Research on Anti-Occlusion Tracking Method for Underground Pedestrian Tracking Based on Adaptive Link Optimization

To address the issue of inaccurate trajectory matching for pedestrians due to frequent occlusions in underground coal mine environments, traditional appearance-based object tracking methods often suffer from misalignment due to occlusions and appearance confusions. This paper proposes an adaptive linkage optimization method for occlusion-resistant tracking of pedestrians in underground settings. First, we introduce nonlinear dynamic features to optimize the trajectory linking method, enhancing the feature processing capabilities of the trajectory linking module in the matching cascade. Based on this, we construct a trajectory linking optimization module using the CPCA attention mechanism. Nonlinear dynamic features are processed through temporal blocks to further improve the accuracy of temporal information capture. Next, we combine the similarity scores calculated by the trajectory linking optimization module with the trajectory state estimates predicted by the Kalman filter, extending the original total cost function to address the inaccuracies in trajectory matching caused by low distinguishability of appearance features. Subsequently, by calculating the variation in the breakage rate and ID switch rate during the tracking process, we introduce an RB factor to define weights and compute the total adjustment amount to dynamically adjust the IOU threshold in the matching decision process, adapting to the tracking interruptions caused by the presence of pedestrian clusters. Experimental results demonstrate that on three types of video sequences with typical underground features, the adaptive linkage optimization method achieves average metrics for tracking coal mine pedestrians of MOTA 76.17%, MOTP 84.13%, and IDF1 74.9%. Compared to the full-scale feature learning model OSNet, the key image feature extraction network YOLOv7-SAM, the fully correlated long-term tracker FuCoLoT, and the convolutional network model DeepSORT, the proposed method improves the averages by 5.75%, 0.77%, and 4.23%, respectively. These results indicate that the proposed algorithm outperforms existing algorithms in validation metrics and effectively resolves the issue of tracking mismatches for targets in occluded scenarios in coal mines.