Poster: Improved DBSCAN Using Modified Epsilon for Static Scene Clustering

Radar is an emerging and attractive sensing modality for automotive applications due to its relatively low cost, robustness in the face of inclement weather, and ability to detect both position and velocity of the target(s). Individual radar observations are grouped with clustering algorithms to identify objects, traditionally dynamic actors in a scene like other vehicles. Traditional clustering methods perform poorly when applied to static point cloud data from automotive radars. DBSCAN is the leading clustering algorithm and was used as the basis for this research. We considered the physical characteristics of the scenes expected from a roadway to improve the clustering performance of DBSCAN. By modifying the shape of the DBSCAN neighbor search, we look for objects that follow the direction of the road, improving clustering of critical features like road edges. Based on our experimental results, this algorithm outperforms DBSCAN clustering of road scenes by as much as 30 %.