LiDAR-Based Human Detection for Agricultural Robots Using Deep Learning

Abstract

Six deep-learning-based obstacle-detection models were developed and evaluated using LiDAR point clouds to assess agricultural robot safety. The models automatically extract object features and provide both classification and distance-estimation end-to-end. LiDAR not only enables object detection but also facilitates type and precise location identification. The best-performing model, Voxel R-CNN, achieved an average precision of 99.55 % at an Intersection over Union (IoU) threshold of 0.5, with inference and testing speeds of 20.63 and 7.39 FPS, respectively. A static test at fixed distances of 3-20 m demonstrated a distance-estimation RMSE below 4.4 cm. Combined with the high detection accuracy of the best-performing model, these results support the potential use of LiDAR-based human detection in agricultural environments.