Road Type Identification Method for Off-Road Vehicles Based on a CNN-GRU-Attention Deep Temporal Network

To address the challenges of complex road types in off-road environments and the susceptibility of visual perception to texture similarities (e.g., similar appearance between washboard roads and concrete roads), this paper proposes a road type identification method based on the time-series analysis of vehicle dynamics response signals. A composite deep network, CNN-GRU-Attention, is constructed: a 1D-CNN is utilized to extract local high-frequency impact features from velocity and acceleration sequences; a GRU is employed to model the long-term dependencies of road excitation evolution over time; and an Attention mechanism is applied to explicitly weight critical impact time steps, enhancing the representation of sparse critical events. Comprehensive ablation studies were conducted to quantify the contribution of each module, including comparisons with LSTM-based variants. Experiments were conducted based on real vehicle data (sampling frequency 100 Hz, segment length 1 s) across five typical road types: concrete, gravel, pebble, washboard, and fish-scale pits. The results demonstrate that compared to single CNN and single GRU models, the proposed method achieves more balanced and higher recognition performance across all road types. It significantly improves the identification stability of easily confused categories (such as pebble/washboard and washboard/concrete), providing reliable road priors for chassis control and suspension adaptation of unmanned vehicles.