Target Tracking Algorithm With Adaptive Fusion Of Multi-layer Convolution Features

In the area of target tracking,the situations such as deformation,occlusion or similar object interference during the tracking process are all difficulties of tracking technology.In view of these problems,the target tracking technology is constantly improving and developing.Correlation tracking algorithm is widely used because of its fast speed and robustness advantages.This thesis mainly studies some mainstream traditional correlation filtering algorithms,and improves them by analyzing their existing defects.Traditional correlation algorithm uses the hand-crafted features such as gradient histogram,color,texture to describe the target information when constructing the target appearance model.These features lack the target semantic information and are difficult to deal with complex environments.In the correlation filtering algorithm,the feature is one of the important factors determining the tracking effect,and the feature is further developed into a convolution feature.Recently,some researchers have exploited convolutional neural networks to extract features which overcome the shortcomings of traditional feature information combining with correlation algorithms.In this thesis,we analyze multiple convolutional layers in the VGG-19 network,and select the first and fifth convolutional layers as the feature extraction layer because of the complementary characteristics of the high-level and low-level convolution features.Two different correlation filter models are obtained by using the convolution feature,and then the convolution response weight is calculated according to the APCE(Average Peak to Correlation Energy)measurement method,and the adaptive weighted fusion response graph is used to determine the final position of the target.In the tracking process,the target often changes in scale.This thesis studies the algorithm of adding scale estimation,and uses the pyramid principle to estimate the target scale.After determining the position,the optimal scale of the target is estimated by extracting the Histogram of Oriented Gradient(HOG)feature of the target.In this thesis,based on the correlation filter tracking framework and multi-layer convolution features,an adaptive filtering tracking method based on adaptive fusion multi-layer convolution features is proposed.The experiment selects 20 challenging video sequences in the public test video set to compare with other tracking algorithms,and it is still robust in complex situations.