Road Surface Classification Using an Accelerometer and MiniROCKET on Ultra-Low-Power STM32

Embedding Machine Learning (ML) algorithms into Advanced Driver-Assistance Systems (ADAS) has been growing in popularity, since they enhance the perception of sensors such as cameras, radars, and LiDAR. While Inertial Measurement Units (IMUs) have a high maturity in applications such as airbag deployment and stability control, the application of ML to IMUs allowing for more complex applications in the automotive industry has been comparatively underexplored. However, deploying ML models in automotive applications presents challenges, including strict power, memory, and latency constraints and to address this, we propose a Tiny Machine Learning (TinyML) based approach for road surface classification using an accelerometer. Our method employs a state-of-the-art time-series random convolutional kernel feature extraction method - MiniROCKET - which combined with a Logistic Regression model achieves a 19 ms inference time, an extremely low 30 mW power consumption - roughly equivalent to a blinking LED - and a classification accuracy of ${9 6. 9 6 \%}$ when distinguishing between Asphalt, Cobblestone and Dirt roads on an STM32L476RG microcontroller. This demonstrates the feasibility of efficient, low-power road surface classification for embedded ADAS applications.