The Study Of Directed Perturbation And Multi-feature Object Tracking Based On Convolutional Neural Network

With the continuous development of computer vision and science and technology,object tracking technology has been widely popularized and applied to every corner of people's life.Areas such as defence surveillance,social security,smart cities and assisted medicine are all dependent on object tracking technology.In recent years,the success of deep learning and the substantial improvement of computer performance enable the deep learning method to be widely used in the object tracking task and achieve better results than the traditional method in a series of data sets.Therefore,the object tracking technology based on deep learning has been highly concerned by scholars at home and abroad.Although object tracking technology has made gratifying achievements,in the process of video target tracking,there are all kinds of interference in video scenes,such as rapid target movement,camera shake,light change,target occlusion and so on.These interferences will decrease the accuracy,stability and success rate of the tracker.Although some trackers have high accuracy and success rate in a limited class or even a class of tracking tasks,such as vehicle tracking,pedestrian tracking,missile tracking and so on.However,the general generalization ability of these trackers is not strong,and it is difficult to apply to the general object tracking task.The general object tracking algorithm still has a lot of room for improvement.Many deep learning based trackers only consider the convolution feature of the object.If you can add other effective tracking features can be better to complete the tracking task.In this paper,the general object tracking problem is studied based on the features of convolutional neural network.In the tracking process,it is corrected by directional perturbation algorithm and multi-feature algorithm,so as to improve the accuracy,stability and success rate of the tracker.The main research work is as follows:(1)In most traditional deep learning based trackers,the default candidate position of the next frame is the gaussian distribution centered on the result position of the previous frame.However,in the case of practical application,once the object moves rapidly,the camera shifts and the object is lost,etc.,the tracker will generate candidate positions in the region relatively far away from the real position of the object,which seriously affects the accuracy,stability and success rate of the tracker.In view of this shortcoming,this paper makes full use of the feature that convolutional neural network can locate,changes the perturbation center of particle filter,and makes directionalperturbation sampling,so as to make the candidate sample closer to the real location,accelerate target recovery,prevent target loss,and improve the precision,success rate and robustness of tracker.(2)Because the tracker is very sensitive to the initialization image,giving different initialization images in the same video image sequence will have a great impact on the tracker performance.In order to solve this problem In this paper,HOG feature and color histogram of regions of object and candidate images in different time periods are extracted respectively on the basis of extracting features of convolutional neural network,and the optimal solution is found by calculating similarity.When HOG feature and color histogram feature are added,the tracker can adapt to more complex scenes and improve the robustness of the tracker.In this paper,the effectiveness of the proposed algorithm is verified on the obt-13 which is a general object tracking benchmark database and compared with some mainstream trackers.This method can also be extended to other trackers.