Fast Multiple Object Tracking Using Relevant Motion Vector

Abstract

Multiple object tracking is a crucial task in the field of computer vision. In conventional tracking algorithms, frequent detections are required to achieve a good tracking performance, which makes the process time consuming and unable to be applied in real-time applications. Since the adjacent frames are highly relevant and the relevant motion vector can be extracted directly from compressed videos without extra calculation, we present a fast tracking algorithm based on the relevant motion vector to reduce the detection frequency. In the proposed algorithm, the video is divided into key and non-key frames. For the key frames, the objects are detected on the RGB images based on detection method. For the non-key frames, the objects are tracked based on transformation information calculated on motion vector. In order to combine the detection results and the tracking results, data association is performed for the key frames based on Hungarian algorithm. Evaluations on a video dataset show that our proposed algorithm achieves better efficiency and comparable accuracy than the previous algorithm.