Research On Object Tracking Based On Convolutional Neural Network

In recent years,with the rapid development of high-pixel imaging equipment and the ability of computer processing data,the field of computer vision has been developed.The research of vision-based target tracking technology is an important direction of computer vision.Target tracking is widely used in intelligent monitoring,human-computer interaction,medical diagnosis and so on.In the tracking process,it is often faced with changes in lighting,deformation of the target,and occlusion of the target.Although it has been studied for decades,the target tracking is still a very challenging issue.At present,designing a real-time tracking algorithm which can track any specified target in any scenario is the focus of the tracking field.Feature extraction is one of the most important parts in target tracking.The apparent feature of the target directly affects the accuracy of the tracking algorithm.Since 2012,the convolutional neural network has shown excellent performance in many fields of computer vision,and its powerful apparent ability has been favored by researchers.This article first carries out classification experiments on the AlexNet network structure,and analyzes the characteristics of features extracted from different convolutional layers through visual operations.With the increase of convolutional layers,features extracted from the network become more and more abstract,and the loss of local features of the target becomes more and more serious.This feature helps the classification of the target.Different from the classification problem,the target tracking problem needs to retain more local information,such as location information.Visual analysis deepens the understanding of the network and facilitates the tracking of the network design in the algorithm.Aiming at the shortcoming that the neural-network trackers trained online cannot meet the real-time requirement,a method based on convolution neural network trained offline and particle filter is proposed.In this algorithm,a generic relationship between object motion and appearance is learnt by a 2-channel CNN with offline training.Without the need to update the network model online,the location of the target and the prediction of the corresponding confidence are directly obtained through network regression.The design of the offline tracking network is used to ensure the real-time performance of the tracking algorithm.At the same time,for the occlusion,rapid movement and other situations that occur during the tracking process,the robustness is ensured by the fusion with the particle filter-based tracking algorithm.Compared with several other tracking algorithms,the experimental results on the VOT2014 dataset show that the proposed tracking method achieves the state-of-the-art performance at a very high speed of 40 frames-per-second.