useful for helping elderly people in daily life. In this paper, we propose a methodology for development and fast-prototyping of a fall detection system based on such a smart camera, which allows to reduce the development time compared to standard approaches. Founded on a supervised classification approach, we propose a HW/SW implementation to detect falls in a home environment using a single camera and an optimized descriptor adapted to real-time tasks. This heterogeneous implementation is based on Xilinx’s system-on-chip named Zynq. The main contributions of this work are (i) the proposal of a codesignmethodology. These methodologies enable the HW/SW partitioning to be delayed using high-level algorithmic description and high-level synthesis tools. Our approach enables fast prototyping which allows fast architecture exploration and optimisation to be performed, (ii) the design of a hardware accelerator dedicated to boostingbased classification, which is a very popular and efficient algorithm used in image analysis, (iii) the proposal of falldetection embedded in a smart camera and enabling integration into the elderly people environment. Performances of our system are finally compared to the state-of-the-art.

International audience; Segmentation is a critical step for many algorithms, especially for remote photoplethysmography (rPPG) applications as only the skin surface provides information. Moreover, it has been shown that the rPPG signal is not distributed homogeneously across the skin. Most of the time, algorithms get input information from face detection provided by a supervised learning of physical appearance and skin pixel selection. However, both methods show several limitations. In this paper, we propose a simple approach to implicitly select skin tissues based on their distinct pulsatility feature. The input video frames are decomposed into several temporal superpixels from which the pulse signals are extracted. A pulsatility measure from each temporal superpixel is then used to merge the pulse traces and estimate the photoplethysmogram signal. Since the most pulsatile signals provide high quality information, areas where the information is predominant are favored. We evaluated our contribution using a new publicly available dataset dedicated to rPPG algorithms comparison. The results of our experiments show that our method outperforms state of the art algorithms, without any critical face or skin detection

SummaryObjectives: We developed a novel system for in home functional capacities assessment in frail older adults by analyzing the Timed Up and Go movements. This system aims to follow the older people evolution, potentially allowing a forward detection of motor decompensation in order to trigger the implementation of rehabilitation. However, the pre-experimentations conducted on the ground, in different environments, revealed some problems which were related to KinectTM operation. Hence, the aim of this actual study is to develop methods to resolve these problems.Methods: Using the KinectTM sensor, we analyze the Timed Up and Go test movements by measuring nine spatio-temporal parameters, identified from the literature. We propose a video processing chain to improve the robustness of the analysis of the various test phases: automatic detection of the sitting posture, patient detection and three body joints extraction. We introduce a realistic database and a set of descriptors for sitting posture recognition. In addition, a new method for skin detection is implemented to facilitate the patient extraction and head detection. 94 experiments were conducted to assess the robustness of the sitting posture detection and the three joints extraction regarding condition changes.Results: The results showed good performance of the proposed video processing chain: the global error of the sitting posture detection was 0.67%. The success rate of the trunk angle calculation was 96.42%. These results show the reliability of the proposed chain, which increases the robustness of the automatic analysis of the Timed Up and Go.Conclusions: The system shows good measurements reliability and generates a note reflecting the patient functional level that showed a good correlation with 4 clinical tests commonly used. We suggest that it is interesting to use this system to detect impairment of motor planning processes.Citation: Hassani A, Kubicki A, Mourey F, Yang F. Advanced 3D movement analysis algorithms for robust functional capacity assessment. Appl Clin Inform 2017; 8: 454–469 https://doi.org/10.4338/ACI-2016-11-RA-0199

International audience; Due to the growing popularity of high-dynamic range (HDR) image and the high complexity to capture HDR image, researchers focus on converting low-dynamic range (LDR) content to HDR, which gives rise to a number of dynamic range expansion methods. Most of the existing methods try their best to tackle highlight areas during the expanding, however, in some cases, they cannot achieve approving results. In this study, a novel LDR image expansion technique is presented. The technique first detects the highlight areas in image; then preprocesses them and reconstructs the information of these regions; finally, expands the LDR image to HDR. Unlike the existing schemes, the proposed approach escapes the complicated treatment to highlight areas in the process of expansion, which makes the expansion straightforward; at the same time, it facilitates the expansion scheme and minimises the formation of the artefacts. The experimental results show that the proposed method performs well; the tone mapped versions of the produced HDR images are popular. The results of the image quality metric also illustrate that the novel approach can recover more image details with minimised contrast loss and reversal, compared with the existing schemes considered in the comparison.

