Pedestrian Multi-Object Tracking Algorithm Based on Fish Eye Camera Distortion Correction

This paper proposes a multi-object tracking algorithm for distorted video images captured by a fisheye camera. The method addresses the issue of residual shortening distortion in calibrated pedestrian images. It utilizes image data augmentation techniques to create a pedestrian image dataset with calibrated distortion features, thereby improving the network’s recognition ability for such erroneous images. Experimental results demonstrate that the network trained using this dataset exhibits a 1.37% improvement in accuracy, a 1.42% reduction in missed detection rate, and a 0.95% reduction in false detection rate compared to other algorithms in recognizing fisheye camera distortion features. Additionally, based on YoloV5, the decoupled head structure is redesigned using the Hybrid Channels strategy, and the CSPStackRep Block is used as the backbone network. Combining the laboratory’s domain-controlled hardware conditions, a hardware-friendly strategy is employed to optimize and adjust the network structure, facilitating hardware deployment and achieving higher efficiency and lower latency. Experimental results show that compared to other algorithms, this approach can improve MOTA by 1.47% and MOTP by 0.51% under a fisheye camera while maintaining the same frame rate, thereby enhancing the tracking performance.