AgSentry: LiDAR Reflectivity Based Obstacle Detection for Safety in Agricultural Environments

LiDAR is widely used for safety in agricultural robots, but the low resolution of cost-effective sensors limits long-range obstacle detection. This paper investigates how LiDAR reflectivity can be leveraged to improve safety perception on the ARWAC autonomous ground vehicle. We first evaluate a voxel-based emergency stop (E-STOP) system, quantifying braking distance, perception range, and processing delay to establish a reliable safety baseline. To enhance long-range obstacle detection, we propose a deep-learning pipeline that performs segmentation on LiDAR reflectivity images and projects the results onto corresponding 3D point clouds. We introduce AgSentry, an 850-image reflectivity dataset with five annotated obstacle classes common on farms. A controlled cross-architecture comparison of one-stage (YOLOv9), two-stage (Mask R-CNN), and transformer-based (RF-DETR) detectors demonstrates that reflectivity-based models consistently outperform their depth-based counterparts, with RF-DETR achieving the strongest overall results (mAP@50-95 of 29.88 vs. 24.53 for depth). The advantage is most pronounced for safety-critical classes: human detection mAP nearly doubles with reflectivity input (27.35 vs. 15.19), while vehicle detection improves by over 70%. These results demonstrate that reflectivity is an effective modality for improving long-distance safety perception in agricultural environments and motivate further integration with formal functional safety frameworks.