Amid the rapid development of agricultural mechanization, with the widespread adoption of autonomous tractors, drones, and other equipment, real-time, precise pedestrian detection in farmland to prevent collisions between machinery and humans has become a critical demand for ensuring productivity and safety. However, existing algorithms often struggle in complex agricultural environments due to challenges such as variable lighting, dense targets, and frequent occlusions, leading to insufficient detection accuracy. How can smart devices efficiently identify pedestrians in scenarios with low-resolution images and densely packed small targets?
A study published by Associate Professor Yanfei Li’s team from Hunan Agricultural University in Frontiers of Agricultural Science and Engineering provides an innovative solution to this challenge (DOI: 10.15302/J-FASE-2025613).
Building on the widely used YOLOv8n model, this research proposes an improved algorithm named YOLOv8n-SS, specifically addressing the shortcomings of traditional models in complex farmland environments. The study enhances model performance through two key technological innovations: the introduction of a Spatial Pyramid Dilated Convolution (SPD-Conv) module and the integration of a Selective Kernel Attention Mechanism (SK). Traditional convolutional neural networks (CNNs) often lose fine-grained details when processing low-resolution images due to strided convolutions and pooling operations. SPD-Conv mitigates this by reorganizing spatial information into depth-wise features, significantly reducing information loss and improving detection accuracy for small targets (e.g., distant or partially occluded pedestrians). Meanwhile, the SK attention mechanism dynamically selects convolutional kernels of varying scales and adaptively enhances features in critical regions, enabling precise target localization in crowded or occluded scenarios.
Experimental results show that the improved algorithm achieves a 7.2% increase in mean Average Precision (mAP) on the public CrowdHuman dataset and a 7.6% improvement in real-world farmland scenarios compared to the original model. For instance, in low-light farmland conditions where the baseline model might miss pedestrians, YOLOv8n-SS accurately identifies all targets. In dense crowds, the new algorithm significantly reduces false detections, effectively distinguishing pedestrians from interference such as luggage or machinery. The researchers also highlights that the model maintains high real-time performance while requiring only moderate computational resources, making it suitable for deployment on agricultural devices with limited processing power.
This technological breakthrough not only provides a reliable tool for the intelligent upgrade of agricultural machinery but also opens new avenues for safety management in smart farms. By precisely monitoring personnel dynamics, the system can issue timely collision warnings or even halt machinery automatically, reducing accident risks. Future research will further optimize the algorithm’s lightweight design and explore target-tracking technologies to enable pedestrian behavior analysis and long-term monitoring, driving the transition of agriculture from “mechanization” to “human-machine collaborative intelligence”.
DOI: 10.15302/J-FASE-2025613