Low-power real-time TinyML approach for pothole and speed bump classification using IMU data on FPGA

Abstract

Monitoring the condition of road surfaces can bring benefits to drivers, ranging from increased safety and comfort, to cost reduction. In this context, accelerometer-based approaches offer a cheap and robust solution to the mitigation of common issues that arise from other types of sensors, like cameras, namely computational cost concerns and inability in low visibility scenarios. Nonetheless, as more applications compete for resources in modern vehicles, it becomes imperative to design such applications considering power efficiency constraints (TinyML): to this end, we propose a lightweight machine learning model to detect road irregularities (speed bumps, potholes, and manholes, including the side of impact) using 4 Inertial Measurement Unit (IMU) mounted near each vehicle wheel. Furthermore, to have a real-time monitoring system, we deploy an XGBoost model on a Field-Programmable Gate Array (FPGA) through the use of the Conifer framework, to translate the model using High Level Synthesis (HLS) language. Our solution achieves an accuracy of 88.1%, with 750 microseconds of inference time, showcasing a robust performance in varied scenarios.