Multi-object Tracking Algorithm Based on Improved CStrack Association Strategy
To address the problems of obvious changes in object appearance,irregular movements that lead to trajectory interruptions and frequent identity switching in complex scenes,the tracker is improved in terms of Re-Identification(Re-ID)features,data association and interpolation,and a multi-object tracking algorithm based on an improved CStrack association strategy is proposed.Firstly,an appearance feature update module is used to reduce the impact of drastic feature changes due to viewpoint changes and object movement,and to enhance the correlation between the features.Secondly,a two-stage association method is proposed,which uses different metrics for secondary association according to the characteristics of high-and low-confidence detection results;in the first association,IoU distance fusion appearance features are used as the cost matrix of matching,and in the second one,extended IoU association is used to alleviate the problems of biased object motion estimation and indistinguishable appearance leading to metric failure;the Gaussian regression algorithm is used to consider motion information to compensate for the missed detection by interpolation.Finally,the test on MOT17 and MOT20 datasets shows that the tracking accuracy is 73.9％and 64.2％,respectively,which indicates that the method has obvious advantages in tracking accuracy and can better adapt to complex scenes.
万方数据
维普
