This content is not included in your SAE MOBILUS subscription, or you are not logged in.

Pedestrian Detection Method Based on Roadside Light Detection and Ranging

Journal Article
ISSN: 2574-0741, e-ISSN: 2574-075X
Published November 12, 2021 by SAE International in United States
Pedestrian Detection Method Based on Roadside Light Detection and Ranging
Citation: Gong, Z., Wang, Z., Zhou, B., Liu, W. et al., "Pedestrian Detection Method Based on Roadside Light Detection and Ranging," SAE Intl. J CAV 4(4):413-422, 2021,
Language: English


In recent years, to avoid the failure of the onboard perception system, intelligent vehicle infrastructure cooperative systems have been attracting attention in the field of autonomous vehicles. Using the perception technology of roadside sensors to provide supplementary traffic information for autonomous vehicles has become an increasing trend. Several roadside perception solutions select deep learning for three-dimensional (3D) object detection. However, deep learning methods have several issues and lack reliability in practical engineering applications. To tackle this challenge, this study proposes a pedestrian detection algorithm based on roadside Light Detection And Ranging (LiDAR) by combining traditional and deep learning algorithms. To meet real-time demand, Octree with region-of-interest (ROI) selection is introduced and improved to filter the background in each frame, which improves the clustering speed. Afterward, an improved Euclidean clustering algorithm was proposed by analyzing the scanning characteristics of LiDAR. Concretely, on account of the vertical and the horizontal angular resolution of the LiDAR, the authors propose a new method for determining the search radius of Euclidean clustering with adaptive distance. This algorithm can improve the problems of insufficient clustering of objects and the under-segmentation of neighboring objects. In addition, the authors’ method carries out an improvement of the sampling mechanism of Pointnet++ to accomplish the classification, and the classification average precision (AP) of Pointnet++ for sparse point clouds is improved. The AP for pedestrian object detection can reach 94.67%, which is higher than that of the other two networks. What’s more, the entire background filtering and clustering process takes 88.7 ms per frame, and the model obtained was deployed on NVIDIA Jetson AGX Xavier, attaining the inference time of 110 ms per frame, which can meet the speed requirement of LiDAR update and achieve the real-time application.