Acoustic winter terrain classification for offroad autonomous vehicles

Autonomous vehicles can experience extreme changes in performance when operating over winter surfaces, and require accurate classification to transit them safely. In this work we consider acoustic classification of winter terrain, and demonstrate that a simple and efficient frequency-space analysis exposed to a small convolutional neural network, rather than recurrent architectures or temporally-varying spectrogram inputs, is sufficient to provide near-perfect classification of deep snow, hardpacked surfaces and ice. Using a dual-microphone configuration, we also show that acoustic classification performance is due to a combination of vehicle noises and vehicle-terrain interaction noises, and that engine sounds can serve as a particularly powerful classification cue for offroad environments.