Surface Multiple Object Tracking: an Accurate HAT-YOLOv8-ADT Tracking Model

<p dir="ltr">With the development of artificial intelligence tech-<br>nology, Autonomous aerial vehicles (AAV) have the ability to<br>sense the environment. Multiple object tracking (MOT) in AAV<br>video is a very important vision task with a wide variety of<br>applications. However, there are still many challenges in MOT<br>in AAV video. First, the movement of the onboard camera in the<br>three-dimensional (3D) direction during the tracking process, as<br>well as the unpredictable measurement noise characteristics of<br>AAVs flying at high speeds, can lead to significant deviations<br>in the prediction of the object’s position. Second, the appli-<br>cability of the traditional detection algorithm decreases when<br>the object is small and dense in the AAV viewpoint during<br>detection. Finally, the traditional intersection over union (IoU)<br>matching approach does not take into account the effects of<br>the height and width of the box, and the matching results are<br>inaccurate for the prediction and detection box. In order to<br>address these challenges, we recommend the adaptive DeepSort<br>(ADT) algorithm to reduce the prediction bias due to camera<br>movement and difficulty in predetermining measurement noise<br>characteristics, the hybrid attention transformer-YOLOv8 (HAT-<br>YOLOv8) algorithm to enhance the detection capability of tiny<br>objects, and the intersection over union of height and width<br>(HWIoU) matching algorithm, which improves the matching ac-<br>curacy and thus the tracking accuracy. Experimental results show<br>that our proposed solution outperforms the baseline solution.<br>It outperforms the current mainstream StrongSort in MOTA,<br>HOTA and IDF1 by 2.86%, 0.9% and 9.36%. Code repository<br>link: https://github.com/networkcommunication/.</p>