International audience; The newly introduced Kubelka-Munk Genetic Algorithm (KMGA) is a promising technique for the assessment of skin lesions from multi-spectral images. Using five skin parameter maps such as concentration or epidermis/dermis thickness, this method combines the Kubelka-Munk Light-Tissue interaction model and Genetic Algorithm optimization process to produce a quantitative measure of cutaneous tissue. Up to the present, variant improved KMGA implementations have been successfully realized using the recent parallel computing techniques. However, all these achievements are based on the multi-core CPUs. This results in a quite high cost and low practicability for the hardware equipment of the clinical system. Fortunately, Embedded Systems (ES) applications have made great progress in recent years, and many highly effective image processing devices, such as DSPs (Digital Signal Processor) and FPGAs (Field Programmable Gate Array), have been made available to engineers at a very convenient price. Nevertheless, today's embedded devices have as well the advantages of high speed, high embedability, low power consumption, more flexibility, etc. Thus, we focus our researches on the embedded KMGA application development. In this paper, we realize the CPU-to-FPGA transplantation of KMGA within a special High-Level Synthesis (HLS) SW/HW Co-design framework. Moreover, several optimizations are made on the algorithm and source code to improve the performances of the final implementation. Compared with CPUs, intensive experiments demonstrate that the proposed approaches can effectively improve the performances of KMGA method both in terms of efficiency and accuracy. (C) 2015 Elsevier B.V. All rights reserved.

International audience

International audience; We studied age-related changes in motor-planning processes through the kinematic features of Sit-to-Stand (STS) and Back-to-Sit (BTS) transfers in aged adults. Our objective was to test the relationship between trunk angle during BTS transfer and functional level. The STS and BTS were analyzed with the Kinect motion capture system in sixty aged patients (33 women; mean age 84 ± 5). The statistical analysis was based on test-retest reliability, and we used Pearson coefficients to measure the robustness of the link between trunk angles and the timed-up and go (TUG) scores. We showed that patients with a lower functional level reached a smaller trunk angle during the BTS. Interestingly, this is even more pronounced in frail aged subjects. We suggest that this decrease of trunk angle during the BTS transfer signs an impairment of motor planning processes that should be checked in the context of frailty.; Il était question ici d’étudier l’impact du vieillissement sur les processus de planification motrice à travers les caractéristiques cinématiques des transferts assis-debout (STS) et debout-assis (BTS) chez des personnes âgées. Notre objectif était d’identifier une relation entre l’inclinaison de l’angle du tronc pendant le BTS et le niveau fonctionnel des sujets âgés. Les transferts STS et BTS étaient analysés à partir du système de capture du mouvement Kinect chez 60 patients âgés (33 femmes ; moyenne d’âge de 84 ± 5). L’analyse statistique était basée sur un test-retest de fiabilité et nous avons utilisé le coefficient de Pearson afin de mesurer la robustesse du lien entre les angles du tronc pendant les transferts et le score obtenu par les participants lors du test fonctionnel du Timed Up and Go (TUG). Les résultats montrent que les patients qui atteignent un faible niveau fonctionnel sont également ceux qui ont un plus petit angle d’inclinaison du tronc pendant le BTS. D’une manière intéressante, cette corrélation est beaucoup plus marquée chez les personnes âgées fragiles. Nous suggérons que cette diminution de l’angle d’inclinaison du tronc pendant le BTS est un marqueur pertinent d’un trouble de planification motrice. Ce paramètre devrait être systématiquement contrôlé dans un contexte de fragilité.

This is an open access article under the CC BY-NC-ND license(http://creativecommons.org/licenses/by-nc-nd/4.0/).Agrosupéquipe BAPPôle GEAPSI; Precision fertilization requires new techniques for determining the spread pattern of fertilizer spreaders. Because of the accuracy and non-intrusive nature, techniques based on digital image processing are most promising. Using image processing, dynamics of particles leaving the spreader can be determined. Combined with a ballistic flight model, this allows predicting the landing position of individual fertilizer particles. In a first approach, a two-dimensional imaging technique was used with small field of view (0.33 m on 0.25 m). In the second approach, a larger field of view (1 m on 1 m) was used. To improve the accuracy of previous technique, binocular stereovision was used to determine three-dimensional information.

