Orchard worker detection using spatial-context attention on 2D LiDAR data

Reliable worker detection in orchards is critical for ensuring safe collision avoidance as autonomous robots navigate narrow agricultural pathways. Although 2D light detection and ranging (LiDAR) is attractive for agricultural robotics, owing to its low cost, energy efficiency, and robustness to environmental lighting, existing methods frequently misclassify worker and tree trunks due to their geometric similarity, particularly when partially occluded. To overcome this challenge, an orchard worker detection (OWD) model is proposed that integrates a spatial-context attention mechanism into a 1D convolutional neural network (CNN), explicitly encoding global positional context to effectively differentiate workers from trees. The proposed OWD model uses a 2D LiDAR scan as input and first employs a streamlined 1D-CNN backbone to extract local geometric patterns (e.g., point density and curvature). Subsequently, a spatial-context attention module jointly encodes polar positional information and global contextual relationships among scan points, allowing the model to reliably distinguish a worker from structurally similar tree trunks, even in repetitive orchard-row environments. The proposed OWD achieved a mean average precision (AP) of 0.88, an equal error rate (EER) of 0.30, and an inference speed of 152.9 frames per second. These results demonstrate comparable detection accuracy and improved computational efficiency relative to the deep-learning baselines under the evaluated offline and embedded-inference conditions. The findings indicate the potential of the proposed model for integration into autonomous agricultural robots, subject to further field validation.