Search

Counting the number of people is a common and basic computing operation in many applications. Most of the people counting techniques need a sensing device like camera and apply image processing methods to track pedestrians. However, counting people with cameras in private places raises a lot of security and privacy issues. The passive infra-red sensor (PIR) can detect the body temperature of the infrared and thus provides another promising solution. Although a single PIR can easily identify the passing situations (i.e., in or out) of a single person, the signals of a single PIR is not sufficient to identify the complex situations of multiple people. In the paper, we design a people counting device with a PIR array to detect the passing situations and generate data records with higher discriminability. In addition, we apply the machine learning classification methods including the CNN, the RBM+LR, Decision Tree, and NaiveBayes on the collected data records to identify the passing situations. To validate our design, we conduct experiments to study the feasibility and classification performance and explore the impact factors. The experimental results show that the CNN outperforms the other and achieves the best accuracy, i.e., about 92%. Also, the results show that the captured data records of the PIR array contain sufficient characteristics for identifying complex passing situations and the configuration of the PIR array including the sensor direction and the field of view (FOV) of a PIR modified by the metal tape can significantly impact the discriminability of the collected data.

Single-instruction set architecture (ISA) heterogeneous multi-processor architecture is promising for developing multi-processor system-on-chips (MPSoCs). In this architecture, all processors execute the same instruction set, yet with various performance and power behavior, since processors may have various micro-architectures. Therefore, systems with this architecture have the advantages of easy to develop new functions as the homogeneous architecture, and easy to customize the resource allocation to achieve high energy efficiency as the heterogeneous architecture. However, for an MPSoC utilizing the target architecture, a key design issue is how to select the set of processors so that the target system can achieve good performance while the cost of the chip is constrained to the expected value. To solve this, in this paper, we propose a processor allocation method for MPSoCs with single-ISA heterogeneous multi-core architecture. The goal of the proposed method is to automatically synthesize the allocation of cores for the given workload so that the performance is optimized while the resource constraint is met. To the best of our knowledge, this is the first work that tackles the processor allocation problem for MPSoCs with the target architecture. To bring out the best performance of a hardware configuration, the proposed algorithm also synthesizes the software design of task mapping for a selected hardware configuration. The experimental results show that, compared with the homogeneous architecture with the least cost and lowest performance cores only, even if the number of core is set to the maximum parallelism degree of the target workload, the proposed method achieves up to 8.25% of performance improvement among all the cases we evaluated while the area constraint is met. Compared with the architecture with all high performance but large cores, when the number of cores is also set to the same as the maximum parallelism degree of the target workload, the proposed method has at most 11.5% of performance degradation, while the area cost is reduced by 60.7%.

When deploying two or more well-trained deep-learningmodels on a system, we would hope to unify them into asingle deep model for the execution in the inference stage, sothat the computation time can be increased and the energyconsumption can be saved. This paper presents an effective method to build a single deep neural network that canexecute multiple tasks. Our approach can merge two well-trained feed-forward neural networks of the same architecture into a single one, where the required on-line storageis reduced and the inference speed is enhanced. We evalu-ate our approach by using MobileNets and show that ourapproach can improve both compression ratio and speedup. The experimental results demonstrate the satisfactory per-formance and verify the feasibility of the method.

It is hard to detect on-road objects under various lighting conditions. To improve the quality of the classifier, three techniques are used. We define subclasses to separate daytime and nighttime samples. Then we skip similar samples in the training set to prevent overfitting. With the help of the outside training samples, the detection accuracy is also improved. To detect objects in an edge device, Nvidia Jetson TX2 platform, we exert the lightweight model ResNet-18 FPN as the backbone feature extractor. The FPN (Feature Pyramid Network) generates good features for detecting objects over various scales. With Cascade R-CNN technique, the bounding boxes are iteratively refined for better results., Comment: Accepted at IEEE 21st International Workshop on Multimedia Signal Processing (MMSP 2019)

