PTNet: an Improved Algorithm for Object Tracking

Pedestrian tracking, as a sub-task of object tracking, is widely applied in smart surveillance. However, the accuracy of object re-identification significantly decreases when objects encounter prolonged occlusion or undergo significant appearance changes. To address this issue, PTNet, a regression-based object tracking framework, is proposed in this paper. Firstly, a combination of DeepSORT and FastReid models is employed to resolve the problem of target re-identification after being occluded for a period of time. Secondly, a self-attention-based feature extraction model named BoTNet, incorporating the Global Multi-Head Self-Attention (MHSA), is integrated into the backbone network to capture global and long-distance features of targets in surveillance videos. Through ablation experiments, it was found that PTNet enhances accuracy by 4.8% compared to the benchmark. When compared to state-of-the-art object detection algorithms, PTNet demonstrates a significant improvement in performance.