Road Obstacle Detection Algorithm And Its Augmented Reality Application Based On Vision

With the continuous development of science and technology,Advanced Driver Assistant System(ADAS)has gradually begun to spread to people's lives.ADAS is mainly devoted to improving drivers'perception of road information,reminding drivers'potential threats in the road in time,and reducing the incidence of accidents.Therefore,ADAS has a high requirement for the accuracy and real-time performance of road obstacle detection.On the basis of the research of traditional obstacle detection algorithm,this paper improves the traditional algorithm with the goal of accuracy and real-time performance.The main focus of this study is in the following aspects:1.This paper first describes the current research status of obstacle detection and Augmented Reality(AR)at home and abroad,determines the research object of this paper,and gives a detailed introduction to the relevant basic theories involved in this paper.At the same time,we studied the algorithm of image preprocessing,including algorithm of region of interest extraction,image grayscale and filtering,and analyzed the advantages and disadvantages of each algorithm based on robustness and real-time.2.In order to improve the real-time and accuracy of the system,this paper uses a hypothesis testing method to detect road obstacles.In terms of candidate area generation,in order to solve the traditional threshold segmentation algorithm in the segmentation of road and obstacles,an improved adaptive threshold segmentation algorithm based on road grayscale histogram is used to segment obstacles and road background.For the divided two value images,different sizes of candidate regions are generated according to the obstacles of different widths.Finally,the robustness and feasibility of the proposed algorithm are verified by comparison experiments.3.In the area of candidate region verification,this paper proposes an obstacle validation method based on convolution neural networks.First,according to the environment studied in this paper,we optimize the parameters of AlexNet network model.Secondly,the Batch Normalization(BN)is introduced into AlexNet network model to improve the accuracy of the model.Finally,the training method and sample data processing of the network model are introduced.The experiment shows that after the introduction of the batch normalization layer,the recognition rate of the network model is increased to 99.1%,compared with the original model by 0.9%,the network convergence error value is reduced by 52.50%,and the average image detection for each frame is 100ms.4.In order to remind the driver of obstacle information in the road ahead,this paper studies augmented reality virtual registration technology.A tracking and registration algorithm based on road scenes for augmented reality is proposed,which improves the BRISK algorithm and describes the feature points by BRIEF algorithm.On the basis of detecting obstacle targets,we track the obstacles in subsequent 3 frames and register the virtual warning signs.The experimental results show that the improved algorithm improves both the robustness of feature point matching and the real-time performance of the algorithm.The algorithm takes about 49ms time and reduces the time compared with the original algorithm by 23.43%.After experimental verification,the road obstacle detection algorithm is designed in this paper achieves the detection of vehicles and pedestrian road obstacles,and can meet the accuracy and real-time requirements of intelligent vehicle auxiliary driving system,providing a certain reference value for the research of intelligent vehicle driving assistant technology.