Online Visual Tracking Based On Convolution Neural Network

Visual tracking has attracted great attention in the field of computer vision because of its huge research and practical value in the direction of video surveillance,self driving,smart security and human-machine interaction.Visual tracking is simply to give the initial state of a target and locate the target in the video sequence.Due to complex background interference such as occlusion,background interference,illumination change and low resolution in tracking process,how to achieve ideal tracking effect in complex environment has always been a challenging problem.Convolution neural network is widely used in image processing field because of its powerful feature extraction function.However,convolution neural network is facing many problems such as high computational complexity,lack of training samples,lack of temporal and spatial information and online learning.In this paper,the convolution neural network as the theoretical basis for research,targeted to solve the key problems,the main research contents are as follows:In this paper,the convolution neural network as the theoretical basis for research,targeted to solve the key problems,the main research contents are as follows:First,according to the requirement of high algorithm hardware environment tracking convolutional neural network problems,lack of training samples and training and so on,this paper introduces Gauss kernel function to speed up robust feature extraction of target,using a simple two layer feedforward convolutional network,put forward a kind of no training simplified convolutional neural network algorithm.In the first frame,it is clustered to extract a set of normalized patches from the target region as initial filters,the simple abstract features is extracted by convolution the background and foreground information of object.Furthermore,all the convolutions of a simple layer form together a deep-level representation.The Gaussian kernel function is used to speed up the convolution operations;the local structural feature information is used to update filters and the tracking is realized by using particle filter tracking framework.The experiment shows that this method can learn the running environment from the tedious depth and improve the tracking efficiency in the complex background.Second,the convolutional neural network abstraction algorithm are lack of spatiotemporal context information.In this paper,combined with spatial-temporal context model as a filter of convolution neural network,a visual tracking algorithm based on online convolution neural network is proposed.Firstly,the initial target was normalized and extract the target confidence map,and then in the process of tracking with spatiotemporal information update spatio-temporal context model,the first layer using spatio-temporal model updated convolution on the input,and the convolution results are sliding window films,the second layer model is used to take the use of temporal convolution results,target extraction of simple and abstract characteristics,expression of deep and convolution simple superposition layer results obtained,finally combining particle filter tracking framework for target tracking.The experiment shows that the deep abstraction features of the spatio-temporal convolution network structure can retain the related spatiotemporal information and improve the tracking efficiency under the complex background.Third,deep networks have been successfully applied to visual tracking by learning a generic representation offline from numerous training images.However,the features of the convolutional neural network abstraction algorithm are lack of spatio-temporal context information and the offline training is time-consuming.It is difficult to cope with a series of changes in target tracking process.Based on the tracking algorithm of convolutional neural network based on temporal and spatial information,an adaptive online sparse learning method combining temporal and spatial information is designed to make the network adaptive online learning according to the change of targets and spatio temporal information in tracking process.In the first frame network,a simple model is trained.In the tracking process,update threshold is set up,and online adaptive sparse update is combined with spatio temporal information to avoid redundant updates and large time complexity caused by updating each frame.The experiment shows that online sparse learning can deal with complex tracking problems,such as scale change and occlusion in low resolution scene.