Multiple Pedestrian Tracking with Improved Object Detection and Adaptive Thresholding Technique

The current state-of-the-art Multi-Object Tracking (MOT) is achieved through the tracking-by-detection paradigm. However, this approach faces two significant challenges: heavy reliance on object detection and the effectiveness of object association. To address these issues, our study explores two key aspects. Firstly, we enhance the object detection network by integrating an attention module and implementing a refined loss function. This enhancement aims to improve the accuracy and robustness of object detection. Secondly, we investigate the incorporation of the adaptive thresholding technique during the object association stage. By dynamically adjusting the threshold, we aim to improve the accuracy of associating objects across frames. We explore Multiple Pedestrian Tracking (MPT), a primary research direction within MOT, using a subset of the VisDrone-MOT. Our proposed method outperforms the baseline by demonstrating superior MOTA and IDFI. Furthermore, we conduct a comparative analysis to highlight the effectiveness of the adaptive thresholding technique compared to a fixed thresholding technique.