Visual Target Tracking Based On Deep Neural Network And Correlation Filter

In the field of computer vision application,visual target tracking is a hot field which has been widely studied.It has been widely used in many major visual fields and achieved good practical application results.However,because the visual target tracking is easy to be interfered by various factors such as illumination change,scale change,background clutter,low resolution and so on,on the premise of meeting the real-time target tracking,this paper mainly makes the following improvements to further improve the accuracy and robustness of target tracking.(1)The visual target tracking algorithm based on correlation filtering theory uses the traditional shallow manual features,and the representation ability of the target to be tracked is insufficient.On this basis,the algorithm proposed in this paper uses the depth neural network model to extract the depth features of the target to be tracked,and because the depth features of different convolutions represent different information,the depth features of different layers will be obtained,The fusion of depth features makes the representation more powerful,and the kernel correlation method improves the tracking speed of visual target tracking algorithm.(2)Based on siamfc(Fully-convolutional Siamese Networks)The main network of feature extraction of visual target tracking algorithm of networks mostly adopts shallow alexnet network model,which can not make full use of the deep network’s representation ability.A new main network of target feature extraction is designed.The deep and wide feature extraction network is built by stacking the proposed network model,so as to replace the Alex used in siamfc Net network model structure.(3)For the siam network visual target tracking algorithm based on RPN(region proposal network),The vggnet network model is used as the depth feature extraction network to track the target;At the same time,the depth output features of different convolution layers are input into four siamrpn(siamese region proposal network)modules for convolution feature fusion,so as to regress and classify the tracking targets and further improve the target tracking accuracy.Finally,qualitative and quantitative experimental analysis is carried out in otb50,otb100 and other data sets.The experimental results show that the algorithm proposed in this papercan achieve a better balance between speed and accuracy,with better tracking effect and robustness compared with other algorithms.