| Title |
Eva: A Distributed Optimization Architecture For Efficient Video Analytics For Smart Cities |
| ID_Doc |
24532 |
| Authors |
Li Z.; Huang Y.; Fu X. |
| Year |
2023 |
| Published |
ACM MobiCom 2023 - Proceedings of the 18th Workshop on Mobility in the Evolving, Internet Architecture, MobiArch 2023 |
| DOI |
http://dx.doi.org/10.1145/3615587.3615990 |
| Abstract |
Intelligent video analytics is resource-intensive, e.g. computation and bandwidth. In this paper, we improve resource efficiency for intelligent video analytics via proposing a distributed architecture named EVA. We noticed that timeliness of anomaly detection is crucial in many video analytics applications along with minimal resource utilization. Therefore, we study the overhead of profiling techniques which find optimal configurations for deep learning during video analytics to achieve good resource-accuracy trade-offs. To minimize this overhead, we propose 2-stage deep learning to build efficient and robust models, and two rate adaption algorithms: Adaptive Frame Rate (AFR) and Adaptive Resolutions (ARR) to achieve good resource-accuracy tradeoff. Evaluations using UCF-Crime dataset show that EVA considerably outperforms baseline Awstream with 71% improvements in resource consumption, 69% in CPU utilization whilst reducing the bandwidth consumption by 70% and storage by 61%. © 2023 ACM. |
| Author Keywords |
deep learning; neural networks; video analytics |
