Terrain Classification Algorithm for Off–Road Vehicle Using 1D CNN–LSTM Network

When driving in unstructured outdoor environments, vehicles may encounter various potential dangers such as wheels sinking, skidding, or even loss of maneuverability, significantly reducing their operational safety. Therefore, identifying terrain accurately is crucial for ensuring vehicle safety and achieving optimal vehicle control. To address this issue, this study designed three different neural network models for road recognition: the Long Short-Term Memory network (LSTM), the one-dimensional Convolutional Neural Network (1D CNN), and the combined 1D CNN-LSTM architecture. In these models, the classification of five typical off-road surfaces-namely asphalt, deep snow, mud, sand, and gravel-is performed by leveraging Controller Area Network (CAN)-bus signals and the computation of vehicle dynamic attributes derived from the vehicle. Algorithm validation experiments based on real vehicle data demonstrated that among the three classification models, The model employing a 1D Convolutional Neural Network (CNN) combined with a Long Short-Term Memory (LSTM) network attained an accuracy rate of 99.2%, which was the highest among the evaluated models.