Localization And Mapping On Urban Area Based On 3D Point Cloud Of Autonomous Vehicles

Unmanned Vehicles are the trending research subject in present intelligent robotic field. Localizing the vehicle in unknown environment and mapping the surrounding environment by simultaneous localization and mapping(SLAM) is an important way to navigation the vehicle, which enables the unmanned vehicle to complete the designed task safely and quickly in the unknown environment. The urban environment is the main driving environment for most vehicles and it’s also an important research direction of future unmanned vehicles. It’s of great value to study on replacing existing vehicles with unmanned vehicle and improving driving experiences.The problem of how to use SLAM for unmanned vehicle in urban environment based on vehicle-mounted 3D Lidar is proposed in this paper. The above problem is solved by preprocessing of the data, the estimate of the attitude of the vehicles, mapping the unknown environment. The contribution of this paper mainly includes:(1) The 3D Lidar data has the property of high density in near environment and low density in distant environment, which make it inaccurate to describe the environment. Moreover, the data is of high frequency and can’t reveal the structural property of the ground. Based on the above property, a preprocessing of the 3D Lidar is proposed in this paper to improve the effectiveness and the computation performances of the algorithms. The data is sampled with the same time interval to reduce the data volume. The point cloud is segmented with voxel and the data in each voxel is sampled with VoxelGrid filter. Finally, the effective point cloud area is generated by segmenting the sampled point cloud data with thresholds to solve the problem that the Lidar data in distant area is too sparse to represent the environment and the lack of structural information for the ground.(2) Based on the classical cloud point matching algorithm- iterative closest point(ICP) algorithm, the transformation matrix and the rotation vector between neighboring data frame is calculated and the rough vehicle attitude is also estimated. However, the ICP algorithm is sensitive to the initial value and an inappropriate initial value will result in bigger iteration times, even a local optimal value which makes the algorithm ineffective. To solve the above problem, the conversion matrix between the neighboring data frame is estimated based on extended Gaussian image is regarded as the initial value of the ICP algorithms. Moreover, the estimate of the vehicle attitude with ICP may be not the real attitude. But it should locate in the high-probability area of the real attitude. Thus just a few particles can cover the valid area by using Gaussian model and update the particles.(3) Quantifying the matching degree of the local map and the global map for each particle. An importance weight is given to each particle to evaluate its performance. Furthermore, the particle with the best performances is regarded as the estimate of the unmanned vehicles and the local map represented by this particle is utilized to update the global map. The matching degree of the particle is calculated by rasterizing the overlap region of the global and the local map represented by the particle. Then the property of empty, scatter, horizontal linear, vertical linear, horizontal plane and vertical plane are given to each voxel based on PCA and GMM. Finally, the respective voxel are matched with each other and the matching degree is considered as an important weight of the particle.(4) The effectiveness of the algorithms is evaluated with experiments with the unmanned platform. The experiments are conducted in unstructured urban environment, simple structured urban environment and complex structured urban environment. For all above mentioned environment, the proposed algorithm in this paper can follow the trajectory of the unmanned vehicle and build the accurate environment map, which proves the effectiveness of the algorithms.