Detection Level and Target Level Road User Classification with Radar Point Cloud

This paper examines the classification of vulnerable road users and vehicles using radar point-cloud data at two tracking levels. While automotive radar offers numerous capabilities suitable for real-world tracking scenarios, it faces challenges due to sparse detection. Consequently, a clustering algorithm could fail to detect a target's cluster due to a limited number of detections. Moreover, it can divide a single object cluster into multiple clusters or merge distinct entities into a single cluster, impacting the accuracy of target-level (cluster-level) classification. On the other hand, the detection-level classification assigns class labels to individual radar detections, which can also be employed without clustering but requires a higher computational complexity. By extracting different levels of information from radar data, this study analyzes detection-level and target-level classification using various combinations of extracted features and classifiers. The performance of the proposed approaches in terms of classification accuracy and computation time for testing is evaluated using a publicly available dataset for automotive radar applications.