Physics-aware learning of vehicle speed from accelerometer data

A key question that arises when dealing with sensors is how to process the available signals to obtain valuable information for monitoring and control, with limited amount of memory and computing power. In this paper, we build a vehicle speed black-box estimation method based on accelerometer data only. The key feature allowing this is the fact that the architecture of the overall algorithm is constructed by looking also at the physics of the problem. As a second contribution, we discuss how the obtained model can be further simplified via frequency-based considerations. Experimental tests carried out on a real vehicle show that the developed estimators are both effective and robust to varying environmental conditions. Specifically, the rms of the estimation error is never greater than 8 km/h.