Research And Implementation Of Road Anomaly Detection Algorithm Based On Deep Learning

Traffic accidents caused by abnormal road surface are a part of road safety problems that can not be ignored.Road depression,bumping and obstacles will not only pose a direct threat to the safety of people,but also bring inconvenience to people’s normal travel.During the process of human driving or automatic driving,if the vehicle system can detect the abnormal condition of the road as early as possible,it can remind the driver to make relevant response or the vehicle automatically make corresponding feedback,so in the auxiliary driving or automatic driving scenario,it can reduce the road safety ricks.Therefore,it is of great significance to research the road anomaly detection algorithm in driving scenes.Because of the complexity and diversity of the abnormal conditions of the road,we mainly detect three kinds of abnormal conditions of the road,such as potholes,deceleration strips and warning posts.In view of the application background of the above problems,based on the image semantic segmentation technology,we propose a road anomaly detection algorithm based on deep learning.The main contributions of this thesis are as follows:(1)In order to generate better semantic segmentation results,the idea of generative adversarial neural network is introduced into the design of the network architecture,and a new neural network architecture is proposed;(2)In order to make the semantic segmentation module output better segmentation mask and get more precise segmentation boundary,we propose a new edge loss function based on the original loss functions during the training process;(3)In order to verify the effectiveness of the road anomaly algorithm,we make a road anomaly data set for driving scenes;(4)We design a prototype system for the detection of road surface anomalies.Several experiments show that the algorithm proposed in this thesis can not only achieve the goal of real-time detection for road anomalies,but also have great advantages in accuracy,which proves the effectiveness of the algorithm.