With advances in precision agriculture, autonomous agricultural machines can reduce human labor, optimize workflow, and increase productivity. Accurate and reliable obstacle-detection and avoidance systems are essential for ensuring the safety of automated agricultural machines. Existing LiDAR-based obstacle detection methods for the farmland environment process the point clouds via manually designed features, which is time-consuming, labor-intensive, and weak in terms of generalization. In contrast, deep learning has a powerful ability to learn features autonomously. In this study, we attempted to apply deep learning in LiDAR-based 3D obstacle detection for the farmland environment. In terms of perception hardware, we established a data acquisition platform including LiDAR, a camera, and a GNSS/INS on the agricultural machine. In terms of perception method, considering the different agricultural conditions, we used our datasets to train an effective 3D obstacle detector, known as Focal Voxel R-CNN. We used focal sparse convolution to replace the original 3D sparse convolution because of its adaptable ability to extract effective features from sparse point cloud data. Specifically, a branch of submanifold sparse convolution was added to the upstream of the backbone convolution network; this adds weight to the foreground point and retains more valuable information. In comparison with Voxel R-CNN, the proposed Focal Voxel R-CNN significantly improves the detection performance for small objects, and the AP in the pedestrian class increased from 89.04% to 92.89%. The results show that our model obtains an mAP of 91.43%, which is 3.36% higher than the base model. The detection speed is 28.57 FPS, which is 4.18 FPS faster than the base model. The experiments show the effectiveness of our model, which can provide a more reliable obstacle detection model for autonomous agricultural machines.