Research On Single Object Tracking Algorithm Based On Multi-layer Deep Feature Fusion

Visual object tracking algorithm tracks the interested target specified at the first frame and estimates the target position and size of each frame in the tracking process,and finally obtains the complete motion trajectory of the target.It has been widely used in real-life scenes such as automatic monitoring,intelligent human-computer interaction and so on.Recently,advanced visual object tracking algorithms have introduced deep learning technology to further improve the tracking accuracy.However,when the appearance of the object changes suddenly or moves rapidly,the existing algorithms will still lose the object.In addition,training a neural network suitable for visual object tracking requires a lot of time and computing resources,which can not adapt to the equipment with insufficient computing resources.In this paper,the pretrained neural network is used to extract features for object tracking.The object is characterized by fusing features of different depths,which improves the tracking accuracy of the algorithm and meets the real-time requirements of tracking in complex scenes.The main research contents are as follows:1.Aiming at the poor anti occlusion ability and robustness of traditional correlation filtering methods in complex environment,an object tracking algorithm based on tree structure correlation filter is proposed.The algorithm maintains multiple correlation filter nodes in a tree structure and stores multiple object appearance template information.The algorithm uses the off-line trained VGG network to extract the features of the image,learns the corresponding correlation filters on the feature layers of different depths,and combines them to form a powerful correlation filter to ensure the accuracy of tracking.In order to maintain the reliability of the correlation filter nodes,an appropriate method is adopted to update the correlation filter nodes in the tree structure to ensure the robustness of the method.The experimental results show that the proposed algorithm can still track the target well when the target is occluded,and has good tracking performance.2.For off-line training a neural network suitable for object tracking task spends a lot of time and computing resources,a real-time tracking algorithm based on pre-trained Dense Net network is proposed.The original intention of Dense Net network design is to realize image classification tasks of 1000 categories,while the object tracking task only needs to identify the object and background,and when it is applied to the field of object tracking,the extracted features will have a lot of redundant information.Therefore,this paper improves the network,and proposes a method to measure the importance of channel features,eliminates redundant features,and multiplies the corresponding weight value according to the different importance of different channel features to obtain the best features.The experimental results show that the proposed algorithm can achieve better tracking accuracy and meet the requirements of real-time in the scene with complex background.