Research On Vehicle Detection,Tracking And Prediction Algorithm Based On Deep Learning

Intelligent Transportation System(ITS),which is an important link in the construction of smart cities,integrate a variety of high-tech technologies,such as artificial intelligence technology and sensor technology,into traffic guidance field and other related fields.In this way,ITS strengthens the relationship between Internet of Vehicles(Io V)and users,which finally help form a safe,efficient and energy-saving comprehensive transport system.Vehicle detection,tracking and prediction are hotspots in the research of intelligent transportation system.However,in practical application environments,it still has some problems such as missing detection of small targets,loss of tracking targets and poor accuracy of vehicle flow prediction.In order to solve the above problems,this paper,regarding driving vehicles in road traffic scenes as research objects,using deep learning technology to research and explore the algorithms of vehicle detection,tracking and traffic flow prediction respectively.Main research contents of this paper shown as follows:(1)Research on vehicle detection algorithm based on the improved YOLO v3.In order to solve the problem of leak detection of YOLO v3 detection model in vehicle detection process,this paper puts forward the improve YOLO v3 target detection algorithm.First of all,in the feature extraction stage,five scale features are designed to improve the network’s learning ability of shallow information of the image and to realize the accurate identification of vehicles.Secondly,in the target detection stage,K-means clustering algorithm is used to optimize the multiple-scale prior frame so as to adapt to different sizes of vehicle and to detect datasets from the two aspects of accuracy and speed.Then,this algorithm is compared with YOLO v3 algorithm and the results show that the average accuracy of this method can reach 92.21% and that its false detection rate is lower than 3.5% and its average detection rate is 48 frames/s.This algorithm is featured by high accuracy and fast speed.(2)Research on Vehicle Tracking Matching Algorithm based on Kalman Filter and Hungarian Algorithm.In order to solve the problem that YOLO v3 detection algorithm may be easily lost after it is used to detect the targets,this research puts forward the suggestion that the Kalman filter algorithm should be used together with Hungarian matching algorithm to deal with the matching relationship between the tracking box and the detection box.In this way,the tracking model is integrated with the detection model.The comparative experiment carried in the tracking sequence of the data sets showed that compared with the single detection algorithm,the detection model integrating the tracking algorithm can effectively restrain such phenomena as tracking lost and missed detection.(3)Aiming at solving the problem that existing prediction models could not fully extract the spatio-temporal features in traffic flow,we proposed an improved convolutional neural network(CNN)with long short-term memory neural network(LSTM)for short-term traffic flow prediction.First of all,a layered extraction method was used to design the network structure and one-dimensional convolution kernel which enabled automatic extraction of spatial features of traffic flow sequences;Second,the LSTM network modules were optimized to reduce the long-term dependence of network on the data;Finally,the optimization algorithm for rectified adaptive moment estimation(RAdam)was introduced to the end-to-end model training process,which accelerated fitting effects of the weight and improved the accuracy and robustness of network output.Experimental results showed that compared with the prediction model of stacked auto-encoders(SAEs)network,performance of the proposed model was enhanced by 3.55% and 8.82% on weekdays and weekends with model running times reduced by 6.2% and 6.9%,respectively;Compared with the prediction model of long-short term memory-support vector regression(LSTM-SVR),its performance was enhanced by 0.29% and1.79% with model running times reduced by 9.0% and 9.7%,respectively.Therefore,the proposed model was more applicable to the short-term traffic flow prediction of different time periods.The vehicle detection and tracking algorithm proposed has a certain universality and can provide a reference for the design of other target detection and tracking models in this paper.