TY - JOUR
T1 - A multi-object soldier tracking algorithm based on trajectory association and improved YOLOv8n
AU - You, Yu
AU - Wang, Jianzhong
AU - Bian, Shaobo
AU - Sun, Yong
AU - Yu, Zibo
AU - Wu, Weichao
N1 - Publisher Copyright:
© 2025 Elsevier Ltd
PY - 2025/8/1
Y1 - 2025/8/1
N2 - In response to the challenges encountered in soldier tracking, including imprecise detection, high computational load, and frequent switching of soldier IDs, we propose a trajectory association and improved YOLOv8n-based soldier tracking algorithm, termed soldier tracking algorithm with YOLOv8-SD and HybridSORT-ST (STA-YH). The algorithm consists of two stages: soldier detection and soldier tracking. In the soldier detection stage, we propose an Efficient Dynamic C2f (ED-C2f) backbone network specifically designed to efficiently capture soldier features. Then, a novel Multi-branched Slim Context and Spatial Feature Calibration Network (MSCSFCN) is constructed to effectively fuse and align multi-scale soldier features. Furthermore, Group-Sparse Dynamic Head (GSDH) network is used to improve the attention of model to the soldier detection area. In the soldier tracking stage, we introduce the OSNet_IBN reidentification network and Adaptive Fading Kalman Filter (AFKF) algorithm into the HybridSORT and improve the state vector of the filter to reduce the frequency of ID switching for tracked soldiers. The results indicate that, in terms of soldier detection, compared with the baseline YOLOv8n, the improved YOLOv8-SD improved precision by 4.18% and mAP50-95 by 4.94% under the same computational load. This means that YOLOv8-SD is more accurate and has fewer missed or false detections. For soldier tracking, compared with the baseline, HybridSORT-ST demonstrates a 25.88% increase in HOTA, a 36.81% improvement in MOTA, and a 13.69% rise in IDF1, significantly improving the stability of continuous tracking of soldier in complex battlefield environments with dense movement and frequent occlusion, while meeting the requirements of lightweight design and tracking accuracy.
AB - In response to the challenges encountered in soldier tracking, including imprecise detection, high computational load, and frequent switching of soldier IDs, we propose a trajectory association and improved YOLOv8n-based soldier tracking algorithm, termed soldier tracking algorithm with YOLOv8-SD and HybridSORT-ST (STA-YH). The algorithm consists of two stages: soldier detection and soldier tracking. In the soldier detection stage, we propose an Efficient Dynamic C2f (ED-C2f) backbone network specifically designed to efficiently capture soldier features. Then, a novel Multi-branched Slim Context and Spatial Feature Calibration Network (MSCSFCN) is constructed to effectively fuse and align multi-scale soldier features. Furthermore, Group-Sparse Dynamic Head (GSDH) network is used to improve the attention of model to the soldier detection area. In the soldier tracking stage, we introduce the OSNet_IBN reidentification network and Adaptive Fading Kalman Filter (AFKF) algorithm into the HybridSORT and improve the state vector of the filter to reduce the frequency of ID switching for tracked soldiers. The results indicate that, in terms of soldier detection, compared with the baseline YOLOv8n, the improved YOLOv8-SD improved precision by 4.18% and mAP50-95 by 4.94% under the same computational load. This means that YOLOv8-SD is more accurate and has fewer missed or false detections. For soldier tracking, compared with the baseline, HybridSORT-ST demonstrates a 25.88% increase in HOTA, a 36.81% improvement in MOTA, and a 13.69% rise in IDF1, significantly improving the stability of continuous tracking of soldier in complex battlefield environments with dense movement and frequent occlusion, while meeting the requirements of lightweight design and tracking accuracy.
KW - Deep learning
KW - HybridSORT
KW - Soldier tracking
KW - Trajectory association
UR - http://www.scopus.com/inward/record.url?scp=105004874241&partnerID=8YFLogxK
U2 - 10.1016/j.eswa.2025.127877
DO - 10.1016/j.eswa.2025.127877
M3 - Article
AN - SCOPUS:105004874241
SN - 0957-4174
VL - 285
JO - Expert Systems with Applications
JF - Expert Systems with Applications
M1 - 127877
ER -