Rules-Based Distracted Driving Detection System Using Facial Keypoints

Abstract

Feasible distracted driving detection systems must be intuitive and non-invasive. Computer vision, a subset of deep learning, provides methods for computer systems to mimic humans in perceiving data from digital imaging. Previous work in distracted driving detection with computer vision has primarily focused on the classification of the entire image, which allows for detection based on body positions and objects in the frame. However, this does not fully isolate the human subject from the background and has possibilities for false positives in certain situations. Keypoint detection is a type of computer vision model capable of plotting points on prominent features of the human body using only a digital camera image. In this work, a rules-based algorithm with Euclidean distance normalization between facial keypoints was developed to determine if driver focus deviates from looking forward while driving. This algorithm also incorporates the steering angle to eliminate false positive detections when looking left and right in acceptable turning situations. This algorithm resulted in 100% accuracy in detecting distracted driving within the testing parameters used. However, future work will incorporate additional vehicle data, different camera types, new visual perception forms, and more practical testing scenarios for increased robustness.