Towards Improving Car Point-Cloud Tracking Via Detection Updates

Most autonomous driving applications leverage RGB images representing the surrounding environment that contain useful appearance features but with a cost in terms of geometric features. On the other side, 3D point clouds generated by LIDAR sensors can provide more geometric 3D information with high accuracy and robustness but with a loss on appearance features. Regardless of the adopted technology, object tracking in autonomous driving scenarios suffers from the so-called error drift in detecting objects over time/frames. This work investigates the car tracking problem in an urban scenario, leveraging 3D point clouds. In particular, we have set our goal to mitigate the typical error drift that characterizes the classic tracking algorithm and, to this aim, proposed a system able to reduce the drift error by detection. An extensive experimental evaluation on the KITTI dataset shows the improvement in our solution's performance compared to state-of-the-art approaches.