Target tracking method based on interference detection

Considering the problems of similarity interference, partial occlusions, and changes in scale during target tracking, a tracking method based on interference detection is proposed, which an improvement over Siamese fully convolutional classification regression neural network (SiamCAR) approach. Under proposed framework, marginal distribution feature maps used to determine presence or absence interferents. When present scene, motion vector composed predicted value obtained through Kalman filter as basis for prediction. Experiments benchmark LaSOT dataset show that algorithm SiamCAR, introduces features, achieves best performance videos with similar object fast motion, small compared classical SiamCAR other excellent algorithms.