Vehicle Re-Identification and Tracking Based on Video Segmentation

Traffic object perception based on cameras is one of the foundations of Intelligent Transportation Systems. In traditional computer vision field, we usually take object detection method to detect and track the vehicle objects using bounding boxes with fixed shape, and some efficient methods based on this such as DeepSORT are used for perception. However, under the situation of dense traffic, vehicles could block each other in the viewpoint of the roadside camera, which severely reduce the accuracy of detection and tracking. Aiming to solve this, we propose our detection and tracking method based on partial feature re-identification and mask segmentation. First we apply segmentation method to separate the pixel-level image of each vehicle, then we use the especially trained CNN-based feature extractor to get the key information from the misshapen images, and finally utilize the masks and the features to track the vehicles. We test our method on CityFlow dataset and prove the validity of our method by visible result. We finally discuss the weakness of our framework and putting forward the future improvement direction of the algorithm.