Accelerometer-based data-driven hazard detection and classification for motorcycles

This article deals with collision and hazard detection for motorcycles via accelerometer measures. A machine learning approach is proposed. A two-phase method is developed that is capable of first detecting non critical anomalies (unusually high accelerations) and critical hazards for which an airbag deployment could be needed. The method is based on Self Organizing Maps and has two may advantages over the classical approach: 1) the machine learning approach easily scales with the number of sensors. 2) It is tuned using normal driving and does not require expensive crash-tests for tuning. In the paper the system is designed starting from data from an instrumented vehicle and validated in simulation.