Abstract
In this paper, a real-time 3D multi-object tracking by using bounding box and depth information has been proposed. Multiple object tracking (MOT) is a key component in the robot industry and autonomous driving. In recent years, many kinds of research focus on how to use a learning-based method to extract feature images and use the features to track the objects. However, the common detected-objects in industrial applications are pretty much the same in appearance and color. In other words, objects having a similar feature have the tendency in making the learning-based modal malfunction. To resolve this problem, the work presented in this paper proposes a new approach using a 2D bounding box of pixel and 1D depth information as input. First, the bounding box from the detector and depth information from the depth sensor can be obtained. Secondly, a combination of the Kalman filter and the Hungarian algorithm is used for state estimation and data association. Lastly, the results are evaluated and compared with others by using KITTI benchmarks. Furthermore, it is found through experiments that in pragmatic tracking of objects, the proposed method and system runs at a rate of 1153.4 FPS, offering a very fast speed among all MOT methods with its MOT accuracy (MOTA) almost the same as others 3D MOT algorithms.
