Lidar-based Adaptive Pedestrian Pose Segmentation Method

Abstract

An adaptive pedestrian target poses segmentation algorithm based on point cloud alignment is proposed for pedestrian target pose segmentation, using which adaptive calibration of threshold parameters can be achieved. In this paper, we introduce a bilateral filtering method to remove outliers and reduce the influence of noise on the target point cloud. Using the Random Sample Consensus method to segment all planar point clouds, using an adaptive threshold calibration method based on the Fast Global Registration and point cloud precision alignment to improve the cluster segmentation module, that will achieve adaptive segmentation of pedestrian postures. Based on the LIDAR dataset with four pedestrian poses for experimental validation, the algorithm in this paper can precisely segment road pedestrian targets by poses. Compared with the traditional Euclidean clustering method and DBSCAN clustering method, the results respectively improve the accuracy of the algorithm in different poses by 1.61% and 2.75%, which proves the effectiveness and progressiveness of the adaptive clustering algorithm.