Research On Intelligent Segmentation Of Zebra Crossing And Automatic Discrimination Of Pedestrian Position

Domestic traffic accidents occur frequently.Most of them are due to the friction between pedestrians and vehicles at traffic intersections.The scene understanding of traffic intersections is necessary for traffic safety.Zebra crossing area is one of the important components of traffic intersections,and is the area in which most interactions between pedestrians and vehicles exist.It is very important for both pedestrians and vehicles to understand the scene of zebra crossing area well.The understanding of zebra crossing area often stays at the level of segmentation or detection,with no further analysis and application of the segmented region.Aiming to solve this problem,this paper studies the zebra crossing area segmentation and its applications.The scheme includes the zebra crossing area segmentation optimization,zebra crossing area scope definition and automatic pedestrian position discrimination based on binocular stereo vision.The main work includes:(1)Firstly,a data set of zebra crossing characteristics is made,which is the one with China urban characteristics for the following research.A variety of semantic segmentation algorithms based on convolutional neural network are analyzed and studied.By experimental comparison,Unet is selected and is optimized by the feature fusion of deep layer information and shallow layer information.An improved CSUnet semantic segmentation algorithm based on channel space feature fusion module is proposed,which effectively improves the segmentation performance of zebra crossing area.(2)Based on the proposed CS-Unet and the zebra line specification,an automatic pedestrian location discrimination system for zebra line areas is designed according to the stereo vision principle.The zebra line area is semantically segmented,and the spatial location range of each zebra line is obtained by stereo matching and zebra line specification.Then,the foot area of the pedestrian is semantically segmented,and the spatial location information of the foot is obtained by stereo matching.An algorithm is designed to identify the specific location of the pedestrian in the zebra line area.And the system is verified by experiments.The experimental results show that the proposed CS-Unet can effectively improve the accuracy of the model with little increase in model parameters.Compared with the original model,the MIOU increased from 94.11% to 95.44% and MPA from 97.07% to 97.84%.The intelligent zebra line area partitioning and automatic pedestrian location discrimination system proposed in this paper can accurately determine the specific location of the pedestrian in the zebra line area,which provides important information for the traffic management.