Retrofitting Video Surveillance Systems using Deep Learning Technologies

Steps toward the development of an embedded system that is able to enhance the capabilities of a video surveillance system are described within the proposed paper. The hardware support for our solution is based on the newly introduced Google Coral development board having, as a prominent feature, the on-board Edge Tensor Processing Unit, capable of execute state-of-the-art mobile vision models such as MobileNet v2 at 100+ FPS, in a power efficient manner. The Google Coral is proven to outperform competing Machine Learning hardware accelerators currently available on the market and so represents the best accuracy/power efficiency choice. The software part includes some deep learning-based modules for video analytics such as detection, classification and tracking objects, including cars and people, using MobileNet + SSD architecture. Also, the experimental setup is described: compiling OpenCV for Google Coral, connecting to the video stream of the surveillance cameras. The system supports video equipment by any manufacturer by interfacing to it using Real Time Streaming Protocol/Open Network Video Interface Forum connections.