Urban Road Understanding Based On Local Spatiotemporal Cues Integration

In recent years,with the development of intelligent driving technology,more research institutions and automobile enterprises pay attention to self-driving vehicles.Road detection is not only a key problem but also a difficult problem for self-driving vehicles' environment perception system.Traditional road detection methods take the lane and the road fence as the main cues.However,these cues would be fragmented with the presence of other traffic participants.For this predicament,this paper studies road detection based on local spatiotemporal cues integration.The main contributions are as follows:1.We design and implement a kind of road boundary detection method based on local geometric feature.Firstly,the calculation and expression methods of local surface normal vector based on 3D point cloud are analyzed,and then this paper implements a road boundary detection method based on the characteristic of surface normal vector mutation.This method promotes the application of stereo vision technology in road recognition.2.In this paper,we propose a algorithm of road environment understanding which can effectively integrate local geometric features and texture features.The Bayes method is used to transform the shape a priori,the surface normal vector,the height and the color feature into the likelihood probability of the road,and forms the comprehensive evaluation of the edge region.Finally,the multi-core support vector machine technology is used to shape the local edge candidate regions,which forms the overall understanding of the road.The experimental results manifest that our algorithm has characteristics of low computation,wide adaptability and high precision.3.Based on other vehicle behavior sequences,this paper proposes a road prediction algorithm.In this algorithm,the hidden relationship between vehicle behavior and road direction is excavated,and the space velocity cues such as velocity,direction and position of the vehicle around the vehicle are integrated.Finally,the understanding of the relationship of the dense traffic flow is formed in the road probability raster map.Experiments show that the method can realize the lane prediction and the lane-level positioning of the unmanned vehicle under the intensive traffic flow.