Research On City Scenes Semantic Segmentation Based On Self-driving

With the rapid development of artificial intelligence technology,the autonomous driving of vehicles is getting closer and closer to people’s life.Autonomous driving first needs to rely on various sensors to collect the surrounding environment data,and then use various algorithms to analyze the environment information that can be perceived by the computer,and then use the perceived information to guide the decision-making of vehicle planning.However,most of the lidar and other sensors used by vehicles are expensive,which is not conducive to the large-scale popularization of autonomous vehicles.In comparison,the camera is cheap and can obtain a large amount of information about the surrounding environment.Therefore,it is of great significance to study the camera-based self-driving perception algorithm.Image semantic segmentation is one of the most important technologies in the perception of autonomous driving.The results of image semantic segmentation can be used to obtain the information of the driving area and obstacles in front of the vehicle.After the development of the deep learning and the convolutional neural network in recent years,many semantic segmentation algorithms based on deep learning have emerged,which can basically achieve end-to-end semantic segmentation output of images.But at present,there are still some problems in the actual use of the automatic driving system :(1)many algorithms cannot be run in real time,which cannot satisfy the automatic driving which need safety firstly;(2)the complexity and variety of scenes in the automatic driving scene are unbalanced in training data samples,and many algorithms are not good at detecting small samples.Therefore,there are still many problems to be solved in the semantic segmentation of images in self-driving city scenes.Therefore,this paper focuses on the real-time semantic segmentation algorithm of images in self-driving scenes.The main research of this paper is divided into two parts: the construction of lightweight real-time semantic segmentation model and the improvement of loss function for the sample disequilibrium problem.Main contents:(1)By analyzing various existing image semantic segmentation algorithms and numerous excellent neural network structures,and fully learning the advantages and exist problems of its network design,the network structure of the algorithm in this paper is built.Multi-branch structure,pyramid pool scale fusion and other structures are adopted to ensure the model has good characterization ability and real-time operation.(2)In view of the serious imbalance in the number of pixels between various types of samples in the automatic driving scene,the loss function is specifically improved to reduce the impact of the sample imbalance.At the same time,the multi-level loss function calculation method is added to improve the efficiency of training and make the model easier to converge.(3)The training was carried out on Cityscapes which is a public data set and a selfbuilt data sets.The training results showed that the algorithm in this paper could run in real time under the condition of input with large resolution,and the model accuracy was in the top of index evaluation.(4)In order to apply the algorithms to the embedded platform,the NVIDIA TensorRT method was applied to accelerate the model,making the trained model run in real time on the embedded platform.And do the experiment by using the real vehicle automatic driving in the actual urban traffic scene,the result shows that the algorithm can satisfy the practical...