A method and device for global initial positioning based on multi-hypothesis tracking

Abstract

The invention belongs to the technical field of global initial positioning, and relates to a global initial positioning method and device based on multi-hypothesis tracking. The method first collects point cloud data of a scene and its corresponding global pose data, constructs a database, and passes the collected data through Preprocessing is stored in the database, and a point cloud map is generated at the same time, and then the point cloud data of the current frame is matched with the data in the database to obtain matching candidate results, and then the candidate results are filtered by the constraint relationship between consecutive frames until the candidate results are obtained. After the results converge, the current frame data and the data in the corresponding database are used to calculate the translation and rotation, so as to obtain the initial value of the global initial positioning, and then match it with the point cloud map to obtain the accurate initial positioning result. The method of the invention combines global description and multi-frame data to perform initial scene localization, and adopts the method of multi-hypothesis tracking to screen the results, which ensures the efficiency of localization and improves the robustness of localization.

Description

technical field [0001] The invention belongs to the technical field of global initial positioning, and relates to a global initial positioning method and device based on multi-hypothesis tracking. Background technique [0002] In driverless positioning, global initial positioning is one of the essential links. In outdoor scenes, GPS is generally used for initial positioning. However, in indoor scenes such as underground garages, there is no GPS signal that can be used for positioning. Therefore, scene recognition and positioning are performed based on vision or point cloud. The classic vision-based methods include "Bags of binarywords for fast place recognition in image sequences", but vision-based methods are sensitive to illumination and perspective, and are only suitable for use in some scenarios. Point cloud-based methods often use local descriptors or global descriptors for scene localization. Local descriptor-based methods include "Fast point feature histograms for 3d...

AI Technical Summary

Problems solved by technology

The classic vision-based method includes "Bags of binary words for fast place recognition in image sequences", but the vision-based method is sensitive to illumination and viewing angle, and is only suitable for use in some scenes

Point cloud-based methods often use local descriptors or global descriptors for scene positioning. Local descriptor-based methods include "Fast point feature histograms for 3d registration" and "Place recognition using keypoint voting in large 3d lidardataset", which have high precision. , but the proposed descriptor has a large dimension and the retrieval efficiency is very low

Methods based on global descriptors include "Scan context: Egocentric spatial descriptor for place recognition within3d point cloud map" and "Intensity Scan Context: Coding Intensity and GeometryRelations for Loop Closure Detection". sensitive to distance

The above methods are all based on single-frame data for scene positioning, which is susceptible to interference from similar scenes, and the positioning results are not robust

Images