Neural networks terrain classification using Inertial Measurement Unit for an autonomous vehicle

This research is focusing on the terrain classification using data from an inertial measurement unit acquired during vehicle motion. The proposed classifier is different from the vibration-based classifier in the fact that it uses the relationship between different axis of input as well as the spectral information to classify the difference between terrains. The data from the inertial measurement unit (IMU) are three axes acceleration and three axes angular velocity. The acquired data are preprocessed and filtered by fuzzy rules, then classified by a neural network into 5 categories: flat plane, rugged terrain, grassy terrain, incline plane and unclassified. The trained networks were experimentally validated with 100 samples in each category. The result shows that the proposed classification method can classify a flat plane, rugged terrain, and incline plane 100% correctly. For grassy terrain, it can be classified correctly about 80%.