Autonomous vehicles need to accurately detect and localize other participants on the road. However, there is often not enough information in the scene to reach the level of accuracy required for safe operation.
Image credits: RAEng_Publications, free license via Pixabay
A recent paper published on arXiv.org examines the possibility of using potentially valuable information from previous traversals of the same route.
The researchers noted that this information reveals where pedestrians, cars and cyclists are in the scene and where signs or other background objects are constantly present throughout the traversal. proposed a simple and efficient way to take advantage of such information without any additional labels that can be incorporated into most modern 3D perception pipelines.
Evaluation confirms that the approach exhibits consistent and significant improvement, particularly on challenging cases (small, distant objects).
Self-driving cars must accurately detect vehicles, pedestrians and other traffic participants in order to operate safely. Small, distant, or highly hidden objects are particularly challenging because LiDAR point clouds contain limited information to detect them. To address this challenge, we take advantage of valuable information from the past: specifically, data gathered in past traversals of the same scene. We believe that these previous data, which are generally discarded, provide rich contextual information to elucidate the above challenging cases. To this end, we propose a novel, end-to-end trainable hindsight framework to extract this relevant information from previous traversals and store it in an easy-to-query data structure, which can be used for future 3D object detection. Can be leveraged to help with same scene. We show that this framework is compatible with most modern 3D detection architectures and can significantly improve their averaging accuracy over many autonomous driving datasets, particularly in challenging cases by more than 300%.
Research Paper: You, Y., “Hindsight 20/20: Leveraging past traversal to aid 3D perception”, 2022. Link: https://arxiv.org/abs/2203.11405
