Research On Object Tracking Based On Feature Learning And Location Prediction

Visual object tracking is widely used in intelligent monitoring,unmanned driving,human-computer interaction and other fields,and is one of the hot research directions in recent years.The difficulty is that the appearance of the target to be tracked is easily affected by disturbances such as illumination,occlusion,deformation,and scale,which leads to inaccurate tracking.This thesis focuses on three parts: feature learning,scale change,and model update,which affect tracking accuracy in target tracking.The main research work is as follows:Firstly,analyzing the problem that the single feature representation in the traditional Mean Shift algorithm causes the object's apparent modeling ability to be weak,and propose a multi-feature fusion feature learning method to fuse the color features and texture features of the target image.The fusion of different features enhances the target's apparent representation and improves the accuracy of the local detection model.At the same time,Kalman filtering is introduced to predict the target motion,combined with the position information of the previous frame to determine the target position,and after determining the position,continue to use the updated model for tracking.Secondly,in view of the problem of weak discrimination of traditional manual features,an online learning convolutional neural network training method is proposed,which is effectively convolved with the target background information obtained by Kmeans sampling.The object features extracted with context are used as the multichannel input of the kernel correlation filter,thereby simplifying the training process of the neural network.In the process of target movement,the scale change will lead to inaccurate tracking,so the scale-related filter is introduced to combine scale change and model update to realize the adaptive adjustment of the scale frame.Finally,in order to solve the problem of low real-time online learning,using offline training convolutional network to extract target features,learning the general relationship between object motion and appearance representation.Image features change from shallow to deep,and deep features contain more semantic information to help distinguish the target from the background.At the same time,the prediction of the object position and size by the network is converted into a regression of the size and position of the rectangular frame,which accelerates the tracking speed.Since object occlusion may cause tracking loss during long-term tracking,a loss-detection algorithm is incorporated to enhance tracking robustness.The methods proposed analyze the effects of feature learning,scale changes,and model updates on tracking effects from different perspectives,and conducts sufficient experimental comparisons on multiple target tracking data sets.Experimental results show that the proposed algorithm has achieved good tracking performance.