A novel lightweight YOLOv8-PSS model for obstacle detection on the path of unmanned agricultural vehicles.

Introduction: The rapid urbanization of rural regions, along with an aging population, has resulted in a substantial manpower scarcity for agricultural output, necessitating the urgent development of highly intelligent and accurate agricultural equipment technologies. Methods: This research introduces YOLOv8-PSS, an enhanced lightweight obstacle detection model, to increase the effectiveness and safety of unmanned agricultural robots in intricate field situations. This YOLOv8-based model incorporates a depth camera to precisely identify and locate impediments in the way of autonomous agricultural equipment. Firstly, this work integrates partial convolution (PConv) into the C2f module of the backbone network to improve inference performance and minimize computing load. PConv significantly reduces processing load during convolution operations, enhancing the model's real-time detection performance. Second, a Slim-neck lightweight neck network is introduced, replacing the original neck network's conventional convolution with GSConv, to further improve detection efficiency and accuracy. This adjustment preserves accuracy while reducing the complexity of the model. After optimization, the bounding box loss function is finally upgraded to Shape-IoU (Shape Intersection over Union), which improves both model accuracy and generalization. Results: The experimental results demonstrate that the improved YOLOv8_PSS model achieves a precision of 85.3%, a recall of 88.4%, and an average accuracy of 90.6%. Compared to the original base network, it reduces the number of parameters by 55.8%, decreases the model size by 59.5%, and lowers computational cost by 51.2%. When compared with other algorithms, such as Faster RCNN, SSD, YOLOv3-tiny, and YOLOv5, the improved model strikes an optimal balance between parameter count, computational efficiency, detection speed, and accuracy, yielding superior results. In positioning accuracy tests, the, average and maximum errors in the measured distances between the camera and typical obstacles (within a range of 2-15 meters) were 2.73% and 4.44%, respectively. Discussion: The model performed effectively under real-world conditions, providing robust technical support for future research on autonomous obstacle avoidance in unmanned agricultural machinery.