When UAV moves,it often leads to trajectory interruption and appearance feature confusion.This paper proposes a multi-object tracking method for UAV based on feature separation.An adaptive Kalman filter is proposed to address the is-sue of motion trajectory interruption.The affine matrix of two consecutive images is calculated based on the image registra-tion algorithm,and the state variables are corrected in time before the prediction stage of Kalman filter,which makes the updating process smoother.To tackle the confusion of appearance features,an efficient multi-class target feature encoder is designed,employing multi-scale convolution streams and channel gating mechanisms to learn distinctive features for each target.By weighting motion costs and appearance costs,the Hungarian algorithm is utilized to solve the global matching re-lationship between trajectories and detected targets.Experimental validation on the widely used VisDrone2019-MOT multi-target tracking dataset shows that the proposed method improves multiple object tracking accuracy(MOTA)and identifica-tion F1 score(IDF1)by 8.1％and 11.2％,respectively,compared to the current best method,UAVMOT,confirming the effectiveness of the proposed approach.
维普
万方数据