Lightweight Terrain Classifier Hits 97.37% Accuracy Using Inertial Sensors
Researchers have developed a lightweight terrain-classification method that reached 97.37% accuracy on six surface types.
The approach uses time-series data from multiple inertial sensors mounted on a robot. It was tested on concrete, grass, pebbles, sand, paving stone and synthetic running track surfaces.
The method starts with sensor data recorded in Cartesian coordinates. It then converts the same data into cylindrical and spherical coordinates and fuses the resulting features before classification.
To reduce dimensionality, the system calculates the average value of each feature over the measurement period. Data augmentation is also applied to an open-source dataset before classification to increase the number of samples available for training and testing.
The model was evaluated with a Weighted Ensemble classifier and achieved 97.37% accuracy. The findings indicate that combining multiple coordinate systems in a reduced feature set can improve terrain recognition while keeping computational demand relatively low, a key requirement for mobile robots operating across changing ground conditions.