Tiny machine learning techniques for driving behavior scoring in a connected car environment

Connected cars and automotive car driving are emerging topics and many researchers are looking to provide new solutions and technologies that could revolutionize the driving experience. However, human car-driving is an activity that will certainly interact, in the near future, with autonomous driven vehicles. Hence, the need to provide an information layer about driving behavior and style that could benefit a traffic management system. Both insurance and car rental providers could also benefit of the knowledge of the driving behavior to custom fit insurance and rental rate. In this work, a machine learning model was deployed on a tiny hardware for performing the classification of car driving style based on inertial measurement data, augmented with in-vehicle diagnostic. The results show that an accurate (98%) model could be successfully deployed on an ultra-low-power STM32L4 Series MCU (MicroContoller Units) based on Arm® Cortex®-M4 core micro controller, capable to infer the proper classification among 4 classes in less than 5 ms after acquiring data from sensors. The system is also able to detect start and stop occurring during normal traffic driving.