A Progressive Vehicle Search System for Video Surveillance Networks

Vehicle search aims to find a specific vehicle appeared in the physical world through surveillance networks. Existing systems usually exploit attributes, like colors and types, or license plate numbers as keywords to find vehicles in the database. However, attribute based search cannot accurately find the target vehicle due to the minor inter-class difference between similar vehicles and the extremely varied environmental factors. Moreover, the license plates usually cannot be correctly recognized in real-world surveillance scenes due to the viewpoints and motion blur. In this paper, we design a progressive vehicle search system, PVSS, for surveillance networks. As a search engine, PVSS contains three main parts: the vehicle crawler, the vehicle indexer, and the vehicle searcher. The vehicle crawler performs vehicle detection and tracking on surveillance videos and sends the cropped vehicle images, camera IDs, and timestamps to the indexer server. The indexer extracts multi-grained visual features, including the appearance features and license plate features, to index the vehicle images in the cameras. In the search stage, given a query vehicle image, the time range, and the spatial scope, the system can search the target vehicle in a progressive manner, i.e., from-coarse-to-fine search with multi-grained features and from-near-to-distant search with the spatiotemporal context. Extensive experiments on the public dataset from a real surveillance network demonstrate the effectiveness of the proposed system.