Personal Mobility Safe Driving System with Knowledge Distillation

The global personal mobility market has been rapidly expanding due to its convenience. However, the increasing number of accidents involving personal mobility devices has become a growing concern, including falls, collisions with objects, and riders being struck by moving vehicles or objects. In this paper, we propose a deep learning-based safe driving system that considers both user and road images to address this issue. Our system employs CNN-based models to detect whether the user is 1) wearing a helmet and 2) looking ahead in user-side images. At the same time, the roadside image recognizes whether the user is 3) driving on the sidewalk and 4) near the intersection. These tasks are simultaneously performed in parallel to identify the overall situation, which is done by determining the final speed as the minimum speed of speed values extracted from all tasks. Additionally, we employ knowledge distillation techniques to compress the models and enable real-time inference on edge devices, resulting in a fast and accurate system that is well-suited to the characteristics of personal mobility.