Search

Our approach to action recognition is grounded in the intrinsic coexistence of and complementary relationship between audio and visual information in videos. Going beyond the traditional emphasis on visual features, we propose a transformer-based network that integrates both audio and visual data as inputs. This network is designed to accept and process spatial, temporal, and audio modalities. Features from each modality are extracted using a single Swin Transformer, originally devised for still images. Subsequently, these extracted features from spatial, temporal, and audio data are adeptly combined using a novel modal fusion module (MFM). Our transformer-based network effectively fuses these three modalities, resulting in a robust solution for action recognition.

This paper provides a modular architecture with deep neural networks as a solution for real-time video analytics in an edge-computing environment. The modular architecture consists of two networks of Front-CNN (Convolutional Neural Network) and Back-CNN, where we adopt Shallow 3D CNN (S3D) as the Front-CNN and a pre-trained 2D CNN as the Back-CNN. The S3D (i.e., the Front CNN) is in charge of condensing a sequence of video frames into a feature map with three channels. That is, the S3D takes a set of sequential frames in the video shot as input and yields a learned 3 channel feature map (3CFM) as output. Since the 3CFM is compatible with the three-channel RGB color image format, we can use the output of the S3D (i.e., the 3CFM) as the input to a pre-trained 2D CNN of the Back-CNN for the transfer-learning. This serial connection of Front-CNN and Back-CNN architecture is end-to-end trainable to learn both spatial and temporal information of videos. Experimental results on the public datasets of UCF-Crime and UR-Fall Detection show that the proposed S3D-2DCNN model outperforms the existing methods and achieves state-of-the-art performance. Moreover, since our Front-CNN and Back-CNN modules have a shallow S3D and a light-weighted 2D CNN, respectively, it is suitable for real-time video recognition in edge-computing environments. We have implemented our CNN model on NVIDIA Jetson Nano Developer as an edge-computing device to show its real-time execution.

A pre-trained 2D CNN (Convolutional Neural Network) can be used for the spatial stream in the two-stream CNN structure for videos, treating the representative frame selected from the video as an input. However, the CNN for the temporal stream in the two-stream CNN needs training from scratch using the optical flow frames, which demands expensive computations. In this paper, we propose to adopt a pre-trained 2D CNN for the temporal stream to avoid the optical flow computations. Specifically, three RGB frames selected at three different times in the video sequence are converted into grayscale images and are assigned to three R(red), G(green), and B(blue) channels, respectively, to form a Stacked Grayscale 3-channel Image (SG3I). Then, the pre-trained 2D CNN is fine-tuned by SG3Is for the temporal stream CNN. Therefore, only pre-trained 2D CNNs are used for both spatial and temporal streams. To learn long-range temporal motions in videos, we can use multiple SG3Is by partitioning the video shot into sub-shots and a single SG3I is generated for each sub-shot. Experimental results show that our two-stream CNN with the proposed SG3Is is about 14.6 times faster than the first version of the two-stream CNN with the optical flow, and yet achieves a similar recognition accuracy for UCF-101 and a 5.7% better result for HMDB-51.

Most conventional fine-grained image recognitions are based on a two-stream model of object-level and part-level CNNs, where the part-level CNN is responsible for learning the object-parts and their spatial relationships. To train the part-level CNN, we first need to separate parts from an object. However, there exist sub-level objects with no distinctive and separable parts. In this paper, a multi-scale CNN with a baseline Object-level and multiple Part-level CNNs is proposed for the fine-grained image recognition with no separable object-parts. The basic idea to train different CNNs of the multi-scale CNNs is to adopt different scales in resizing the training images. That is, the training images are resized such that the entire object appears as much as possible for the Object-level CNN, while only a local part of the object is to be included for the Part-level CNN. This scale-specific image resizing approach requires a scale-controllable parameter in the image resizing process. In this paper, a scale-controllable parameter is introduced for the linear-scaling and random-cropping method. Also, a line-based image resizing method with a scale-controllable parameter is employed for the part-level CNNs. The proposed multi-scale CNN is applied to a food image classification, which belongs to a fine-grained classification problem with no separable object-parts. Experimental results on the public food image datasets show that the classification accuracy improves substantially when the predicted scores of the multi-scale CNN are fused together. This reveals that the object-level and part-level CNNs work harmoniously in differentiating subtle differences of the sub-level objects.

