Autonomous driving systems have benefited from recent progress in computer vision and deep learning. In particular, recent object tracking algorithms have used deep learning in various ways to improve tracking performance. However, several challenges related to autonomous driving systems and object tracking remain, and popular evaluation criteria do not always incentivize the development of object trackers that perform well in autonomous driving systems. This work first identifies several challenges related to object tracking in autonomous driving systems. Next, three different types of object trackers are evaluated in a series of experiments involving an autonomous vehicle in a simulated environment. As part of this evaluation, a framework is described and provided for others to be able to reproduce the results of the experiments and integrate, deploy, and evaluate alternative object trackers. An analysis of the results demonstrates that no single type of object tracker performs best in all driving scenarios. Finally, future research directions regarding object tracking and autonomous driving systems are proposed.




Download Full History