Task-attentive LiDAR Filter: Adaptive ROI and Semantic-based Ground Segmentation

Region-of-interest (ROI) and ground point filter is a vital pre-processing step that 1). reduces computation and memory complexity of following perception tasks, and 2). identifies traversable areas and obstacles above the ground. This paper proposes a task-attentive LiDAR filter that combines an adaptive ROI filter and a semantic-based ground segmentation filter. The proposed adaptive ROI filter dynamically estimates a ROI area that ensures driving safety while removing many unrelated points. The proposed semantic-based ground segmentation filter combines semantic prior from HD map and LiDAR scans to accurately estimate ground planes in complex ground environments. This enables the perception system to detect small obstacles near the ground surface. Experiments are conducted on real-world data collected by the WATonoBus projects in challenging urban environments, where the proposed task-attentive filter reduces data points by 5.6 times and decreases the runtime by 5.6 times of clustering-based object detection algorithm.