Improved FairMOT for multi-pedestrian tracking in complex environments

This paper proposes a multi-person tracking algorithm based on improved YOLOX and Kalman filter to address challengs such as occlusion, lighting changes, and nonlinear movements in complex environments. The FairMOT algorithm is used as a baseline and improved the detection accuracy by introducing an enhanced version of YOLOX as the detector. The dual attention mechanism is embedded in the backbone network of YOLOX to enhance the detection accuracy. The features extracted from the YOLOX backbone feature network are fused at multiple scales to effectively reduce the missed detection rate of small and blurry people. The Unscented Transform is integrated into the Kalman filter to convert nonlinear pedestrian movements into linear ones for prediction, thereby reducing issues such as interruption of trajectories and frequent ID switching during tracking. The algorithm is validated on the MOT16 and MOT17 public dataset, the experimental results indicate that the improved algorithm shows performance improvements on both the MOT16 and MOT17 datasets. Specifically, on the MOT16 dataset, compared to the traditional FairMOT algorithm, the improved algorithm achieves a 2.4% increase in MOTA, a 1.5% increase in IDF1, and a reduction in ID switching from 1069 to 967, effectively improving the accuracy of multi-person tracking.