International audience; Better characterization of the fertilizer spreading process, especially the fertilizer pattern distribution on the ground, requires an accurate measurement of individual particle properties and dynamics. Both 2D and 3D high speed imaging techniques have been developed for this purpose. To maximize the accuracy of the predictions, a specific illumination level is required. This paper describes the development of a high irradiance LED system for high speed motion estimation of fertilizer particles. A spectral sensitivity factor was used to select the optimal LED in relation to the used camera from a range of commercially available high power LEDs. A multiple objective genetic algorithm was used to find the optimal configuration of LEDs resulting in the most homogeneous irradiance in the target area. Simulations were carried out for different lenses and number of LEDs. The chosen configuration resulted in an average irradiance level of 452 W/m2 with coefficient of variation less than 2%. The algorithm proved superior and more flexible to other approaches reported in the literature and can be used for various other applications.

International audience; Glass workpieces are mainly planar and obtained by diamond cutting however if free-form surfaces are required, manufacturing process is usually based on shaping grinding wheels. Multi-axis waterjet cutting, an other means of obtaining planar workpieces, could also be used to machine complex shapes with appropriate manufacturing strategies. Water jet pocketing could also be achieved but it brings challenging issues since the high pressure jet composed of water and abrasive particles must be contained inside the machined pocket. Water jet glass machining optimization requires understanding of numerous parameters such as interaction between the jet and the brittle material behavior or the identification of the jet himself. We focus our investigation on bottom pocket surfaces to study these parameters. Pocket bottom surfaces are characterized by multi-scale defects: shape defects influenced by the tool path and the manufacturing strategy, macro-craters at waviness scale due to the jet interactions, micro-craters at roughness scale due to particles impacts, sub-surface damages (SSD) at micro- roughness scale. This paper focuses on the study of macro and micro craters. We propose a decomposition of the multi-scale defects using modal filtering. Residual topography will then be analyzed to characterize the surface damages.

Display Omitted An algorithm for face detection has been implemented on CPU.An acceleration of this algorithm on GPU migration.Performance of GPU implementation shows the effectiveness of this implementation.Another optimization method on GPU are operated. Face detection is an important aspect for various domains such as: biometrics, video surveillance and human computer interaction. Generally a generic face processing system includes a face detection, or recognition step, as well as tracking and rendering phase. In this paper, we develop a real-time and robust face detection implementation based on GPU component. Face detection is performed by adapting the Viola and Jones algorithm. We have developed and designed optimized several parallel implementations of these algorithms based on graphics processors GPU using CUDA (Compute Unified Device Architecture) description.First, we implemented the Viola and Jones algorithm in the basic CPU version. The basic application is widened to GPU version using CUDA technology, and freeing CPU to perform other tasks. Then, the face detection algorithm has been optimized for the GPU using a grid topology and shared memory. These programs are compared and the results are presented. Finally, to improve the quality of face detection a second proposition was performed by the implementation of WaldBoost algorithm.

