Driver Monitoring Using Face Detection and Facial Landmarks

Abstract

The influence of fatigue on accidents has been demonstrated over time by conducting several research papers, whose conclusions came to confirm the actual situations faced by drivers every day. Globally, between 10% and 25% of road, accidents are caused by fatigue, and studies have shown that professional drivers are more prone to the risk of being involved in an accident than casual drivers. Nowadays a high percent of automobile accidents are caused by human error. Human errors reflected by drivers lacking necessary vigilance levels or reacting improperly causing inadequate control of their vehicle.The critical role of human factor in an accident motivated the development of a new series of countermeasures aimed to prevent or mitigate human error while driving [1]. Concept systems developed and implemented to monitor in real-time the performance of driver help to achieve the goal of reducing the overall rate of accidents. Current systems that use image processing and computer vision techniques are able to measure driver fatigue and perceive distraction or engagement levels [2], [3]. These driver assistance systems analyze driver head movements, face and eye movements to evaluate the driver engagement level while driving [4]. The work presented in this paper describes a system developed to monitor driver face and eye movements with the purpose of identifying incipient clues describing driver drowsiness and distraction from driving task. A video camera captures image frames with the driver face supplied as input to the system. The proposed image processing system analyzes the video, frame by frame, locates driver face, detects the eyes regions, measures the movement of eyes and eyelids and then evaluates drowsiness and distraction levels. Performance characteristics of the proposed driver assistance system have been tested using video sequences acquired during a test drive with a car in a real world environment.The first part of our work challenges the implementation of a face detection procedure. Although this problem is common and was tackled by various approaches [5], it remains of high complexity imposing interesting challenges at system level. One of the challenges is to make a robust implementation that is able to detect or recognize a human face in environments described by a high dynamic range of luminosity, occlusion of scale. The complex problem of such a driver assistance system partially tacked in our work by highlighting the main issues, their impact and proposing dedicated solutions that address these challenges. Our implementation proposes solutions tested in real world scenarios to design challenges that influence the outcome when deciding the concept of a face detection or recognition system for a real world application. The solutions proposed to address the design decisions process when conceptualizing such a system convey the usage of methods for feature extraction, approaches to holistic matching and other hybrid methods described with more details in the following sections.The second part of our work addresses challenges encountered when developing a concept for driver drowsiness or driver vigilance monitoring. In the more general use case of such a system it would be of high interest to correlate the environment with what driver is perceiving or what is aware of while driving. This high-level functionality would allow the system to warn the driver when is not aware or it does not perceive something in the surrounding environment. Our work evaluates the performance of a driver vigilance method that uses data obtained by measurement of face and eye region movements. The method employs a face-matching step and a histogram projection approach to extract driver vigilance data from face and eye regions. Head rotation used by our method to gather data about the visual focus point of driver correlated to the orientation of driver face. This approach uses the following advantages of the method: low complexity and efficient detection of head rotation using face template matching, robust features extraction from the eye region. The extracted features employed by our method evaluate the driver drowsiness and distraction levels.