Abstract: In order to solve the problem of difficulty in balancing accuracy and real-time performance of obstacle recognition algorithms for grassland inspection robots due to complex external environments and insufficient computing power,a lightweight detection model for grassland obstacles based on YOLOv8 is proposed,which utilizes an efficient multi-scale attention module to enhance network feature extraction capabilities.At the same time,1X1 convolution is added to the neck structure of the network for dimensionality reduction mapping processing,reducing the number of parameters in the network.This paper also replaced the loss function of the original network with WIoU,reducing the impact of low-quality images on the model during training.Experiments conducted on self-built datasets have shown that the improved model has an F1 score of 93% and an average accuracy value(mAP) of 96.2%,which is 1 and 1.9 percentage points higher than the original model.The model parameter size is 1.96×10⁶,which is 34.7% lower than the original model.Finally,the model was ported to an embedded platform and FP16 quantization was performed,resulting in a 35% increase in running frame rate.The proposed method can balance accuracy and real-time performance,and is a lightweight detection method suitable for embedded platforms,providing technical support for obstacle detection of grassland inspection robots.

Key words: Grassland inspection robot, Obstacle recognition, Attention mechanism, Lightweight detection methods, Embedded platform