Automated Pedestrian Tracking Based on Improved ByteTrack

In order to augment the robustness of pedestrian tracking in video sequences, we offer an enhanced automatic pedestrian tracking method that is based on the ByteTrack framework. The objective of the proposed approach is to tackle the issue of missed detections and trajectory loss in pedestrian tracking due to dense occlusion. The achievement of multi-object pedestrian tracking is realized through the integration of YOLOX-CF, an enhanced iteration of YOLOX, in conjunction with the BYTE tracking approach. In order to improve the ability of the network to detect pedestrians in various places, we have incorporated the coordinate attention (CA) module into the feature extraction network of YOLOX. In addition, we want to tackle the complex issue of crowd occlusion in pedestrian objects by proposing the utilization of focus loss as a confidence loss function. The above function aims to achieve weight balance between positive and negative samples, hence enhancing the network's attention on problematic samples. The experimental results obtained from the MOT17 dataset demonstrate a notable enhancement in both the mean Average Precision (mAP) and Multiple Object Tracking Accuracy (MOTA) as compared to the first approach. We observe a notable enhancement of 3.1 percentage points in mAP and 3.4 percentage points in MOTA. Furthermore, with the transformation of the model into TensorRT, the rate of inference improves to 126 frames per second (FPS) when executed on a single 2080Ti GPU. The proposed methodology offers enhanced efficacy in real-time pedestrian tracking within the context of autonomous driving, beyond the capabilities of the original.