Pedestrian Tracking Algorithm Based on IRB-JDE

There are several problems in the existed pedestrian tracking algorithms, such as missing detection and false detection of small-size pedestrians and occluded pedestrians, frequent pedestrian identity jumps, and track loss. In this paper, an improved pedestrian tracking algorithm called IRB-JDE (Joint Detection and Embedding Improved by Res2NeXt50 and BYTE) is proposed. In specific, an improved backbone network, namely Res2NeXt50 (I_Res2NeXt50), is proposed by improving the residual network structure of ResNet50. The improved backbone network has developed the tracking accuracy of occluded pedestrians and reduced the number of pedestrian identity jumps. Secondly, a multi-scale fusion network is presented to get the location ability of small-size pedestrians better by integrating the PAN structure on the FPN structure. Finally, a data association algorithm named BYTE is introduced to improve the reliability of data association and decrease the quantity of track loss. Ablation experiments demonstrate the efficiency of the abovementioned improvements in increasing the accuracy of pedestrian tracking. In comparison with the existed algorithms, the IRB-JDE performs better on the MOT16 dataset according to the commonly-used evaluation indicators.