A Research Of Visual Tracking Algorithm Based On Deep Learning

Visual tracking is an important research direction in the field of computer vision,and it has a wide range of applications in areas such as automatic driving,military reconnaissance,video surveillance,human-computer interaction,and unmanned aerial vehicle.The main purpose of visual tracking is to predict the position,scale,motion state and other information of the target in the subsequent video sequence according to the given initial target information in the video sequence.In recent years,more and more researchers have been engaged in the research of tracking algorithms.Various tracking algorithms have been improving in tracking performance,but the tracking performance in complex scenes such as target occlusion,lighting changes,fast motion,and target deformation is still insufficient,so more in-depth research on tracking algorithms is still very important.At present,deep learning technology is widely used in visual tracking.In the tracking algorithm based on correlation filtering,the pre-trained deep network model is used to extract the features of the target to enhance the ability of the tracker to represent the target model.In the tracking algorithm based on the siamese network,the whole framework is built by the deep network,and the end-to-end design is realized.Combined with deep learning technology,visual tracking algorithm has made a series of progress,but in complex scenes,such as target deformation,rotation,illumination change,scale change,etc.,it is easy to lead to tracking failure.How to perform appropriate model updates and build more robust target model is a problem that needs to be solved.Aiming at the above problems,this paper carries out research from two aspects of model update and feature enhancement under the framework of correlation filtering and siamese network.The main innovations and achievements of this paper are as follows:(1)Aiming at the problem that the tracking algorithm based on siamese network lacks online updating of the model,an online learning based siamese network tracking algorithm is proposed.Firstly,the target in the first frame is regarded as the static template,and the dynamic template is obtained by using the high confidence updating strategy in the subsequent frame.Then,in the online tracking process,the fast transform learning model is used to learn the apparent change of the target from the double template,and the likelihood probability map of the target in the search area is calculated according to the color histogram features of the current frame,which is fused with the depth features to carry out background suppression learning.Finally,the response graphs obtained by the two templates were weighted and fused to obtain the final tracking results.Experimental results on OTB100 and TC128 data sets show that the online updating of the model can effectively improve the algorithm performance.(2)Aiming at the problem that the tracking algorithm based on siamese network does not express the feature of the target in the sence of target deformation and scale change,a siamese network tracking algorithm based on asymmetric convolution and response graph confidence is proposed.Firstly,the square convolution in the original siamese network is replaced by d×1,1×d and d×d asymmetric convolution,which is used to enrich the feature space of the target.Then,at the last layer of the network,three sets of asymmetric convolution kernels are added in parallel to obtain three response graphs generated by three sets of target features.Finally,the confidence of each response graph is calculated respectively,and the weight is allocated adaptively for fusion,so as to select the optimal target prediction location.Experimental results on OTB100 data set show that the introduction of asymmetric convolution module can effectively enhance the feature representation ability of siamese networks.(3)Aiming at the problems of target rotation and deformation that are easy to occur in UAV video scenes,this paper proposes a cognitive features based on target and from coarse to fine search UAV target tracking algorithm: Under the framework of correlation filtering,first use VGG19 network for feature extraction in the process of the training,the network of the fourth and fifth layer output as the target characteristics.Then,in the first frame,a regression loss function is used to screen out the feature channels that can sense the target,which is used to guide the channel selection in the subsequent tracking.Finally,in order to further improve the tracking performance,the depth feature is used to obtain the rough location of the target,and the search window is updated,and then the manual feature is used to accurately locate the target.Experimental results on several UAV data sets show that the proposed algorithm has high success rate and accuracy,and has good tracking performance under complex scenes such as target rotation,deformation and background interference.