Adaptive Target Tracking Based On Bounding Box Regression Model And Feature Optimization Learning

Target tracking is an important direction in the field of computer vision,which is widely used in UAV control,video surveillance,human-computer interaction and so on.The main task of target tracking is to accurately and reliably predict the position and size of the target in subsequent image sequences after a given target to be tracked.The difficulty of target tracking is that the target is arbitrary,the number of positive samples used for model training is small,and the change of target is unknown and diverse during tracking.In this paper,the problems of small number of training samples,target deformation,illumination change and occlusion are studied.The main results are as follows:An adaptive target tracking method based on the boundary box regression model is proposed.This method generates training samples by Gaussian mixture model,and the depth feature of the training sample is extracted by the deep convolution neural network VGG-M,obtains the reduced dimension multi-resolution depth characteristics by interpolation and convolution factorization operations,trains the spatial penalty correlation filter by using this feature,using the filter to get the initial position of the target,and determines the final position of the target by the initial position regression of the boundary box regression model.The training samples are updated according to the peak value and oscillation degree of the filter response graph.Experimental results show that the tracking speed of the algorithm can meet the real-time requirements and the tracking results have a certain accuracy.In this paper,a feature-based optimal learning method for target tracking is proposed.A multi-domain network with three layers of convolution layer and three layers of fully connected layer is constructed by using some layers of VGG-M network and trained by multi-video sequence,the image features extracted from multi-domain networks are offline to guide the training of feature optimization networks,in the stage of online tracking,uses multi-domain network and feature optimization network to build a biclassification network,uses biclassification network to preliminarily detect the target,according to the detection location,uses boundary box regression model to determine the final target.In order to avoid over fitting problem of tracking network,the algorithm selects difficult negative samples to train network.Experimental results show that the algorithm can complete the task of target tracking in a variety of scenarios.A SVM detector-based target tracking method for twin feature networks is proposed.By using the deep neural network VGGNet-19 some layers to construct the twin feature network of five-layer convolution layer and two-layer pooling layer,the ridge regression and scale estimation function are used as the optimization functions of the twin feature network.The perceptual features of the target are extracted by twin feature network,and the location of the target is determined by template matching algorithm.In order to avoid tracking drift,the confidence value of the target is determined by SVM detector,and the perceptual features of the target are updated adaptively according to the confidence value.The experimental results show that the algorithm can be used to track the objects with appearance deformation,occlusion and blur,and the tracking performance of the algorithm is good and stable.