Research And Application Of Monocular Vision For Navigation On Intelligent Vehicles

Positioning and navigation are integral parts of intelligent driving,in which visual simultaneous localization and mapping is a paramount and core technology.Based on vision sensors,it estimates position of vehicles and establishes a large-scale 3D driving scene.It is one of the research hotspots in the current field,which can decrease the cost on the premise of bring out localization and mapping.In this paper,monocular visual sensor is adopted to study the application of monocular visual SLAM system on intelligent vehicles.The main contents are as follows:1.Constructing monocular visual SLAM model on intelligent vehicles and parameterizing image information and estimating camera pose.Pose projection model and inverse depth parameterization model of images are established to parameterize images.The new image is transformed into the radiation value and its image pyramid is layered.A method of five-frame initial optimization is proposed to enhance its success and smooth inverse depth.It is established that a photometric error model of two frames based on direct method and the camera pose is estimated by an iterative method combined with the optimization of pixel intensity,which contribute to positioning and location tracking of vehicles.2.Calling and verifying keyframes and observation points reciprocally in the front-end location tracking.The polar line search of the point to be observed is performed in key frames to minimize the photometric error,whose inverse depth determines whether it can be converted to an observation point.Used key frames,camera pose,observation points,photometric errors are further optimized and updated cumulatively based on Gauss-Netwon method,which calls each other with the back-end optimization.3.Optimizing the camera pose and photometric error cyclically at the back-end based on the sliding window and graph optimization method combined with the latitude and longitude information of the vehicle.The algorithm is designed to traverse and linearize the photometric error of map points.And the system photometric error is calculated cumulatively.In addition,the map positioning points in keyframes are coordinate-converted by matrix multiplication to create a 3D point cloud map and the location verification is performed corresponding to the latitude and longitude information of the vehicle by the time line.4.Setting up an experimental platform to verify the monocular vision SLAM system in this paper.Functional experiments such as image preprocessing and observation point recognition are carried out and promising.Furthermore,offline video verification is examined.Eventually,the low-speed vehicle experiment in the test field is fulfilled to verify the locationing accuracy and the practicability of 3D point cloud reconstruction.In conclusion,the system can run well on the intelligent vehicle platform in real time and realize the navigation route tracking of the vehicle and 3D map reconstruction.It demonstrates good real-time,robustness and positioning accuracy.