Research On Vehicle Image Recognition In Haze Environment

With the rapid development of automobile production in recent years,automobile has became a necessity of people’s daily life,especially the rapid development of electric vehicle,intelligent vehicle and driverless vehicle.The rapid development of science and technology needs technical support.The vehicle detection system plays an important role in the intelligent transportation system.With the rapid development of computer technology,computer vision technology has gradually appeared in people’s field of vision,image processing technology has became one of the most hot research topics,vehicle detection technology based on image processing has been widely concerned,in good weather,vehicle detection and recognition rate is high,and the robustness is also very good.But how to carry out real-time detection and recognition of many vehicles on the road in the haze environment is a key step in the development of intelligent transportation.By using the edge detection and license plate location methods and the rapid development of depth learning algorithm,the vehicle detection and recognition in the haze environment,play their respective advantages,can improve the accuracy of vehicle detection and recognition and detection efficiency to improve the intelligent transportation system.First of all,aiming at the problem that the color of some vehicles on the road and the road surface will be close to the haze,this paper proposes a method to improve the dark primary color failure by introducing parameters to optimize the transmittance of the demisting algorithm.Based on the atmospheric scattering model and the prior theory of dark primary color,the image collected by the vehicle camera is de fogged.After optimizing the defog parameters,the problem of color distortion and contrast decrease can be reduced.Then,aiming at the problem that the light condition is poor and the image is fuzzy in the haze environment,the lane line and the shadow at the bottom of the vehicle can not be used to detect the vehicle,this paper puts forward a method to initially determine the existing area of the vehicle through the edge features of the vehicle,and then combine with the location of the license plate to realize the accurate location of the vehicle.In the process of vehicle edge detection,aiming at the problem of fuzzy vehicle edge information in the haze environment,the edge information in the image can be well displayed by reducing the sensitivity threshold of the edge detection algorithm.For the problem that the sensitivity threshold of edge detection algorithm will be reduced,which will lead to the false detection of billboards,guide signs and other objects on the road into the possible areas of vehicles,the vehicle license plate location can be realized by locating the possible areas of the detected vehicles.In order to solve the problem that the location of license plate can not be achieved by the gray-scale characteristics of license plate and wavelet analysis,this paper uses the edge operator to detect the edge of license plate,and then uses morphology to process the initial location of license plate;through the color statistics of the possible areas of license plate,the upper and lower left and right frames of license plate are obtained accurately,and the license plate is refined Accurate positioning,so as to achieve the accurate positioning of vehicles,improve the recognition rate of vehicles.Finally,choose the yolov3 deep learning network which has better performance in image processing to identify vehicles on the road.It can detect large,medium and small vehicles in real time,and the detection speed is fast.Because yolov3 can detect 80 kinds of objects,and because there are many kinds of objects in the original data set coco,its prior frame scale and size can not well coincide with the actual intelligent transportation system and the vehicle boundary frame in the intelligent driving environment.To solve the above problems,this paper uses K-means clustering algorithm to optimize the anchor box design,and then determines the anchor box of the final target detection through non maximum suppression,so that the vehicle target and the boundary box coincide.In order to improve the training speed and recognize the vehicles in the haze environment,the haze images captured by the camera and the randomly selected open data set images are made into more complex training sets and test sets.During the training,the special data set Kitti of vehicle test is used for the first training to obtain the training weight,and then the obtained weight is used as the initial weight of the training set(including the self collected image and the Pascal VOC data set image)made by the training itself for the second training.By selecting the deep learning framework,the activation function and the loss function,a deep learning neural network with high recognition rate is obtained to realize vehicle detection,and the detection rate reaches 86.82%.