| In recent years,the field of intelligent transportation has been continuously developed with the promotion of artificial intelligence.Vehicle tracking technology,as a basic problem in the field of intelligent transportation,has attracted more and more attention.Vehicle tracking methods based on deep learning have become hot topics in the field of intelligent transportation,such as target tracking models based on convolutional neural networks.Although this type of model has made breakthrough progress in tracking accuracy,it has poor real-time performance.Therefore,the target tracking method based on the Fully-Convolutional Siamese neural network(SiamFC),modeling the target tracking task as a similarity matching problem,well balanced the tracking speed and accuracy,and became the focus of research in this field.However,the SiamFC model lacks robustness to target deformation and complex background interference,and has insufficient discrimination and generalization capabilities.Reduced and proposed an Improved Fully-Convolutional Siamese neural network vehicle tracking model.The main research contents of this article are:(1)The current mainstream algorithms in the field of target tracking based on deep learning are studied,mainly including target tracking models based on convolutional neural networks and target tracking models based on Fully-Convolutional Siamese neural network.The process of training and tracking,as well as the principles that can achieve target tracking,and the advantages and disadvantages of the two class models are analyzed.(2)Based on the Fully-Convolutional Siamese neural network,an Improved Fully-Convolutional Siamese neural network target tracking model is proposed.The main improvements are as follows: During the training phase,the Triplet loss is adopted to replace the logistic loss function in the SiamFC tracking model.The Triplet loss contains more elements,which helps to mine more potential relationships between the exemplar,the positive instance and the negative instance.VGG16 is used instead of the CNN network used in the SiamFC model for feature extraction,which can be pre-trained on large image datasets to obtain powerful feature expression capabilities.And,a channel attention module is introduced,which can learn different feature channel weight coefficients in the model training process,thereby achieving adaptive tracking of different objects.In the online tracking phase,a Distractor-aware module is introduced,which effectively suppresses the distractors in the background,and makes the model more robust to the tracking object.(3)Based on the improved SiamFC algorithm,a vehicle tracking model system is designed and implemented,and the training set is replaced by the data augmentation technology to replace the shortage of the vehicle tracking dataset.VOT2016 and OTB50 target tracking benchmarks contain many vehicle videos in complex road scenarios.In the experimental analysis,multiple vehicle tracking video sequences in the two benchmarks were replaced to test the Improved SiamFC vehicle tracking model instead of qualitative and quantitative analysis of the test results.The analysis results show that the Improved SiamFC vehicle tracking model is High accuracy and real-time performance are obtained on vehicle tracking in various complex road scenes. |