Accounting for customers' perception of manufactured goods has become a major challenge for the industry. This process is to be established from early design to retail. Customers are nowadays more aware and detail oriented about perceived quality of products. This allows one to set not only an estimated price but also the expected quality of the product. Surface appearance analysis has therefore become a key industrial issue. Two approaches are proposed here to formalize the detection methodology and provide objective criteria for experts to evaluate surface anomalies. The first proposed approach is based on surface metrology. It consists in analyzing the measured topologies in order to bind aspect to geometric characteristics. A multi-scale procedure based on Discrete Modal Decomposition is implemented and allows an effective separation of geometric variations. Accordingly, appearance anomalies can be isolated from other geometrical features. This method enables the calculation of surface curvatures in a simplified and robust manner. It is shown that such geometric information is relevant and bound to visual aspect. The presented work also emphasizes the influence of raw data in aspect analysis. Two main metrological difficulties are investigated: the presence of outliers (High frequencies) and the presence of non surface-related geometric defects, generated by the measuring device (Low frequencies). An innovative method for identifying outliers in surface metrology is presented. It is based on a multi-scale statistical approach. Finally, the issue of geometrical variation due to positioning tables is also addressed. A calibration protocol based on DMD that intends to correct this phenomenon is proposed. The second proposed approach, more global, is based on the interaction of a surface with its light environment. It aims at providing experts with assistance, specifically during the anomaly detection phase. The presented work uses Polynomial Texture Mapping. This technique consists of calculating the reflectance at each point of the surface and simulating its appearance while the lighting angles vary. A surface Inspection Support Device based on this principle is presented and detailed. Finally, an industrial study is proposed that shows how these two academic approaches can be combined within a global industrial methodology dedicated to surface appearance quality., Parmi tous les leviers à disposition des entreprises, la prise en compte de la perception par les clients est aujourd'hui centrale, dès la conception des produits. En effet, le consommateur est aujourd'hui mieux informé et attentif à ce qu'il perçoit de la qualité d'un produit et cette perception lui permet d'établir une valeur d'estime de la qualité esthétique des produits, mais aussi de ses fonctionnalités techniques. La méthodologie de l'analyse de la qualité d'aspect des surfaces est donc un enjeu essentiel pour l'industrie. Deux approches de la fonctionnalité des surfaces sont proposées afin de formaliser la méthodologie de détection, et d'apporter aux experts des critères objectifs d'évaluation des anomalies. La première approche proposée est basée sur la métrologie des surfaces. Elle consiste à analyser les topographies mesurées pour lier la fonction aspect aux caractéristiques géométriques extraites. Une approche multi-échelle basée sur la Décomposition Modale Discrète est mise en oeuvre afin de séparer efficacement les différents ordres de variations géométriques d'une surface, et ainsi d'isoler les anomalies d'aspect. D'autre part, cette méthode permet la mise en oeuvre du calcul des courbures sur une surface de façon simplifiée et robuste. On montre que cet attribut géométrique apporte une information supplémentaire et pertinente en lien avec la fonction aspect. Enfin, ces travaux ont mis en évidence l'importance de la qualité des données sources pour analyser l'aspect, et particulièrement deux difficultés d'ordre métrologiques, liées à la présence de points aberrants (hautes fréquences) et de variations géométriques non intrinsèques aux surfaces, générées par le moyen de mesure (basses fréquences). Une méthode innovante d'identification des points aberrants dédiée à la métrologie des surfaces et basée sur une approche statistique multi-échelle est proposée. La problématique des variations géométriques liées aux tables de positionnement du moyen de mesure est traitée au moyen de la Décomposition Modale, et un protocole pour corriger ces variations est présenté. La seconde approche, plus globale, est basée sur l'interaction entre les surfaces et l'environnement lumineux. L'objet de cette approche de l'analyse de l'aspect est d'apporter une aide aux experts pour mieux détecter les anomalies. Les travaux présentés sont basés sur la technique Polynomial Texture Mappings et consistent à modéliser la réflectance en chaque point des surfaces afin de simuler le rendu visuel sous un éclairage quelconque, à la manière de ce que font les opérateurs en analyse sensorielle pour faciliter la détection. Un dispositif d'aide à l'inspection des surfaces basé sur ce principe est présenté. Enfin, une approche industrielle est proposée afin de montrer comment ces 2 axes de recherche peuvent être complémentaires dans le cadre d'une méthodologie globale, industrielle, de l'analyse de la qualité d'aspect de surfaces.

International audience; This paper presents a metrological analysis of hip prosthesis components. When changing ceramic prostheses, the surgeon sometimes finds traces of alloy lying in the insert or the femoral head. These traces can be thin and accurate or as a wide band. From the measurements made on the contact areas of hip prosthesis components, we analyse these phenomena by highlighting the defects of form, waviness and roughness of these surfaces using a novel geometric parameterization (namely modal parameterization). The aim of this work is to isolate these defects to characterize the prostheses components. We show that this parameterization allows a multi-scale analysis of surfaces regardless of the type of wear of the prosthesis, and that the results offer some relevant explanations to the analysis of visible damage on the prostheses. In a later study, we are going to analyse the defects influence on the alteration of the performance of hip prostheses.

Since several decades, the camera spatial resolution is gradually increasing with the CMOS technology evolution. The image sensors provide more and more pixels, generating new constraints for the suitable optics. As an alternative, promising solutions propose Super Resolution (SR) image reconstruction to extend the image size without modifying the sensor architecture. Convincing state-of art studies demonstrate that these methods could even be implemented in real-time. Nevertheless, artifacts can be observed in highly textured areas of the image. In this paper, we propose a Local Adaptive Spatial Super Resolution (LASSR) method to fix this limitation. A real-time texture analysis is included in a spatial super resolution scheme. The proposed local approach enables adjustments to intrinsic heterogeneity of image texture. A FPGA-based implementation of the proposed method is then presented and enables high quality 4k super resolution videos to be performed at 16 fps, from a standard 2K video flow.