Research On Structure Road Detection And Tracking Algorithm Based On Image

When the number of vehicles is increasing around the world,traffic accidents are also increasing.With the threat to property and safety posed by traffic accident,the assisted driving system and the unmanned system get more and more attention.As one of the key technologies of the two systems,road detection has been widely studied by researchers.The current research mainly focuses on the detection of structured roads.The structure road is the roads with lane markings or clearly boundary,such as urban roads and highway.At present,the general process of structured road detection includes feature extraction,road model establishment and road detection and tracking.Firstly,feature extraction is performed in bird's-eye image in which the road boundaries are parallel.Feature extraction results often contain many abnormal points and this paper proposes a method to remove the abnormal points to reduce its' influence on road detection.Secondly,a mathematical model is established to describe the boundary shape of the road.The model used in this paper does not describe the geometric shape of the road boundary,but rather describes the relationship between road boundary points.So it can adapt to different shapes of road boundary.Thirdly,the road is detected by the particle filter and the road detection result is two sets of road boundary points.This paper establishes a low-dimensional particle state using the distance information which effectively reduces the number of particles needed in the particle filter and improves the efficiency of particle filter.At the same time,in addition to taking into account the distance information,the observation model in this paper accounts the detection results of the previous states to improve the accuracy of road detection.When tracking the road,the detection result of the previous frame is used to provide more information for the road detection of the current frame.The method proposed in this paper is tested.The detection accuracy of the method is tested in the caltech lane dataset.And the running time of the method is tested in the embedded devices.