Research On Vehicle Detection And Traffic Flow Analysis Methods In Complex Environments Based On Deep Learning

Vehicle detection and traffic flow analysis based on deep learning,which uses deep learning and computer vision algorithm to locate and identify the target vehicle in the video or image,is an important research topic in the field of intelligent transportation.Although in recent years the target detection algorithm in automatic driving technology,intelligent transportation systems,and other fields has made by leaps and bounds in the development,the detection speed and accuracy than traditional image processing algorithm,but in the real traffic complex environment,because of the vehicle scale size differences,mutual occlusion vehicle,trees shade and light to each other and the interference of pedestrians,based on the deep study of vehicle detection and traffic analysis algorithm in accuracy and speed is still insufficient,difficult to meet the requirements of the reality.Aiming at vehicle detection and vehicle flow detection in complex environment,this paper carries out research on vehicle detection and vehicle flow analysis method based on deep learning,and the specific research content includes the following aspects:Firstly,in view of the traffic road middle and small scale vehicles detection,multiscale prediction was carried out by convolution algorithm of neural network and the fusion research,through to lower the details of the network to extract features and highlevel semantic characteristics of network extraction,multi-scale prediction scales the size of the vehicle,so as to improve the accuracy of vehicle for small scale,the average accuracy of 84.6%,solve the problem of small scale vehicles were missing.Secondly,in view of the mutual occlusion vehicle segmentation,multiple expansion was carried out by convolution and predict the NIo U algorithm research,by increasing the receptive field,improve the ability of feature extraction,predict the size of the mutual occlusion vehicle at the same time,the judge blocked contact ratio of the vehicle,to improve the localization and the segmentation accuracy,the average accuracy of 82.3%,solving the mutual occlusion vehicle vehicles instance segmentation incomplete mask or repeated integral problem.Then,aiming at the problem of vehicle recognition under low light,the paper analyzes that the contrast of vehicle information is not obvious in low illumination environment,and proposes an adaptive image enhancement network,which can effectively enhance the contrast of image,so as to improve the accuracy of vehicle recognition.The accuracy reaches 86.1%.Finally,under the environment of dense traffic detection research,due to the dense shade environment in vehicle,dimension is different,even in low light environment,so you need to study before fusion detection algorithm of small size scales,keep out vehicle segmentation algorithm,low light recognition algorithm,improve the dense environment of vehicle detection precision,at the same time in order to reduce the interference of surrounding environment,put forward a kind of background prediction algorithm,by predicting the environment background,reduce the interference of background categories,to improve traffic detection accuracy.The accuracy rate is above 90%...