Research On Fused Omnidirectional 3D Vision Theory And Application Of Lane Detection And Localization

This dissertation is grounded upon a major research plan of the National Natural Science Foundation of China-"Cognitive Computing for Audio-Visual Information".Combining the omnidirectional camera and lidar with 360 degree field of view,the panoramic images and accurate 3D position of environments are provided for Unmanned Ground Vehicle(UGV).The lane detection and localization is researched based on omni-directional 3D visual data.The main results and innovations of the thesis are summarized as follows:1.The effect of inverse perspective transformation on lane marking feature extrac-tion is explored and the performance of the proposed three lane marking feature extraction methods based on omnidirectional image are evaluated on original and bird's-eye images.Experimental results on omnidirectional image databases show that the performance of the same lane marking feature extraction method on bird's-eye images is better than it on original images from the sensitive,precision,robustness,best possible performance and average accuracy.2.A lane detection method based on distorted omnidirectional image is presented.The parameters of lane in the vehicle coordinate system can be estimated directly by fit-ting curves in omnidirectional image coordinate system.Experimental results show the validity of lane detection on distorted omnidirectional images is demonstrated.3.A lane model fitting method based on the parameters optimization of lane model is presented to fit straight lane and curve lane in uniform lane model.According to the parabola lane model,the straight lines and curves of feature maps can be represented as straight lines in a linear space coordinate system,in which the straight lines and curves are described using straight lines.Then lane model fitting can be treated as an optimization question in linear space and the parameters of lanes can be estimated by minimizing the objection function.The proposed lane detection methods based on the omnidirectional data are tested on omnidirectional image datasets and public available monocular image datasets,and long distance experimental results show the accuracy rates of proposed lane detection method reach 99%.4.A multiple map localization algorithm using appearance map and 3D feature map is proposed.Localization based on map can be divided two steps.Firstly,the coarse localization is realized using the appearance map and localization based on 3D feature map is acquired by combining the results of coarse localization.Experimental results show that the localization performance of appearance map based on position distance is better than it constructed according to the feature distance and the accuracy of localization of appearance map is improved using the 3D feature map.