Abstract: In response to the current needs of farmland obstacle detection in the automatic driving of agricultural machinery, a method of using three-dimensional LiDAR to detect field obstacles was proposed. Firstly, the collected environmental point cloud was preprocessed. The voxel grid down-sampling method was used to filter the dense point cloud without losing feature information. A bounding box was used to segment the region of interest for fast calculation. The random sample consensus algorithm（RANSAC） was used to detect the farmland ground, and the ground point cloud was removed from the whole point data so that the obstacle points were extracted. Then the obstacle point cloud was clustered by Euclidean distance based on the K-d tree, and the distance threshold of clustering was 0.6 m in this test. Finally, the size of the cluster and the volume of the circumscribed cuboid were judged, and invalid clusters that were too large or too small were filtered out to obtain obstacles. A LiDAR with 32 channels was used to collect field obstacle point cloud at National Experiment Station for Precision Agriculture in Beijing Xiaotangshan. The algorithm was used to detect agricultural implement, haystack, field ridge, low houses, roadside trees, and field pedestrian. The test showed that the algorithm was suitable for the field common obstacles detection. When detecting pedestrians in the field, the people crossed the front view of the LiDAR and the distances from the LiDAR respectively were 5 m, 10 m, 15 m, 20 m, 25 m and 30 m to test the effect of the algorithm at different distances. The results showed that the average detection rate of dynamically walking people in the field within 30 m was 96.11%. This algorithm can be used to detect obstacles in the field environment and can provide a basis for the research of obstacle avoidance strategies in agricultural machinery autonomous driving.