Revisiting Standard-Definition Map Based Motion Prediction in the Era of End-to-End Autonomous Driving

Zhiyuan Zhang, Xiaosong Jia, Qifeng Li, Guanyu Chen, Junchi Yan

Keywords: motion prediction, standard-definition maps

Abstract: Most motion prediction models use maps as environmental context. For a long time, high-definition (HD) maps are preferred as they provide detailed lane-level information and often lead to significantly better performance compared with standard-definition (SD) maps. However, offline HD maps require extensive manual annotation, making them costly and unscalable. Online mapping-based methods still require HD map annotation to train the online mapping module, which is costly as well and may suffer from the issue of out-of-distribution map elements. In this paper, we look back to SD maps in the era of end-to-end autonomous driving and focus on narrowing the performance gap between HD and SD maps. We initially extend anchor-based and anchor-free motion prediction models in an end-to-end manner and find the performance gap narrowed with the introduction of raw image features. Furthermore, we discover the unique challenges that the coarse and misaligned SD maps bring to feature fusion of the anchor-free model and on anchor generation of the anchor-based model. Thus, we design two novel modules named Enhanced Road Observation and Pseudo Lane Expansion to address these issues. With these insights, we reduce the performance gap between HD and SD maps by 84%, making SD map based motion prediction achieve comparable performance as HD map based one.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 6897