This paper deals with the problem of training convolutional neural networks (CNNs) with facial action units (AUs). In particular, we focus on the imbalance problem of the training datasets for facial emotion classification. Since training a CNN with an imbalanced dataset tends to yield a learning bias toward the major classes and eventually leads to deterioration in the classification accuracy, it is required to increase the number of training images for the minority classes to have evenly distributed training images over all classes. However, it is difficult to find the images with a similar facial emotion for the oversampling. In this paper, we propose to use the AU features to retrieve an image with a similar emotion. The query selection from the minority class and the AU-based retrieval processes repeat until the numbers of training data over all classes are balanced. Also, to improve the classification accuracy, the AU features are fused with the CNN features to train a support vector machine (SVM) for final classification. The experiments have been conducted on three imbalanced facial image datasets, RAF-DB, FER2013, and ExpW. The results demonstrate that the CNNs trained with the AU features improve the classification accuracy by 3%-4%.

SMD (Singapore Maritime Dataset) is a public dataset with annotated videos, and it is almost unique in the training of deep neural networks (DNN) for the recognition of maritime objects. However, there are noisy labels and imprecisely located bounding boxes in the ground truth of the SMD. In this paper, for the benchmark of DNN algorithms, we correct the annotations of the SMD dataset and present an improved version, which we coined SMD-Plus. We also propose augmentation techniques designed especially for the SMD-Plus. More specifically, an online transformation of training images via Copy & Paste is applied to solve the class-imbalance problem in the training dataset. Furthermore, the mix-up technique is adopted in addition to the basic augmentation techniques for YOLO-V5. Experimental results show that the detection and classification performance of the modified YOLO-V5 with the SMD-Plus has improved in comparison to the original YOLO-V5. The ground truth of the SMD-Plus and our experimental results are available for download.

Conventional image resizing problems demand hard conditions on size and aspect ratio, which must be met with no tolerance. In this paper, a generalized optimization framework is presented, which can handle soft conditions as well as the hard ones. The soft condition can be given by an allowable range of the image parameter, which is incorporated as an inequality condition in the constrained optimization framework. Given the soft constraints, the proposed framework seeks to find the set of image parameters that minimize the cost function. A constrained optimization via a linear programming framework is employed to manage a diverse combination of soft and hard conditions for the target image. The optimization is based on the image line, which optimally selects a set of image lines (columns and rows) to be deleted for size reduction in accordance with the cost function and the constraints. As a case study, the line-based optimal image resizing method based on the linear programming framework is applied for the pre-processing of VGG-19 convolutional neural network (CNN). Although the target input size is a hard condition of 224 × 224 for the VGG-19 CNN, the proposed optimization framework with a soft condition on the image size firstly finds an optimal near-square image with a tradeoff against the saliency level of image features. Then, the optimal near-square image is linearly scaled to the final image size to meet the hard condition.

Depth maps taken by the low cost Kinect sensor are often noisy and incomplete. Thus, post-processing for obtaining reliable depth maps is necessary for advanced image and video applications such as object recognition and multi-view rendering. In this paper, we propose adaptive directional filters that fill the holes and suppress the noise in depth maps. Specifically, novel filters whose window shapes are adaptively adjusted based on the edge direction of the color image are presented. Experimental results show that our method yields higher quality filtered depth maps than other existing methods, especially at the edge boundaries.

Calibration between color camera and 3D Light Detection And Ranging (LIDAR) equipment is an essential process for data fusion. The goal of this paper is to improve the calibration accuracy between a camera and a 3D LIDAR. In particular, we are interested in calibrating a low resolution 3D LIDAR with a relatively small number of vertical sensors. Our goal is achieved by employing a new methodology for the calibration board, which exploits 2D-3D correspondences. The 3D corresponding points are estimated from the scanned laser points on the polygonal planar board with adjacent sides. Since the lengths of adjacent sides are known, we can estimate the vertices of the board as a meeting point of two projected sides of the polygonal board. The estimated vertices from the range data and those detected from the color image serve as the corresponding points for the calibration. Experiments using a low-resolution LIDAR with 32 sensors show robust results.

