MDS-PointPillars: A lightweight obstacle identification method in farmland based on three-dimensional LiDAR for autonomous navigation

Abstract

To improve the autonomous navigation and operation of smart agricultural machinery in complex farmland, this study proposes a lightweight obstacle identification method, MDS-PointPillars, based on three-dimensional LiDAR to enhance perception capabilities. The MDS-PointPillars model primarily consisted of two components: the pillar feature net (PFN) and the backbone. In the PFN component, a multi-pooling encoding module (MPE) was designed, which integrated max-pooling, average-pooling, and attention mechanisms to improve the extraction of multi-scale point cloud features. In the backbone component, a depthwise separable convolution block (DSB) was designed to reduce computational complexity while enhancing the perception of both local and global features. Additionally, the model incorporated a parameters-free simple attention module (SimAM), which adaptively strengthened the focus on key point cloud features, further improving the identification accuracy of rare and hard-to-classify obstacles. Experimental results showed that MDS-PointPillars achieved a mean average precision (mAP) of 90.8 % on the test set of person, agricultural machinery, and utility poles in farmland, with an inference speed of 20.1 FPS and a model size of only 13.1 MB, significantly reducing computational burden. Robustness test revealed that the MDS-PointPillars maintained precision (P), recall (R), and mAP above 88.4 %, 87.2 %, and 88.6 %, respectively, across different scenarios, demonstrating its excellent adaptability and robustness in complex agricultural environments. Compared with mainstream models, MDS-PointPillars reduces parameters by 91.1 %, 76.1 %, and 73.2 %, and improves speed by 645.5 %, 70.8 %, and 368.6 % compared to Pv-RCNN, SECOND, and PointRCNN, respectively. This highlights its greater application potential in resource-limited farmland.