Mapping And Localization Method Based On Visual Semantic Information In Indoor Parking Lots

High-accuracy mapping and localization are among the most crucial techniques in the automated valet parking system.The fact that the large amount of disturbance of vehicles and pedestrian in the parking lot leads to the frequent updates of 3D map,which tremendously enlarges the burden of computation resources.Besides,the constrained environment of the global navigation satellite system further increases difficulties for the implementation of the low-cost,high-precision localization.Considering the above factors,this paper incorporates fisheye cameras,inertial measurement units,wheel encoders and proposes an approach based on the semantic information of ground markings for high-accuracy mapping and localization,which implements real-time,robust localization with high accuracy.The high-accuracy mapping and localization system in this article mainly contains three parts: mapping,global initialization and localization.The main contributions of this paper are as follows:Since the misclassification error is detrimental to the precision of the map constructed in the mapping process,this paper designs a semantic point cloud fusion method that exploits multiple frames to correct the labels of map points and hence reduce the effect of segmentation noise.To eliminate the accumulated drift caused by the long-term local dead reckoning,this paper proposes to leverages both semantic category information and visual feature descriptors in the process of loop-closure.By matching the sub-map to verify the loop-closure results,the detection performance can be further improved.Furthermore,to ameliorate the problem that the optimization methods for the global pose diagram is not capable of ensuring the consistency of the point cloud,this paper revises covariance matrix of the global pose graph and constructs globally consistent,precise maps of the ground markings.The empirical experiments in the real-world indoor parking lots verify that the mapping methods proposed can guarantee the maps’ precision and consistency.For the global initialization based on the ground landmark,it is difficult to guarantee the performance of real-time and accuracy at the same time.Combining particle filtering and descriptor matching based on semantic information,this thesis designs an effective algorithm to achieve a fast and accurate global initialization.By integrating the results from the descriptor matching into the particle filtering framework,the algorithm,making full use of the high precision of the particle filtering and the fast speed of descriptor matching,ensures the performance of the global initialization and simultaneously improves the algorithm efficiency.The experiments in the real-world indoor parking lots demonstrate the fast global initialization method proposed have both favorable real-time and accuracy performance.Aiming at the problems of demand of higher localization accuracy during the parking process,and low localization accuracy when the ground landmark is blocked,this paper proposes a point cloud registration approach that combines semantic and geometric inference so as to improve the registration accuracy of point clouds.Moreover,by integrating inertial measurement unit and wheel encoder measurements,the split covariance intersection filter is applied to fuse the global poses from multiple data sources to obtain global localization results with higher accuracy.While running on the embedded platform in real-time performance,the mean square error of the localization process is less than 10 cm,which demonstrates the algorithm proposed in the paper is capable of meeting the localization demand of automated parking system in the indoor parking lots.