The Study Of Key Technique Of Image Processing In Smart Traffic System

The collected road traffic video data are rich and could be used to calculate traffic parameters in the entire road intersection based on image processing technology.However,there are difficulties in collecting vehicle flow and vehicle density data in intersection for complex traffic environment.The key contributions of the study are listed below:(1)We proposed a joint training method for vehicle detection and vehicle attribute recognition based on deep learning methodology.First,based on the SSD framework,the multi-scale feature map is extracted by the convolution neural network.Moreover,at each point of the feature map we used a 3x3 convolution kernel to predict vehicle foreground and background.A fully connected layer is used to predict vehicle type,color,and pose attributes.In the loss function,the whole error consists of the classification error of the vehicle,the background,the position error of the vehicle,and the vehicles' attribute error.In the training process,the online hard case mining technology is applied to enhance the learning of vehicle samples and to improve the accuracy of vehicle detection in real traffic scenarios.Which provides a guarantee for vehicle detection and fine grained attribute prediction in complex scenes.(2)We proposed a method of training vehicle feature based on triplet loss function to solve the problem of vehicle re-association after occlusion or failure detection.First,the vehicle are tracked by edge feature matching and data association,and then the vehicle dataset in the tracking process is constructed.The dataset includes continuous multiple angles of the vehicle in the running process,which satisfies the changing of vehicle's angle and the complex illumination condition.The trained deep model converts the vehicle image to the lower dimensional space.In the low dimensional space,the triplet loss between the vehicle images are computed.The extracted vehicle feature could not only achieve good performance in continuous tracking,but also in complex scene.(3)A lane estimation algorithm in the intersection is proposed.Based on vehicle tracking trajectory,the traffic flow curve is computed and the phase switching time of traffic lights are calculated at each intersection.Then the delay time of each vehicle is calculated and the turning type of vehicle(turning left,turning right,and going straight)is identified.The approach of lane estimation algorithm is computed based on the vehicle's trajectory which is adaptive to the complex situations at different time and junctions.(4)We improved algorithm of vehicle density estimation based on a static image.First,in the manually tagged vehicle distribution picture,we try to fit the main direction of vehicle distribution based on the location of each marked vehicle.In addition,we merged an adaptive geometric Gaussian distribution for all the labeled vehicles to form the vehicle density map.The convolution neural network is applied to learn a mapping from original image to the corresponding vehicle density map.Finally,a density map estimation method based on residual learning is designed to obtain higher fitting accuracy.The vehicle density estimation algorithm presented in this study could not only be applied to vehicle counting under ideal scene with high resolution and small vehicle density,but also could be applied to vehicle counting for images with imperfect image resolution,high traffic density and badly occlusion.