Heterogeneous Sensor’ Online Calibration And Mapping Relocation Based On Road Semantic Information

In recent years,benefits from the rapid development of sensors and the automotive industry,autonomous driving has become a research hotspot.Achieving relatively stable and accurate data conversion and pose estimation in the SLAM link at the front end of the technical process is a critical point in the entire autonomous driving task.The range of autonomous driving activities covers the outdoor environment of the whole city.The road conditions are complex.There are problems such as excessive storage,difficult maintenance of map,and light change.At the same time,it needs to run at the speed of people’s daily life.Therefore,it has a high pursuit of algorithm processing speed and safe driving.Aiming at improving stability and reducing the amount of calculation and storage,this thesis divides it into three subproblems: online calibration between lidar and camera based on semantic information,construction of lightweight semantic map,and internal relocation in a priori semantic map,and puts forward solutions.The main work and research results of this thesis are as follows:1.Online calibration based on semantic information: To solve the problem of errors in external parameters from offline calibration caused by frequent and irregular vehicle jitter,this thesis is proposed to use the correlation between the point cloud and the image semantic segmentation detection results to calibrate the external parameters between the lidar and the camera online.According to the analysis of the representation form and detection accuracy of point cloud and image semantic segmentation results,the algorithm constructs the data association between sensors and designs three errors for different feature semantic information.It uses the position transformation between the before and after frames to realize the reprojection of the point cloud between multiple frames,and proposes three inter-frame optimizer design schemes that can optimize the multi-frame external parameters simultaneously.In the actual scene,three kinds of error ablation experiments and optimizer scheme comparison experiments are carried out,and the calibration results are quantitatively and qualitatively analyzed to verify the effectiveness of the algorithm,and the accuracy is higher than the existing calibration methods of the cooperative company.2.Construction of lightweight semantic map: This thesis proposes an algorithm for con-structing lightweight semantic maps to overcome the huge challenges of map storage and maintainability when autonomous driving operates in large-scale scenes.Compared with dense maps,lightweight semantic maps can only store the structural information of semantic objects.With the 3D information supplement of the original lidar data to the image detection,the reconstruction of the 3D model of the global lane lines,signs,and light poles is realized,and the key features are saved.The experimental results show that the constructed lightweight semantic map has high accuracy and can be used for follow-up relocation research.3.Internal relocation in a priori semantic map: To improve the robustness and accuracy of the localization algorithm in complex and changeable scenes,an algorithm of relocalization in the prior map mentioned above using semantic information is proposed.Combined with the rough pose estimation of the inertial navigation system,the algorithm designs the correlation method between the semantic detection results of each frame and the prior map.The correlation result is used as a global measurement to update the pose estimation in the optimizer,which effectively avoids the sensor failure or pose jump in sparse scenarios and realizes the accurate estimation of the vehicle pose.The accuracy and robustness are tested on the prior map constructed above,and the algorithm can achieve better localization results than common open-source frameworks.