SM-YOLO: an improved off-road trafficability detection network in complex field environment for autonomous driving

Abstract Off-road trafficability determines the in-depth application of autonomous driving in complex field environment, which creates an urgent need for the real-time assessment of soil strength. However, traditional models are inadequate in meeting dynamic demands while existing deep learning methods continue to pose substantial challenges in complex environment. A novel assessment framework of vehicle trafficability is proposed by online monitoring of soil moisture. To achieve non-contact detection of soil moisture content, an improved detection model named Soil Moisture-YOLO (SM-YOLO) is developed. Firstly, a nonlinear feature fusion mechanism is introduced to mitigate the feature degradation in complex scenes. Then, this study designs a multi-scale squeezing excitation module, which enhances the focusing capability on the target soil region. The attention mechanism is optimized based on Mamba, which improves the inference efficiency. Finally, comparative experiments are conducted based on a specialized field soil image dataset. Simulation results show that SM-YOLO offers a good trade-off among accuracy, lightweighting and real-time inference.