Research On Deep Learning-based Vehicle Identification And Positioning Application In Complex Environment

With the increasing throughput of traffic network to vehicles,more and more people pay attention to the severe traffic forms and frequent road vehicle safety problems.Therefore,it is urgent to develop intelligent transportation.One of the most important steps to develop intelligent transportation is to identify and track vehicles accurately and quickly in complex road scenes,which can effectively solve practical problems such as vehicle collision,traffic accident identification and traffic congestion determination.1)this paper starts with the study of the classic vehicle tracking algorithms at home and abroad,obtains and marks the data set of vehicle image,prepares the data for neural network training and verification,studies the depth model based on multi feature subspace distribution and the vehicle recognition algorithm based on online migration learning,builds the bottom layer with multi-layer restricted Boltzmann mechanism,and constructs the depth belief network The deep model of the superstructure uses the transfer learning to identify the vehicles in the complex road scene.The advantages and disadvantages of the algorithm are analyzed in terms of recognition accuracy and recognition speed,so as to further find a vehicle recognition algorithm that can meet the needs of complex scenes.2)A fast and accurate vehicle recognition algorithm(FAVR)is proposed based on the deep transfer learning of multi feature subspace distribution and Yolo neural network.Firstly,the model is trained and the parameters are optimized by inputting a variety of self collected vehicle samples and KITTI vehicle data set in complex environment to identify vehicles in complex scene.Then,combined with the vehicle tracking algorithm,the vehicle tracking is stable,accurate and real-time.The accuracy,accuracy,accuracy,recall rate and F1 of the algorithm are 92.18%,80.73%,78.67% and 79.95% respectively.Through the experiment of the trained network model,the accuracy and efficiency of the recognition algorithm and the vehicle positioning algorithm based on Kalman filter are verified.The FAVR algorithm proposed in this paper uses batch normalization regularization,which can avoid over fitting and accelerate the convergence of the network;delete the pooling layer to alleviate the information loss caused by pooling;cancel the full connection layer to avoid the deep network gradient disappearance;predict the target from three scales to solve the problem of small target false detection rate of the Yolo algorithm and improve the network push At the same time,high precision is obtained on two kinds of data sets.