Multi-Camera Vehicle Tracking with Powerful Visual Features and Spatial-Temporal Cue

Vehicle re-identification and multi-camera multi-object vehicle tracking are important components in the field of intelligent traffic, which is attracting more and more attention. In the NVIDIA AI City Challenge, we propose our solutions to solve these issues. In Track1 task, clustering loss and trajectory consistent loss are introduced into the vehicle re-identification training framework to train more suitable trajectory-based features for the clustering task. Besides, spatial-temporal cue is fully excavated to make up the deficiency of appearance feature and constrained hierarchical clustering is introduced into the pipeline to get the final cluster results. In Track2 task, we propose an effective vehicle training framework and trajectory-based weighted ranking method, which greatly improve the performance. Furthermore, an efficient way to mining the additional data to train more robust features is proposed to enlarge the training data. Finally, our algorithm achieves the state-of-theart performance in the competition.