Face pose estimation with automatic 3D model creation for a driver inattention monitoring application
AutoresJiménez Molina, Pedro
IdentificadoresEnlace permanente (URI): http://hdl.handle.net/10017/9743
DirectorBergasa Pascual, Luis M.
Fecha de publicación2011
Sistemas de asistencia a la conducción-Simulación por ordenador
Visión por ordenador
Electrónica-Aparatos e instrumentos
Texto en inglés y resumen en inglés y español
Tipo de documento
Derechos de acceso
Recent studies have identified inattention (including distraction and drowsiness) as the main cause of accidents, being responsible of at least 25% of them. Driving distraction has been less studied, since it is more diverse and exhibits a higher risk factor than fatigue. In addition, it is present over half of the inattention involved crashes. The increased presence of In Vehicle Information Systems (IVIS) adds to the potential distraction risk and modifies driving behaviour, and thus research on this issue is of vital importance. Many researchers have been working on different approaches to deal with distraction during driving. Among them, Computer Vision is one of the most common, because it allows for a cost effective and non-invasive driver monitoring and sensing. Using Computer Vision techniques it is possible to evaluate some facial movements that characterise the state of attention of a driver. This thesis presents methods to estimate the face pose and gaze direction of a person in real-time, using a stereo camera as a basic for assessing driver distractions. The methods are completely automatic and user-independent. A set of features in the face are identified at initialisation, and used to create a sparse 3D model of the face. These features are tracked from frame to frame, and the model is augmented to cover parts of the face that may have been occluded before. The algorithm is designed to work in a naturalistic driving simulator, which presents challenging low light conditions. We evaluate several techniques to detect features on the face that can be matched between cameras and tracked with success. Well-known methods such as SURF do not return good results, due to the lack of salient points in the face, as well as the low illumination of the images. We introduce a novel multisize technique, based on Harris corner detector and patch correlation. This technique benefits from the better performance of small patches under rotations and illumination changes, and the more robust correlation of the bigger patches under motion blur. The head rotates in a range of ±90º in the yaw angle, and the appearance of the features change noticeably. To deal with these changes, we implement a new re-registering technique that captures new textures of the features as the face rotates. These new textures are incorporated to the model, which mixes the views of both cameras. The captures are taken at regular angle intervals for rotations in yaw, so that each texture is only used in a range of ±7.5º around the capture angle. Rotations in pitch and roll are handled using affine patch warping. The 3D model created at initialisation can only take features in the frontal part of the face, and some of these may occlude during rotations. The accuracy and robustness of the face tracking depends on the number of visible points, so new points are added to the 3D model when new parts of the face are visible from both cameras. Bundle adjustment is used to reduce the accumulated drift of the 3D reconstruction. We estimate the pose from the position of the features in the images and the 3D model using POSIT or Levenberg-Marquardt. A RANSAC process detects incorrectly tracked points, which are not considered for pose estimation. POSIT is faster, while LM obtains more accurate results. Using the model extension and the re-registering technique, we can accurately estimate the pose in the full head rotation range, with error levels that improve the state of the art. A coarse eye direction is composed with the face pose estimation to obtain the gaze and driver's fixation area, parameter which gives much information about the distraction pattern of the driver. The resulting gaze estimation algorithm proposed in this thesis has been tested on a set of driving experiments directed by a team of psychologists in a naturalistic driving simulator. This simulator mimics conditions present in real driving, including weather changes, manoeuvring and distractions due to IVIS. Professional drivers participated in the tests. The driver?s fixation statistics obtained with the proposed system show how the utilisation of IVIS influences the distraction pattern of the drivers, increasing reaction times and affecting the fixation of attention on the road and the surroundings.