| Parking AGV(Automated Guided Vehicle)is a novel mobile robot and it is designed to solve the parking problem in urban development.Because parking AGV can save parking space and improve parking efficiency,it is more and more widely used in flat parking lots and stereo garages.However,drivers are required to park and adjust the pose of their vehicles on the AGV mechanical platforms,which would increase parking difficulty and parking risks,reduce parking experience,and limit the flexibility of parking.In this thesis,the scheme of vehicle pose detection system is determined,which consists of global vision system and local system of parking AGV.In order to ensure that the target vehicle is within the field of view of the sensors on parking AGV,the global camera is used to roughly estimate the pose of the target vehicle in parking buffer zone.Considering the safety of parking,the dynamic objects in parking buffer zone are detected based on the mixed Gaussian model.When the parking buffer zone is in a static state,the target vehicle contour is segmented.Based on the central axis of the vehicle contour,the convex parts such as rearview mirrors are filtered and the rectangle feature of the vehicle is extracted.The pose of the target vehicle in parking buffer zone is estimated based on the corner points of parking buffer zone border and target vehicle rectangle.Then,the pose of the parking AGV is adjusted and a rough alignment with the target vehicle has been done.Since the mounting height of sensors on parking AGV are very low,the traditional algorithms of vehicle pose estimation cannot be applied.Thus,a novel method is given based on the symmetry feature of wheel point cloud in this thesis and the feasibility of the scheme is verified based on 2D ladar.To obtain higher precision and more stable results,a further research is done based on 3D ladar.Since the vehicle point cloud is a part of the non-ground point cloud,the ground segmentation algorithm combining cell-based method and SVM(Support Vector Machine)classifier is proposed.After filtering ground point cloud,the wheel point cloud can be effectively extracted from the sector region of interest based on suspended object extraction algorithm and Euclidean clustering algorithm.The symmetry of the wheel point cloud is corrected by ICP(Iterative Closest Point)algorithm.Besides,in this thesis,SAC-IA(Sample Consensus Initial Alignment)algorithm is combined with ICP algorithm to register the wheel point cloud to target point cloud and the precision of the vehicle pose estimation method proposed in this thesis is evaluated based on the Euclidean fitness score of the registration.The experimental platform has been set up to verify the feasibility of the algorithms,and the human-computer interaction interface is designed with ROS(Robot Operating System)and QT. |
