Detection-Driven Gaussian Mixture Probability Hypothesis Density Multi-Target Tracker for Airborne Infrared Platforms

Recent advancements in the unmanned aerial vehicle remote sensing field have highlighted the effectiveness of infrared sensors in detecting and tracking time-sensitive ground targets, particularly within the domain of early warning and surveillance. However, the limitations inherent in airborne infrared platforms can lead to irregular imaging and inadequate textural features. This study presents a multi-object tracking system specifically designed for weak-textured infrared targets, aimed at enhancing detection accuracy and tracking stability. Initially, improvements are made to the YOLOv10 model through the incorporation of modules such as DSA, c2f_fasterblock, and NMSFree, which collectively enhance detection accuracy and robustness for weak-textured targets. Subsequently, the detection results are employed in conjunction with GM-PHD tracking, enabling rapid and stable target tracking. The proposed methodology demonstrates a 2.3% improvement in detection accuracy and a 3.8% increase in recall when assessed using publicly available infrared tracking datasets. Notably, the key tracking metric, MOTA, achieves a value of 90.7%, while the IDF1 score reaches 94.6%. The findings from the experiments indicate that the proposed algorithm surpasses current methodologies regarding effectiveness, accuracy, and robustness in the context of infrared multi-target tracking tasks, thereby meeting the requirements associated with airborne infrared target tracking tasks.