Detection And Recognition Of Urban Intersection Based On Vision

The detection and recognition of urban intersections belong to the scene understanding problem,which is one of the challenges that Unmanned Ground Vehicle(UGV)is faced with.The complexity and variety of intersection scene set higher demands for self-driving technology.However,combining the positioning system and HD map is faced with occa-sional position error,especially when signals are blocked.Therefore,on-board sensor based intersection detection and recognition method is a significant and important research for the improving of self-driving in urban scene.In this paper,a visual feature and probabilistic model inference based intersection detection and recognition method is proposed utilizing an on-board camera.The aim of the proposed method is to assist UGV for online intersection detection and recognition,reduction of the positioning deviation and improving perceptibility.The paper is demonstrated in overall structure,image features analysis and intersection probability inference.Firstly,a intersection model is established,which includes the geometry and components.The model composition can be divided into static features and dynamic features.The static features include the traversable directions(numbers and distribution),arrow-road markings and traffic lights.The dynamic feature is short tracklets of moving vehicles.The proposed method uses a monocular color image as input for analysis of static-dynamic features.In the field of image processing,deep learning technology is widely used because it is quicker and more accurate than traditional methods in feature extraction.In the aspect of static image feature analysis,pixel-level semantic segmentation is firstly carried out,and a two-dimensional dense conditional random field is introduced to optimize the semantic segmentation results by combining features like the position and color of pixels.Then,the inverse projection transformation and skeleton extraction are carried out to extract traversable directions.At the same time,the arrow-road markings and traffic lights are detected and classified to assist in intersection detection.Dynamic image features are short tracklets of moving vehicles.The bounding boxes and orientations of vehicle objects are detected and classified,followed by continuous frame tracking.In this paper,the Hungarian algorithm is used to achieve vehicle objects matching between adjacent frames.Based on the continuous movement trend of tracklets,the potential traversable directions are analyzed.Hidden Markov Model is introduced to combine the scene image features with intersection model.The scene image features are regarded as observable variables and the intersection types are regarded as implicit variables,and they are integrated to infer the probability and classify the type.The training of deep neural network is based on the open dataset KITTI,and the experiment and test are based on actual scene dataset collected by IN~2Bot.In order to verify the validity and evaluate the accuracy,online experiments and real scene vehicle experiments are carried out,and the experimental results have proven the effectiveness of the proposed method.