Research On Detecting Passengers On Two-wheel Vehicle And Counting Traffic Flow Based On Faster R-CNN

In the intelligent transportation system,two-wheel vehicle has become an indispensable means of transportation.However,the illegal driving of two-wheel vehicle,especially the irregular carrying of people,is very easy to cause traffic accidents.So the detection of passengers on two-wheel vehicle is very important.The traditional detection method is realized by manual extracting feature and matching target,which is not suitable for the complex traffic scene;while the detection method based on deep learning can automatically learn the target features in different environments,with strong generalization ability,which is suitable for the detection task in complex traffic environment.In this paper,the Faster R-CNN algorithm is used to detect the two-wheel vehicle and passengers on the vehicle,and then the vehicle flow statistics is carried out through the vehicle tracking and counting method.Finally,the Django application framework is used to build platform of detecting passengers on two-wheel vehicle and counting traffic volume.The specific research contents are as follows:1.The Fast R-CNN algorithm is selected as the detection model for the detection of passengers on two-wheel vehicle.The original Fast R-CNN algorithm is not effective for the detection of passengers in traffic pictures.There are two main reason.One is that the pixel size of the passenger's head is too small,which leads to the missed detection of the detection algorithm.The other is that the occlusion between multiple passengers on the two-wheel vehicle will also increase the difficulty of the detection algorithm.In this paper,the size of the anchor frame is modified with reference to the experimental data set.For different size targets,the detection model adopts multi feature fusion structure.Due to the missing detection of passengers with high overlap,the Fast R-CNN algorithm selects Soft-NMS to filter the candidate boxes.Finally,the experiment verifies the excellence of the improved Faster R-CNN in the detection of two wheeled vehicles and passengers.2.Vehicle tracking and statistics solves the tracking and counting problem of two-wheel vehicles in the video.In the aspect of counting traffic volume of two-wheel vehicle,this paper uses the mean shift algorithm to track the two-wheel vehicles in the video,and then applies the target area counting method to count the traffic volume.In view of the problems of target tracking loss and target tracking error in mean shift tracking algorithm,the target model is updated with Fast R-CNN algorithm,and the effectiveness of the improved method of tracking and counting two-wheel vehicle is verified by experiments.3.According to the different detection requirements of traffic video and pictures,the platform of detecting passengers on two-wheel vehicle and counting traffic flow constructs the traffic video detection function module and user-defined function module.The platform uses Django application framework and neo4 j database.In the experiments,the platform can realize the detection of passengers on two-wheel vehicle and the counting traffic flow in the traffic video,and also can provide users with services of picture detection and data query online.