Research On Vehicle Detection Based Vehicle Video

With the rapid development of society,the traffic problems generated by automobiles are increasing rapidly.Improving the perception ability of automobile and avoiding traffic accidents has become the frontier area of academia.On the basis of the vehicle camera,the high quality vehicle detection algorithm can increase the visual perception of the car and reduce the vehicle collision accident.It has a wide application in the fields of automatic driving and intelligent transportation.In recent years,traditional vehicle detection algorithms have been used in actual use,but the performance of vehicle video performance is general because of the many factors such as the variety of the vehicle type,the large background environment,the movement and the limitation of the detection range.Under the support of the key science and technology project(2017H6009)of Fujian Province,this paper proposes a vehicle detection algorithm based on the convolution neural network,and proposes a vehicle state analysis method based on the complicated calculation process,low detection accuracy and limited scope.The work completed in this article is as follows:First.We design a vehicle video dataset with different perspectives,models,scales,normalization and target tagging,which are applied to training and testing vehicle detection algorithms.In order to study the vehicle status analysis,the vehicles tagged in the vehicle video data are intercepted,and the brightness shift of the color space is made,and the vehicle state data set is designed.Second.A side evaluation method is designed,and the influence of activation function on network design is evaluated.Different from the traditional algorithm,the convolution neural network algorithm can achieve end to end computation,in which the characteristic transmission between each data layer must be mapped by the nonlinear activation function.In the case of other structures,the network of different activation functions is designed to train the vehicle detection model,and the convergence of the model and the accuracy of the test are compared,and the comparison results are used to reflect the effect of the activation function on the vehicle detection network.The experiment shows that the performance of vehicle detection network using Relu activation function is the most stable,and the Leaky activation function will produce a network that is superior to Relu when the parameter adjustment is suitable.Third.Using the above conclusion,combining the advantages of the existing target detection network,the codec grid regression network is designed to do vehicle detection,and the vehicle status classification model is trained with the ultra fast AlexNet network.First stage vehicle detection,firstly,the feature extraction network structure is used as the coding region,the coding region is structured and the optimal feature extraction network is fused.This paper uses the VGG16 feature to extract the network.Then,the candidate frame output structure is used as the decoding area,the decoding area combines the Anchor mechanism of the Faster-RCNN network and the grid regression of the YOLO network.The mechanism makes the candidate box fast regression and trimming.In the second stage vehicle state analysis,the vehicle on the road is divided into two types:the forward driving vehicle and the opposite vehicle;then the vehicle status is analyzed in the form of AlexNet classification.The experiment shows that in different data sets,the average accuracy of the codec grid regression network is over 5%higher than the average accuracy of other algorithms,and can be detected in real time.The average accuracy of the AlexNet vehicle state analysis model is 80%.To sum up,this paper presents a new convolution neural network structure for vehicle detection based on vehicle video.It improves the accuracy of vehicle detection algorithm and analyzes the vehicle running state.It has important research significance and practical value.