Research On Road Detection In Complex Scene

Semantic segmentation is an active application and research field in the field of computer vision technology.The main task is to assign the labels of each pixel in the image.Using semantic segmentation algorithm to detect scene in road images can effectively help autopilot system to understand and analyze its environment,so as to provide a prerequisite for the application of driverless system.Therefore,this paper focuses on the research of road scene detection,the main contributions are the following two aspects:In view of the existing road semantic segmentation model spatial information is not rich enough,the feeling field is not big enough and so on.This paper studies the semantic segmentation algorithm of road image based on deep learning.By analyzing the research status of image semantic segmentation,a dual pyramid segmentation model DPNet is proposed.Specifically,the design and modification of encoder and decoder are introduced in detail,based on which DPNet is constructed.A large number of experiments are designed to verify the effectiveness of the design and DPNet's performance in Cityscapes data set,Camvid data set and security monitoring data set is reported.In view of the difficulty in detecting the vanishing point caused by the occlusion of pedestrian and driving on the road boundary line in the complex road scene,and the difficulty in case segmentation and under segmentation,an improved case segmentation road detection method is proposed from the perspective of scene construction.Firstly,the road area is extracted by case segmentation,and then the crowning algorithm is used to compensate the occlusion of pedestrians and traffic on the road boundary line.Finally,the crowning algorithm is used to synthesize the trapezoid model which conforms to the scene structure,so as to optimize the detection of the road.To sum up,this paper provides two ideas for the road detection task in complex scenes: one is to improve the existing road semantic segmentation model from the network architecture level to improve the accuracy of road detection;the other is to propose a trapezoid model in line with the complex road scene from the scene structure level to optimize the detection of complex roads.The method in this paper has been experimentally tested on three real scene road datasets and compared with a variety of existing cutting-edge road detection algorithms.Experimental test results show that thedetection algorithm in this paper has relatively high robustness and accuracy in a variety of complex situations.Compared with the existing detection algorithms,the experimental results are more reliable and have practical significance.