Feature Extraction based on Deep Learning for LIDAR Position Estimation of Autonomous Vehicles

摘要

In one embodiment, a method of extracting a point cloud feature used when estimating a position of an autonomous vehicle (ADV) includes the steps of selecting a first group key point from an online point cloud-the online point cloud is used for the prediction pose of the ADV. Created by ADV''s LiDAR device-; And extracting the first group feature descriptor for the first group key point using the feature learning neural network executed in the ADV. The method comprises the steps of determining a location for a second group key point in a pre-built point cloud map-each of the second group key points corresponds to a key point in the first group key point; Extracting a second group feature descriptor from a pre-built point cloud map; And estimating the position and direction of the ADV based on the first group feature descriptor, the second group feature descriptor, and the predicted pose of the ADV.