Research On Adaptive Moving Target Tracking Method Based On Convolutional Neural Network And Related Filtering

Moving object tracking is an important research subject in the field of computer vision and image processing.It is a task that predict the position of the tracked object in a continuous video frames with a given initial position.With the improvement of computer technology and the rapid development of deep learning,visiual object tracking has made great progress and correlation filters-based tracking algorithm has achieved significant performance.However,due to the existence of complex scenarios such as deformation,occlusion,illumination variation,motion blur and so on,the existing object tracking algorithms are difficult to reach great accuracy and robustness.Aiming at some problems and shortcomings of current object tracking algorithms,this thesis mainly conducts theoretical analysis and research on the correlation filters-based tracking algorithm,makes full use of its performance advantages and makes a series of innovations and improvements in terms of feature selection,model updating and redetection.It works as follows:(1)We propose a correlation filtering tracking algorithm based on multi-feature fusion.This algorithm makes full use of the performance advantages of many different features to solve the problem that a single traditional feature cannot fully characterize the object and cannot adapt to complex scenarios such as illumination variation,deformations and background clutter.Firstly,HOG(Histogram of Oriented Gradient)features and color histogram features are fused with fixed weights on the output response map,and then they are adaptively fused with the response map of convolution features.Finally,the position of object is determined according to the maximum peak of the fused response map.In addition,in order to adapt to the scale variation of object,we peopose a way of separately learning the scale discrimination correlation filter.In terms of model updating,the algorithm implements sparse update of the model by presetting a fixed threshold as update condition,preventing models from being polluted by incorrect updates.The proposed algorithm is evaluated on the OTB-2013 benchmark dataset.The experimental results show that the algorithm has a significant improvement in accuracy and success rate,and has robust and real-time performance under the complex scenarios such as deformation,illumination variation,scale variation and so on.(2)We propose a multi-feature correlation filtering tracking algorithm combined with re-detection mechanism.Aiming at the problem that the above-mentioned adaptive fusion multi-feature tracking algorithm cannot deal with challenges such as occlusion,out-of-view and motion blur in the long-term tracking process,a re-detection mechanism including multi-peakdetection and online SVM classifier is proposed.In addition,in order to further improving the model quality,this algorithm also proposes an adaptive model update strategy,which uses double thresholds that are peak value and APCE(Average Peak-to Correlation Energy)value of the response map as the model update criteria.The proposed algorithm is evaluated on OTB-2013 and OTB-2015 benchmark datasets.The experimental results show that the algorithm has greatly improved compared with most state-of-the-art tracking algorithms,and it is more robust in complex scenarios such as complete occlusion,out of view,motion blur and so on.