Research And Implementation Of Target Tracking Algorithm Based On Road Traffic Scene

In recent years,with the rapid improvement of computer hardware performance,computer vision and pattern recognition theory has been developed vigorously,people for road traffic intelligent security system demand continues to increase,recognition and tracking in surveillance video of pedestrians and vehicles on the road and other traffic participants has become one of the research hotspots in academia and industry.The traditional correlation filtering tracking method could not solve the problem of object and scale zoom,border effect exists seriously.In addition,the prediction of traditional Siamese-FC network will be the convolution of the target area and the search area,easily disturbed by background objects.In order to solve the above problems,this thesis focuses on the study of tracking method in video monitoring of road traffic participants.Aiming at the problem of traditional filtering method of zoom scale and boundary effect,the paper introduces Box Regression Correlation Filtering(BRCF)traffic scene target tracking method based on local region,mainly including: 1)A mining method of positive and negative samples based on local region,effectively reduce the boundary effect in strengthening the classification capability of classifier at the same time;2)A bounding box regression method combined with a key points matching method for the scale prediction,the method firstly estimate a target object scale range,and then fine tune the bounding box by a regressor,lead to a fast and efficient calculation;3)An adaptive Gauss window function for object scale and an adaptive model update strategy based on the analysis of the response.With a Gauss window tightly wrapped the target area,the method effectively reduce the spectrum leakage.With the numerical analysis of the output filter response,the method can judge tracking environment for the adaptive updating model.The experimental results show that the proposed method can effectively deal with the problem of object size changes and background interference and other adverse conditions,the success rate and accuracy rate were higher than Discriminative Scale Space Tracking(DSST)by about 2 percentage points in the laboratory of traffic data set.In terms of processing speed,the proposed method is higher than DSST by nearly 40 Frames Per Second(FPS),can achieve the effect of real-time processing.Aiming at the problem that the Siamese-FC network is vulnerable to background interference,the paper proposed tracking method based on background semantic features of attention model of fully convolution siamese neural network.The algorithm is based on deep similarity neural network,supplemented by a human attention mechanism for self selection of deep features.It mainly includes the following contents: 1)A separation method for sample template and semantic background region which can seperate target and background in the initial frame,so that the sample template can be input to the network at any scale,increasing flexibility of the net;2)An A-Siamese-Net fused attention model with the adaptive selection of the deep features for siamese convolution.The semantic channel weight is generated by background region,and used for feature selection of the template and search region to measure similarity between the image blocks,causing higher separability of our model;3)A two stage method for model training.Firstly let the model pretrained on large scale public datasets,and then fine tuned on our traffic dataset for joint training,solving the problems of the dataset is not enough.The experimental results show that the proposed method improves the tracking accuracy without an increase in tracking time,strengthen the robustness of algorithm under the condition of dense background region.The overall ASiamese-Net in traffic data sets on the success rate is higher than Siamese network by 10 percentage points,the overall accuracy is higher than Siamese network than the 8 percentage points.