Vehicle Re-Identification for Video-based Surveillance

Abstract

Due to the advancement of means of conveyance and recent rise in traffic volumes, there has risen a need for automated surveillance systems to keep traffic scenarios under control. Vehicle detection and re-identification holds immense importance in tracking the traffic and maintaining law and order situation. However, several modern techniques involve installation of a system (like GPS) or a gadget (like trackers) within the vehicle. This can be simplified by monitoring vehicles through the video footages received from surveillance cameras. Conventional methods of such monitoring are performed manually and incorporate time and labor. This process can be automated using Computer Vision techniques. We propose one such system. Detection of vehicles from frames of videos is done through deep learning based solution. Once detected, vehicles are stored for re-identification to be performed on them, later. In the re-identification module, features of query image are extracted and compared to those of detected and stored ones. The vehicles with top 5 similarity indices are shown as results and potential matches of the query image, hence a solution to the re-identification problem. Re-identification is performed through ResNet and Siamese Neural Netwrok. This kind of framework can be improved by combining it with license plate recognition systems to achieve even better re-identification.