Research On Multi-cues And Re-detection Kernel Correlation Filtering Object Tracking

There are many challenges in the tracking process,and these challenges often cause the tracking performance of the algorithm to be much worse than expected.In order to achieve accurate and robust tracking,this thesis combines some excellent algorithms in recent years,on the basis of correlation filtering framework and deep features,to address some of the deficiencies in the tracking process,the following works are done:To solve the problem of drift in the tracking,in this thesis,a re-detection object tracking algorithm based on fast multi-scale estimation is proposed.In the tracking process,the tracker locates the target position through the maximum value of the response and detects the reliability of the current position using a new self-adaptive detection criterion.In contrast to other detection criteria,our new detection criterion reduces the dependence on the maximum response value.If the current location is determined to be unreliable,our method can generate target candidate boxes by using the edge boxes algorithm and select the best target location by applying the non-maximum suppression(NMS)methods.Furthermore,we proposed an adaptive updating method to reduce the errors caused by tracking failure.Experimental results show that the proposed redetection algorithm based on the kernel correlation filtering framework has achieved the desired results in terms of accuracy and success rate of OTB data sets.In order to make full use of the advantages of different features and explore the relationship between multiple features,a multi-cue object tracking algorithm based on kernel correlation filtering is proposed.The HOG feature,conv4-4 and conv5-4 layer features of convolutional neural network were arranged and combined to form multiple expert trackers,so as to improve the diversity of the model.At the same time,this chapter puts forward the double evaluation mechanism,which can select the best clues produced by experts more accurately.In addition,the best clues generated will be used to update all experts in an adaptive update manner,which can alleviate the tracking drift of weak trackers.The experiments in the OTB dataset show that the multi-cue object tracking algorithm proposed in this paper has good robustness compared with other algorithms.