Improved Radar Data Clustering for Extended Target Tracking with Camera Fusion and Micro-Doppler Effect

Abstract

High-resolution automotive radars widely used nowadays yield multiple measurements per frame from a single target, which contrasts with the point target assumption in most tracking algorithms. Clustering radar detections based on their sources is a crucial step for subsequent tracking. It will significantly impact the environmental perception capability of autonomous driving systems, which in turn affects their overall performance. However, accurate clustering is challenging in complex automotive driving scenarios with miss detections, false alarms and clutter. Many clustering methods have been proposed for this topic in the literature.

In this thesis, an improved DBSCAN with range-adaptive extended criteria is first presented to reduce clustering error. Based on this, we further developed a camera-assisted clustering method to enhance the clustering process with target class and bounding box information from camera data. This method also addresses several issues in radar-camera fusion, including object occlusion and partial visibility, uncertainty in radar measurements, overlapping bounding boxes, and coverage discrepancy between radar and camera.

In addition, the effect of micro-Doppler motion has been ignored in most Doppler-assisted clustering algorithms, which may introduce significant splitting errors. In this thesis, the distribution of micro-Doppler measurements is formulated to predict the possible measurements from the tires. A two-step iterative clustering approach is first proposed to reduce the splitting error caused by the micro-Doppler effect in Doppler-assisted clustering, but its computational time is significant. Therefore, another clustering approach with tire position and range rate probability constraint is developed accordingly, which improves clustering accuracy and significantly reduces computational time compared with the two-step iterative method.