Multi-Processor Systems-on-Chips (MPSoCs) with 3D-stakced memories frequently work under thermal emergent status due to its high power density. Several thermal-aware task allocation or data placement methods have been proposed for 3D ICs to reduce the number of time-consuming dynamic thermal managements techniques being invoked. However, we observe that, these thermal-aware software designs all consider task allocation or data placement only. Studies show that, with the increasing number of stacked memories and the widening of vertical buses, heat generated by memories is comparable to processors, and synergistically performing thermal control on both processors and memories is a must since vertically aligned modules have the greatest thermal impacts to each other. So, in this paper, we propose the first thermal-aware task and data co-allocation method for MPSoCs with 3D-stacked memories. The proposed method synergistically places data and task considering the heterogeneity of cores and memories to optimize system performance under the given thermal constraint. Among all our test cases, compared to a performance-aware software design, the proposed method has at most 26% performance degradation while the system temperature are kept under the threshold and the performance-aware method has 108.4°C over the threshold. Compared to thermal-aware design that respectively considers data allocation and task allocation only, the proposed method achieves 9.76% of performance improvements on the average.

Location management is an important issue for object tracking applications of wireless sensor networks. Since locations obtained by most positioning techniques are inherently imprecise, users may send approximate location queries with precision constraints to trade for energy consumption of sensor nodes. Therefore, we propose an Adaptive Location Management scheme called ALM to process approximate location queries. ALM employs a two-tier storage architecture (i.e., centric storage node and local storage node) to facilitate approximate query resolving and to reduce the energy consumption of sensor nodes. We propose a storage node relocation and replication mechanism, which can create, remove replicas of storage nodes and adjust the positions of storage nodes according to the location queries and updates. As such, the number of forwarding location update and query messages is reduced by the proposed storage node relocation and replication mechanism, thereby conserving more energy. The experimental results show that compared with EASE (Xu et al. (2008) [18]), ALM is able to reduce the number of transmission messages and the energy consumption of sensor network, and prolong the network lifetime.

The IEEE 802.11e MAC standard can be used to support the quality of service (QoS) for wireless networks, but the bandwidth resource management is not efficient enough for the wireless multimedia network application. In this paper, a novel cross-layer architecture based on both the bandwidth measurement and queue adaptation is proposed to improve and adjust the video transmission over wireless multimedia networks. In other words, an average bandwidth measurement approach is used to adjust the video delivery immediately. Furthermore, the proposed cross-layer approach uses a dynamic queue adaptation for the current communication protocol of wireless multimedia networks. In addition, it is shown by NS2 simulations that the method of average bandwidth measurement does not consume any network resource, and the proposed method with dynamic queue adaptation can provide a more acceptable service quality by dynamically adapting the transmission rate.

Pedestrian and motorbike detection are two important areas in obstacle detection on road. Most state-of-the-art detectors are constructed with new features or learning methods on Histograms of Oriented Gradients (HOG) features. However, few researches focus on analyzing which features are complementary for the aforementioned detection. According to our study of pedestrians and motorbikes, there are three major properties including shape, texture, and self-similarity. We design a Shape, Texture and Self-Similarity (STSS) feature for these properties. The features we have employed here are HOG, Local Oriented Pattern (LOP), Color Self-Similarity (CSS), and Texture Self-Similarity (TSS). The STSS detector which combines Shape, Texture, and Self-Similarty features achieves 31% log-average miss rate. At the same time, 93% detection rate at 10-4 false positives per window on INRIA Person Dataset has also been concluded. Besides, we also have evaluated our detector on Caltech Motorbike Dataset and Caltech Pedestrian Dataset, and found the detector outperforms HOG detector in these datasets. As a result, we have shown that these features are complement to each other and useful in pedestrian and motorbike detection.

Pedestrian detection is an important part of intelligent transportation systems. In the literature, Histogram of Oriented Gradients (HOG) detector for pedestrian detection is known for its good performance, but there are still some false detections appearing in the cases with flat area or clustered background. To deal with these problems, in this research work we develop a new feature which is based on pairing comparison computations, called Comparison of Granules (CoG). The idea of CoG is to encode the textural information of local area describing how different the pixel intensities are distributed within a region. It is shown that the special characteristics of CoG feature are “small” and “efficiency” relative to HOG. By incorporating this new feature, we propose a HOG-CoG detector which through our validation experiment achieves 38% log-average miss rate in full image evaluation and 90% detection rate at 10−4 false positives per window on INRIA Person Dataset. Another contribution of this work is that, we also present a training scheme that can be applied on huge database for training a detector. Such training scheme can reduce the number of hard samples during bootstrap training.