Deep Learning-based Vehicle Re-identification Using Temporal Information In Urban Traffic

With the rapid development of navigation, communication, and sensing technologies, Unmanned Aerial Vehicles(UAV) based vision applications have grown increasingly popular. As a full spatial and temporal coverage of a large network is often infeasible with a limited number of UAVs. Therefore, re-identifying a vehicle from one video frame to another becomes important for various traffic estimations, including continuous trajectory extraction. Vehicle re-identification(ReID) works as a key technology since it aims at localizing and tracking the queried targeted vehicle from a large volume of vehicles. However, two major challenges exist in the previous UAV-based vehicle ReID studies. The first challenge is the limited perception of the vehicle from the Bird's Eye View(BEV) environment. The drone's camera captures only the top view of the vehicle, losing uniquely identifiable information. This leads to low image-to-image matching accuracy. The second challenge is highly dependent on the appearance-based method. To address these challenges, this study proposes a deep learning-based vehicle re-identification method which combines both visual and temporal information. Specifically, this method employs the Strong Baseline method, an advanced technique to extract the latent feature vector that embeds the appearance information of the vehicle. Then the predicted arrival time of the vehicle is used to further improve the matching accuracy. Our results show that this method notably improves vehicle ReID accuracy in the BEV environment, outperforming existing models. The suggested method can be applied to complement the vehicle tracking, merge the separated trajectories, and reduce the cost of data collection in the limited monitoring area.