Studies on a Video Surveillance System Designed for Deep Learning

This paper proposes a new video surveillance system designed for Deep Learning. The proposed system uses three steps to transfer RTSP streams to pictures for Deep Learning. First it decapsulates the streams, then decodes and converts color space & extracts frames. The proposed system has two ways to decode RTSP streams, hardware decoding and software decoding. By checking the processor's version of CPU firstly, system chooses a better way to decode. The proposed system has GPU and CPU. CPU is used to process RTSP streams, extract frames and do human-machine interaction. GPU is used for computing and analyzing the algorithms of Deep Learning. So the complex computing does not run on the CPU. The proposed system runs on Linux system and has Python interface, so it can easily connect with the models of Deep Learning. By running on multiple machines, the result shows that the proposed system can process up to 16 channels of stream. After 7*24 hours of testing on several machines, this system can run continuously without downtime and the delay time is less than 7 seconds.