Comparison of different LiDAR sensors for safe object detection using deep learning algorithm

Abstract

This study investigates the use of different light detection and ranging (LiDAR) sensors for object detection tasks using deep learning algorithms in autonomous driving applications. Three LiDAR sensors - LS LiDAR, Livox, and Ouster - were tested by collecting point cloud data from various road scenes involving cars and pedestrians. The data was labelled using MATLAB’s Ground Truth Labeler and used to train a Complex YOLO-V4 neural network model. The performance of the trained model was evaluated on test data from each sensor using mean Intersection over Union (IoU) scores, Average Orientation Similarity (AOS) and Average Precision (AP) metrics. Results showed that LS LiDAR achieved a mean IoU of 0.322 for cars but 0.229 for pedestrians, while Livox scored 0.397 and 0.265 respectively. Ouster had the best results with 0.471 for cars and 0.332 for pedestrians, demonstrating its strong object classification capabilities. Point clouds from Ouster also exhibited higher localization performance compared to other sensors based on IoU value graphs. Ouster’s high-resolution 3D point clouds worked optimally with the YOLO-V4 model to achieve the highest accuracy for both vehicle and pedestrian detection among the three LiDARs tested. The study provides insights into selecting the appropriate LiDAR sensor for autonomous driving applications based on object detection performance.