Real-time 3D multiple occluded object detection and tracking

Abstract

This paper describes a real-time object detection and tracking method which is improved to occluded measurements. 3D point cloud data describes surrounding objects on one measurement frame. If two objects cross in front of an observing system, equipped 3D LIDAR on a vehicle, the object recognition mechanism has inherent problems. In this paper, we define that the system process consists of a detection part and a tracking part. The first problem occurs on the clustering process in the detection part. Two objects are liked as one object by a neighbor point approach. In our case, RBNN (Radial Bounded Nearest Neighbor) clustering method works with two parameters, radial bounded range and point number. The second problem happens on the tracking process. The previous detection process hands on wrong the object number. The tracking system also shows the ambiguity results with the fault tracking. It is frequently happened on occluded measurements. We proposed the enhanced object detection and tracking method using an object class and relative position. In this paper, we propose the system which has an enhanced tracking mechanism for approaching successful object tracking which prevents wrong detection and tracking. In experiment, we implement two humans and an autonomous ground vehicle to achieve a leader-following system.