Real-time Multi-Object Tracking Algorithm for UAV Based on Deep Learning

Unmanned Aerial Vehicles (UAVs) with multi-target tracking technology are of significant importance in various key areas, including the management of modern urban life and military strategic deployment. This paper proposes an advanced Multi-Object Tracking (MOT) method with the combination of TStsm-yolov10 and an improved version of ByteTrack, which is well-suited to real-time multi-target tracking tasks for UAVs. In the realm of object detection, this paper introduces a tiny detection head to replace the existing large detection head, aiming to enhance the detection capabilities for small objects while constraining the parameter count. In the neck section, utilizing an improved TFPN for feature fusion across three distinct scales aims to improve the overall feature representation capability. Additionally, the SCSA attention mechanism has been integrated into the C2f module, designated as C2f-SCSA. The newly designed small object detection model is named TStsm-yolov10. For UAVs tracking, this paper enhances the Kalman filter by incorporating an adaptive noise estimation method to alleviate the impact of detection results. Experiments conducted on the VisDrone2019 dataset demonstrate that the improved methodology improves the detection performance, with a 5.7%, 5.6%, and 6.9% improvement in precision, recall, and mAP50, respectively. In terms of tracking performance, experiments on VisDrone2019-MOT-test and VisDrone2019-MOT-train demonstrated excellent tracking performance of the proposed method, with HOTA, MOTA and IDF1 improvements of 4.083% and 4.013%, 3.205% and 5.001%, and 5.221% and 5.988%, respectively. Furthermore, in comparison with the current popular two-stage tracking algorithms such as StrongSort and OCSort, the proposed method not only demonstrates excellent performance but also ensures real-time performance.