Performance Study of Deep-Learning Based Surveillance Systems

Abstract

Recent advances in deep learning have transformed video surveillance systems into authentic visual intelligence ecosystems capable of detecting and recognizing objects. Existing literature generally analyses the performance of surveillance in terms of detection accuracy and inference time, not the overall process. Moving towards Internet of things (IoT) trend, real-time monitoring has become a new paradigm for automation safety systems. This motivates us to investigate the holistic performance of deep-learning based surveillance systems in an embedded platform. Specifically, this paper presents a real-time surveillance system which encodes video clips only when human is detected, and subsequently pushes notifications for video retrieval. The implementation runs on Raspberry Pi Zero W (RPZW) with a pre-trained network of single shot detector (SSD) algorithm and a Mobile Net architecture. We further accelerate the hardware performance by utilizing Movidius Neural Compute Stick (NCS). Results show that RPZW with NCS can process live video stream at a rate of 3frames per second (fps).Experiment evaluations also reveal that how different video codecs impact the video quality and notification time. These findings offer insights into the quality of footage in terms of potential crime evidence and emergency response time for an IoT surveillance system.