Segmentation of human bodies in images is useful for a variety of applications, including background substitution, human activity recognition, security, and video surveillance applications. However, human body segmentation has been a challenging problem, due to the complicated shape and motion of a non-rigid human body. Meanwhile, depth sensors with advanced pattern recognition algorithms provide human body skeletons in real time with reasonable accuracy. In this study, we propose an algorithm that projects the human body skeleton from a depth image to a color image, where the human body region is segmented in the color image by using the projected skeleton as a segmentation cue. Experimental results using the Kinect sensor demonstrate that the proposed method provides high quality segmentation results and outperforms the conventional methods.

Mobile robot operators must make rapid decisions based on information about the robot’s surrounding environment. This means that terrain modeling and photorealistic visualization are required for the remote operation of mobile robots. We have produced a voxel map and textured mesh from the 2D and 3D datasets collected by a robot’s array of sensors, but some upper parts of objects are beyond the sensors’ measurements and these parts are missing in the terrain reconstruction result. This result is an incomplete terrain model. To solve this problem, we present a new ground segmentation method to detect non-ground data in the reconstructed voxel map. Our method uses height histograms to estimate the ground height range, and a Gibbs-Markov random field model to refine the segmentation results. To reconstruct a complete terrain model of the 3D environment, we develop a 3D boundary estimation method for non-ground objects. We apply a boundary detection technique to the 2D image, before estimating and refining the actual height values of the non-ground vertices in the reconstructed textured mesh. Our proposed methods were tested in an outdoor environment in which trees and buildings were not completely sensed. Our results show that the time required for ground segmentation is faster than that for data sensing, which is necessary for a real-time approach. In addition, those parts of objects that were not sensed are accurately recovered to retrieve their real-world appearances.

Terrain classification allows a mobile robot to create an annotated map of its local environment from the three-dimensional (3D) and two-dimensional (2D) datasets collected by its array of sensors, including a GPS receiver, gyroscope, video camera, and range sensor. However, parts of objects that are outside the measurement range of the range sensor will not be detected. To overcome this problem, this paper describes an edge estimation method for complete scene recovery and complete terrain reconstruction. Here, the Gibbs-Markov random field is used to segment the ground from 2D videos and 3D point clouds. Further, a masking method is proposed to classify buildings and trees in a terrain mesh.

Remote vehicle operator must quickly decide on the motion and path. Thus, rapid and intuitive feedback of the real environment is vital for effective control. This paper presents a real-time traversable ground surface segmentation and intuitive representation system for remote operation of mobile robot. Firstly, a terrain model using voxel-based flag map is proposed for incrementally registering large-scale point clouds in real time. Subsequently, a ground segmentation method with Gibbs-Markov random field (Gibbs-MRF) model is applied to detect ground data in the reconstructed terrain. Finally, we generate a texture mesh for ground surface representation by mapping the triangles in the terrain mesh onto the captured video images. To speed up the computation, we program a graphics processing unit (GPU) to implement the proposed system for large-scale datasets in parallel. Our proposed methods were tested in an outdoor environment. The results show that ground data is segmented effectively and the ground surface is represented intuitively.

Moving objects of interest (MOOIs) in surveillance videos are detected and encapsulated by bounding boxes. Since moving objects are defined by temporal activities through the consecutive video frames, it is necessary to examine a group of frames (GoF) to detect the moving objects. To do that, the traces of moving objects in the GoF are quantified by forming a spatiotemporal gradient map (STGM) through the GoF. Each pixel value in the STGM corresponds to the maximum temporal gradient of the spatial gradients at the same pixel location for all frames in the GoF. Therefore, the STGM highlights boundaries of the MOOI in the GoF and the optimal bounding box encapsulating the MOOI can be determined as the local areas with the peak average STGM energy. Once an MOOI and its bounding box are identified, the inside and outside of it can be treated differently for object-aware size reduction. Our optimal encapsulation method for the MOOI in the surveillance videos makes it possible to recognize the moving objects even after the low bitrate video compressions.