• computer vision
• deep learning
• detection
• neural network
• tracking
• vehicle dataset
• vehicle re-identification

Advanced on deep learning research and the availability of a lot of data to be trained, thanks to the growth of the internet, has allowed progress in many fields of computer vision, such as object detection, object tracking, and object re-identification.

Tracking vehicles over multiple cameras placed at different positions is not a single task, but the composition of three distinct research problems: detection, single-camera-tracking, and re-identification.
In this thesis work, we realize a system capable of tracking and re-identifying the same vehicle from different cameras using state-of-the-art approaches for detection, tracking, and re-identification.

A new vehicle re-identification baseline, V-ReID-KTP-Baseline, that exploits the use of vehicle keypoints, traklets, and license plate information for re-identification, is deployed. In particular, a new re-ranking method based on license plate information is designed specifically for this task. We also present a new labeled dataset, V-ReID-AB-Dataset, created and employed to test the use of license plate information for vehicle re-identification.
Test on this new dataset suggests that the availability of license plate information can make a considerable improvement in results for the task of vehicle re-identification.