Edge Computing Platform-Based Farmland Obstacle Detection and Distance Estimation System Using 3d Lidar and Rgb Camera

Abstract

High precision detection of obstacles and accurate estimation of their distances are essential for the autonomous operation of agricultural machinery. Using a single sensor typically cannot simultaneously detect obstacles and estimate their distance. Currently, methods based on multi-sensors fusion are utilized predominantly on deployed on large computing platforms and are difficult to be applied to mobile robots. To address these issues, this paper proposes obstacle detection and distance estimation utilizing the Jetson Xavier NX edge computing platform. This approach combines a 3D LiDAR and an RGB camera to detect obstacles and estimate their distances. The method employs an improved YOLO v5s model for obstacle detection and optimized point cloud data for distance estimation. Experimental validation reveals that this improved YOLO v5s model achieves high detection accuracy while significantly reducing the computational demands, floating-point operations and parameters are lowered by 48.05% and 52.05%, respectively. Following the optimization of the enhanced YOLO v5 model and point cloud processing, the proposed detection and estimation method attains a processing speed of 22 frames/second on the Jetson Xavier NX platform, satisfying real-time operation requirements. The system detects farmland obstacles at distances exceeding 17 m, ensuring the safe functioning of agricultural machinery in field environments. The proposed system leverages the combined capabilities of 3D LiDAR and an RGB camera and offers the advantages of compact size, affordability, and real-time performance, fulfilling practical application needs in agricultural settings.

Keywords: Farmland obstacle detection, Distance estimate, Sensor fusion, Improved YOLO v5s, 3D LiDAR