Multiple Scale Object Tracking Combining Correlation Filters And Deep Network

Object tracking is a significant research content in the field of computer vision.It is a task that obtain the target trajectory in the video sequence,by given the initial position of the target.With the rapid development of deep learning technology,great progress has been made in the field of object tracking.Object tracking algorithms combining correlation filters(CF)with deep learning has attracted wide attention.However,due to the existence of occlusion,scale variation and motion blur,object tracking algorithms are still challenging.Based on the research that the object tracking algorithms combining correlation filters and deep learning,the main work this paper is as follows:(1)We propose a novel object tracking method that combining correlation filters with ResNet.Because the traditional convolutional neural networks do not have residual structure,the features from last layer of network lack detailed information,it is difficult to cope with the challenges of illumination change and target occlusion.Meanwhile,CF-based methods usually update filters at every frame even occlusion occurs,which reduces the discrimination ability of the algorithm.Firstly,we use the features extracted from different layers of ResNet to produce response maps,and then,in order to locate the target more accurately,these response maps are summed and fused based on AdaBoost algorithms.Secondly,to prevent method from updating when occlusion occurs,an update strategy based on occlusion detection is proposed.Finally,a scale filter is used to estimate the target scale.This algorithm can track the target accurately under the challenges of illumination change,target occlusion and scale variation.(2)We propose a real-time target tracking algorithm based on Siamese network.DCFNet integrates correlation filters into the network as a layer of neural network,which can achieve real-time speed.However,it only use the features from one layer of network,while limiting the robustness and accuracy.In this paper,our algorithm combines early convolution layer and last convolution layer,then train with correlation filter layer,so that the network can extract robust features.The features from different layers are suitable for different situations,in order to select features and enhance the robustness of the algorithm in different situations,response map are fused based on the confidence of response map.For target scale variation,the proposed method get candidate regions by multi-scale sampling,candidate regions combined into a batch,and then input to network to track the target scale adaptively.The proposed algorithm can track the target accurately in complex scenes at real-time speed.(3)Our algorithms is validated on OTB and VOT.Compared with the current mainstream algorithms,the experimental results show that our algorithms improves the success rate and precision in the case of occlusion,motion blur and scale variation,and can run at real-time speed.