SIG Chair: Shao-Yi Chien (Professor, Dept. of Electrical Engineering, National Taiwan University)
Five sub-projects pertain to SIG Sensing:
The Aims and Objectives of SIG Sensing:
In SIG-Sensing, we focus on four major and important aspects of a sensing platform for augmented collective beings (ACB), including sensing, communication, security, and middleware. For these four aspects, four projects are designed and executed to develop high-potential key techniques. The project “Distributed Intelligent” focuses on the large-scale video sensing for ACB. The project “Next-Generation Camera Communications” focuses on the communication system with LED light sources as the transmitters and cameras as the receivers for transmitting digital information as well as positioning. The project “Enhancing Security and Privacy in Augmented Collective Beings” focuses on the analysis and defense techniques for the ACB systems. The goal is secure and privacy-preserving rule-based automation in ACB. The project “WuKong: middleware of ACB,” focuses on the middleware optimization for ACB devices.
After background survey and several studies with experiments, in the next year, the plans for these three projects are described as follows.
For Distributed Intelligence, an object based video summarization system and a quantization technique for deep neural network have been developed. In addition, a baseline re-identification based multiple object tracking system is also developed. In the future, we will continue two directions: object tracking with re-identification, and optimized CNN. For the re-identification sub-project, we plan to further improve the spatio-temporal model with the developed dataset, and also consider to re-train the neural network on the fly. For the optimized CNN, we will put our focus on binarized network and the associated hardware architecture design.
For Next-generation Camera Communications, through the development of several experimental prototypes, in the first year we have built confidence that the polarization of light can indeed be utilized to carry digital information and facilitate positioning, effectively addressing several unsolved issues in prior arts – most notably, the short operation range of a CamCom system. Our novel polarization intensity modulation allows the service range of the system to reach 40 meters, a significant improvement over any state-of-the-art. In the future, we are confident that through further development we will be able to optimize the developed techniques and improve the throughput. We also intend to investigate suitable use cases which can utilize the developed techniques in the indoor and outdoor, automotive contexts.
We also had some positive results in developing the techniques to identify a certain object with a polarized marker. Such a tag emits light with varying polarization waveform, which corresponds to the ID of the object. Our latest results have demonstrated that a small 3.6cm x 3.6cm marker can still be reliably detected with some camera movements. In the next few quarters, we aim to further reduce the computational complexity of the detection algorithm, such that it can be applied in real-time applications.
The project team also continues the effort in standardizing its Rolling Shutter – Frequency Shift Keying (RS-FSK) camera communication waveform (developed in the predecessor project) in the IEEE 802.15.7m standard committee, with the goal of generating wider impact in future commercial products. Furthermore, the team has also been communicating with several Taiwan-based LED or lighting equipment manufacturers to seek further collaboration opportunities in the near future.
For Enhancing Security and Privacy in Augmented Collective Beings, a simple FSM-based analysis module has been developed with a simple example. We are collecting more concrete and real use cases for further investigations. In the next year, with the collected use cases, we plan to develop the analysis module to identify the potential attacks against rule-based automation. Moreover, we plan to develop at least one technique for offline (ex. patching) or online defense (ex. runtime monitoring and filtering).
For WuKong: middleware of ACB, we developed techniques to reduce this code size overhead by 52% and performance overhead by 80%, resulting in an AOT compiler that produces code that is on average only 1.7 times slower and 2.1x larger than optimized native C. Many opportunities for future work remain. For the mark loops optimization, a good heuristic is needed to make a better decision on the number of registers to pin for the markloop optimization, and we can consider applying this optimization to other blocks that have a single point of entry and exit as well. A more general question is what the most suitable architecture and instruction set is for a VM on tiny devices. The VM not only provides a virtual run-time environment but also the safe and robust run-time environment for sensor nodes. However, the limited computation capacity on sensors node suffer the performance. Based on our approach on enhancing run-time performance, we will also extend the approach for safety and robustness on VMs for sensor nodes.
In the next year, the preliminary modules of these key techniques will be developed. We will further facilitate more interaction and cooperation between SIGs. Distributed Intelligence can provide a computing platform for SIG-Leaning. and All these four projects can support SIG-Interaction. (updated in Feb, 2017)