Research On Moving Object Tracking Based On Deep Learning

Object tracking is the task of locating and modeling the object in the first frame according to the context information of the video image sequence,and then performing continuous tracking in subsequent frames.Target tracking has important uses in areas such as intelligent traffic supervision,public safety monitoring,and autonomous driving.The emergence of deep learning related methods provides more options for the research of object tracking.Although many new algorithms have made great progress in target tracking in recent years,complex backgrounds,illumination changes,occlusion,deformation,etc.are still the main factors that affect the accuracy and accuracy of target tracking.In recent years,object tracking algorithms based on deep learning methods have made great breakthroughs in solving the problem of object tracking.They have achieved excellent tracking performance on a series of public data sets such as OTB(2013-2015)and VOT.However,due to the lack of training data and the complexity of the actual scene,the object tracking algorithm related to deep learning technology can not achieve the ideal tracking effect.How to achieve a proper balance between the amount of calculation required by the deep network to achieve powerful characterization functions and the realtime requirements of object tracking is the difficulty of the current object tracking algorithm research.By studying the object tracking algorithms related to deep learning,this paper uses convolutional neural networks to construct a real-time tracking framework to predict the target position in the video sequence,and propose two effective object tracking algorithms.The main innovations of this article include:(1)Aiming at the occlusion and similarity problems of vehicle object tracking in intelligent traffic environment,this paper proposes a object tracking algorithm based on multi-domain convolutional neural network.First,use the Mask R-CNN algorithm to segment the object to be tracked,and clearly express the foreground and background regions.Secondly,input images and perform network training in a multi-domain learning manner,so that each convolutional layer can learn the common features of the tracking object in a single video sequence,thereby improving the feature extraction ability of the algorithm.Finally,the parameters of the convolutional layer are fixed,and the object-related features are extracted from the previous sequence images in the interval period,so as to ensure that the fully connected layer of the tracking model is completely updated,so that the algorithm can realize real-time tracking.The algorithm also improves the accuracy and success rate of the tracker by adapting to changes in the appearance of the moving target itself.(2)Aiming at the problem that the deep network cannot be applied in the object tracking task,this paper proposes a new moving object tracking algorithm based on the Siamese network and combined with the residual network.First,input a pair of pictures into the model,and the algorithm calculates the similarity between the object template in the picture pair and the given search area through the twin network,retains the highest similarity area and records it as the location of the tracked object.Then,by fusing the traditional features with the semantic features extracted by high-level convolution,the semantic information is complemented.At the same time,meta-learning is introduced in the training phase to solve the problem of insufficient training data,and weighted crosscorrelation is used to overcome the limitations of the Siamese network,so that the model can achieve more robust results on less data.Finally,the algorithm improves the adaptability to object changes by reducing the degree of model fitting(3)Aiming at the problem that the convolution operation only extracts features from local neighbors and lacks global information,this paper introduces the extrusion and excitation module and improves it.First,use the attention mechanism to mine the position information in the video frame,and cascade it with the convolutional neural network.Then,a Squeeze and Excitation structural unit is introduced to model the relationship between the picture channels and adaptively match the relevant feature responses of the picture channels.Finally,the channel feature information is merged to improve the representation ability and feature expression ability of the entire network.The module also selectively enhances useful functions and suppresses useless functions by learning global information,so that the network can perform function recalibration.