Visual Tracking Algorithm Based On Feature Extraction And Classification Learning

With the development of computer vision technology,target tracking technology,as one of the research hotspots,has a wide range of applications in intelligent transportation,behavior recognition,medical diagnostic systems,robot vision and other fields.Although experts and scholars at home and abroad have contributed various algorithms to target tracking technology in recent years,due to the complex and changeable target tracking application scenarios,they still face many challenges,such as motion blur,illumination changes,and in-plane(Outer)Reversal,object occlusion,etc.These challenges make the tracking effect of the existing target tracking technology and the experts' expectations still have a certain distance.Therefore,how to improve the efficiency and stability of tracking technology is still a current research hotspot.In view of the constantly changing characteristics of the target and background in the target tracking process,this paper uses the feature transfer ability of TRL transfer characterization learning to construct a target tracking system from the perspective of improving the adaptability of the tracker to changing features.The main work carried out is as follows:1)The first work is to use the transfer representation learning in the discriminant-based target tracking system to construct the feature model of the target tracking system.The proposed feature model can use the transfer characterization learning to minimize the distribution distance of the target feature to perform feature transfer learning on the target and background image blocks that change over time,so that the target feature in the unknown frame and the target feature in the known frame are distributed The distance is minimized,so as to provide a stable and reliable target feature expression for the tracker's classification model.Moreover,unlike model migration,feature migration can be independent of the performance of the classifier,so that the target tracking system has a more sufficient choice space.In addition,this paper adopts a dynamic sample update strategy to further improve the robustness of the tracker.2)The second work is based on the proposed feature model,using the transfer representation learning algorithm based on the TSK fuzzy system to further improve the ability of the tracker feature model to express image features.The proposed method uses the feature learning ability of TRL-TSK-FS to provide the tracker with a compact shared feature space with knowledge transfer characteristics and good discrimination capabilities.In addition,in order to integrate the timeliness requirements of target tracking,this paper optimizes the setting of the construction of the maximum averaged difference between the target domain and the source domain features in TRL-TSK-FS.Then,Gaussian Naive Bayes classifier is used to quickly divide the target and background in the shared feature space.The method proposed in this paper is analyzed and verified experimentally on OTB2013 and OTB2015 video tracking data sets.The results show that the tracker proposed in this paper has significant advantages over existing algorithms.3)The third work is a target tracking system constructed by ELM-AE and transfer representation learning.On the basis of the second work,aiming at the problem that too high number of TSK-FS rules will lead to insufficient timeliness of the tracker,this paper uses ELM-AE's rapid learning ability for complex image features,and the feature model of the proposed target tracking system Make improvements.Then use the feature learning ability of TRL to improve the adaptability of the random feature space.And adopt a dynamic sample update strategy to improve the stability and effectiveness of the tracker.The algorithm has carried out a large number of experimental analysis and verification in the 11 target tracking challenge scenarios proposed by OTB.From the experimental results,it can be seen that the target tracking algorithm proposed in this paper has obvious advantages over existing target trackers.