Design and Implementation of On-Device AI for Dangerous Driving Behavior Recognition

Recognizing dangerous driving behaviors is crucial considering the significant number of casualties in road accidents. This paper proposes an on-device AI for recognizing drivers' dangerous driving behaviors during vehicle operation. The driving behaviors are recognized using deep learning and implemented as a compact device inside the vehicle. The deep learning model utilizes the vehicle internal sensors to classify dangerous driving behaviors in real-time, distinguishing a total of 17 driving behaviors, including 9 normal driving behaviors and 8 dangerous driving behaviors. Due to the use of low-spec processors in the on-device AI system, lightweight techniques are applied to reduce the size of the deep learning model by over 50%. Nonetheless, the system achieves an accuracy of 83.24% and an F1-Score of 0.81. The lightweight deep learning model implemented in the on-device AI demonstrates an inference time of 13.6ms, enabling real-time operation. This allows for various applications, such as recognizing dangerous driving behaviors in real-time and providing drivers with warnings and encouraging safe driving.