This article describes how due to the diversification of electronic equipment in public security forensics, vehicle surveillance video as a burgeoning way attracts us attention. The vehicle surveillance videos contain useful evidence, and video retrieval can help us find evidence contained in them. In order to get the evidence videos accurately and effectively, a convolution neural network (CNN) is widely applied to improve performance in surveillance video retrieval. In this article, it is proposed that a vehicle surveillance video retrieval method with deep feature derived from CNN and with iterative quantization (ITQ) encoding, when given any frame of a video, it can generate a short video which can be applied to public security forensics. Experiments show that the retrieved video can describe the video content before and after entering the keyframe directly and efficiently, and the final short video for an accident scene in the surveillance video can be regarded as forensic evidence.
A Conceptual Model for Surveillance Video Content and Event-Based Indexing and Retrieval
