Combining the 2D and 3D world: A new approach for point cloud based object detection

Recently, dense 3D sensor data is used to perceive the environment of objects. Hence, perception systems take advantage of these sensors to detect arbitrary objects in the surrounding of the system. Due to their arbitrary position, size and reflection characteristics objects cannot be detected easily. Often over-clustering occurs which leads to wrong tracking results and failures in the environment model. This publication addresses the problem of over-clustering and object detection using a combination of 2D object pathway tracks and 3D clusters. Subsequently, the detected objects are tracked and the detection and tracking results are evaluated on a real data sequence. For evaluation the subpattern assignment (OSPA) metric is used to evaluate the object detection as well as the tracking results.