| Target tracking technology plays an important role in the construction of smart factory and security system,but the changes of illumination,background clutter,target occlusion and its own scale will affect the accuracy of tracking.Therefore,the research on real-time and accurate target tracking method in complex scenes has important academic research significance and practical engineering application value.Kemelized correlation filter(KCF),KCF algorithm is the most representative real-time traditional target tracking algorithm at present,but there are two problems in this algorithm.One is the lack of target tracking loss judgment mechanism.After the target leaves the field of view or is lost,the tracker continues to track the wrong target unconsciously,which leads to unstable tracking;The second is to update the template linearly with the target features of the current frame.When the target features of the current frame are mixed with the background information,the template will drift to a certain extent,resulting in subsequent tracking errors.In order to solve the problem of tracking instability in KCF algorithm,this paper proposes a target loss judgment mechanism based on triple network,which uses the target detection result of Faster R-CNN detector,the similarity between KCF tracking result and template to confirm the tracking target.This mechanism combines with the detector to improve the stability of the tracker.In order to solve the problem of target template drift in KCF algorithm,this paper proposes a target template updating algorithm based on incremental learning.Incremental kernel principal component analysis(IKPCA)is used to extract salient features from the historical features of the target,and reduced set(RS)is used to extract salient features,extension technology compresses the target features after incremental learning to achieve constant speed template updating.The algorithm not only uses the historical salient features of the target to prevent template drift,but also achieves the invariance of template update speed.In order to verify the effectiveness of the proposed method,extensive experiments are carried out on the open target tracking dataset TC-128 and OTB.The experimental results show that the average success rate and accuracy of the proposed method are 94.8%and 97.3%respectively on the four sequences selected from TC-128 dataset.Compared with the classical KCF tracking method,the success rate and accuracy of the proposed method are improved by 49.0%and 40.1%respectively;Experimental results on 11 challenge factors of OTB dataset show that the average success rate and average accuracy of the proposed algorithm are improved by 25.7%and 26.0%respectively compared with KCF algorithm.Compared with the advanced tracking algorithms such as TLD,DCF,OCT-KCF and SiameseFC in recent years,this method has also achieved competitive results.The experimental results show that the improved KCF target tracking algorithm has high tracking accuracy and success rate.This method can be applied in smart factory and security system. |
