The Research On Road Detection Algorithm For Intelligent Transportation

In the field of intelligent transportation,the information on the road ahead is intricate,and the targets on the road include vehicles,pedestrians,and road surfaces.The computer vision system can obtain information on the road,and complete the detection and recognition of the target on this basis,so a set of efficient road information detection and recognition algorithms needs to be solved urgently.The selection and optimization of neural networks are the focus of current research.The continuous detection and improvement of detection accuracy in videos are also the focus of current algorithm optimization,and the selection and processing of datasets will also affect the efficiency of road detection algorithms,which is highly versatile.The datasets can improve the generalization ability of the model.This article analyzes the road detection algorithm,focusing on three aspects: vehicles,pedestrians and road surfaces.For vehicle targets,complete vehicle detection and tracking in the video.For pedestrian targets,obtain the joint position of the pedestrian in consecutive video frames,and predict the pedestrian’s posture.For road surface targets,in the image obtained by the sensor,the classification of the road surface is recognized.(1)In the continuous video frame,according to the front road information obtained by the driving recorder,the position of the vehicle is calibrated by the YOLO algorithm.Because in the video,all vehicles in each frame will be detected as a new target,so It is necessary to match the position of the vehicle in the previous frame with the position of the vehicle in the current frame.Complete the optimization of the serial number conversion of the vehicle,and solve the problem of being assigned a new serial number when the vehicle in the occluded situation is recognized again.(2)Aiming at the pedestrian target in the video,through the cropped VGG network,the detection of the joint points of the pedestrian is completed,and the confidence and affinity of each joint point are calculated according to the joint point position,and the relationship between each pair of joint points is calculated.These two parameters are matched to complete the connection of all related nodes of a single human body.All relevant nodes of each person are predicted through Kalman filtering to predict the position of all relevant nodes in the next frame,and the posture of the human body is judged according to the predicted position of the next frame.(3)Regarding the road ahead information acquired by the vehicle sensors,the largest proportion of the image is the road surface,so it is necessary to use an improved VGG network to complete the classification of the road surface.In order to improve feature reuse and effective information extraction,it is necessary to add a dense connection strategy between feature maps of the same size to improve feature reuse between feature maps of the same size.At different levels of the network structure,the features of each level are allowed to participate in the final result classification,so auxiliary classifiers are also added.At different stages of the model,auxiliary classifiers of different proportions are added to allow higher-level features to classify The important role of the lower-level features can also play a supporting role.Based on the identification and detection of road targets,the above algorithm can complete the tracking of vehicle targets,the warning of dangerous attitude of pedestrian targets and the classification of road objectives.The algorithm in this paper can promote the development of driverless and intelligent driving technology.