Research On The Key Technologies Of Driver’s Gaze Estimation Based On Binocular Vision

Driver’s fatigue and distraction is the main causes for traffic accidents. Therefore, we have been focused on judging driver’s attentional state through gaze estimation to develop the relative safety assistant driving devices. We mainly study the following key technologies of driver’s gaze estimation based on binocular vision which include camera calibration, eye feature extraction, eye feature stereo matching and three-dimensional reconstruction, etc.(1) We constructed the binocular vision system hardware platform and then calibrated the system, which laid the foundation of driver’s gaze estimation. We mainly studied two common traditional calibration methods:three-dimensional method and two-dimensional method. By the comparison of both accuracy, speed and complexity, we get the conclusion that it can improve the reliability and accuracy of the calibration process by using the two-dimensional method with high precision and speed.(2) Only by accurate eye location can we extract effective eye features and then get precise gaze estimation results. At first, face location was conducted to make the eye location more fast and accurate. It can make make the feature regional orientation more simple, fast and efficient by introducing the image preprocessing, by the comparison of various smoothing methods we chose the Gaussian Blur method with clear and better image smoothing effect. Afterwards, we introduced the Haar feature detection in combination with the Adaboost algorithm to conduct driver’s face and eye original location, which not only meet the characteristics of driver face and eyes but improve the speed and accuracy of the feature detection.(3) Effective eye feature extraction is the foundation of driver’s gaze estimation. We utilized the single threshold method to extract the pupil area with obviously different pixels, and then used the morphological method to smoothly preprocessed the pupil contour edges. By the comparison of various ellipse fitting methods, we fitted the pupil contour curve based on the least square method with high accuracy, speed and stability. We selected the corner detection methods to get the Purkinje spot position because it had obvious grey-scale feature. The Harris Corner Detection algorithm is the optimum choice to get accurate Purkinje spot position because it has high speed, stability and precision, which lays the foundation of driver’s gaze estimation.(4) We conducted driver’s gaze estimation based on the pupil contour and the Purkinje spot information. Firstly, we improved the speed and accuracy of eye features stereo matching through the establishment of epipolar constraint between two images. By the comparison of two common methods, we selected the Bouguet algorithm with less distortion and simple calculation to simplify the stereo matching process. Among three common stereo matching methods, the BM algorithm has the best comprehensive effects so that it’s utilized to carry out the stereo matching of eye features. We used the three-dimensional gaze estimation method to get high accuracy and wide range, so the three-dimensional reconstruction of eye features is necessary to get the space coordinates of the pupil and the Purkinje spot. Above all, based on the three-dimensional space vectors and the relative location relationship of the pupil and the Purkinje spot, we accomplished the driver’s gaze estimation. Finally, the relative experiments were carried out to verify the reliability and accuracy of the methods used in this paper.The algorithm and software system which are utilized to analyze driver’s attention states based on the binocular vision have been developed by Visual Studio2008and OpenCV2.3.1. Experiments with the infrared face images of different drivers were conducted to analyze driver’s attention states, which is of positive significance for the further researches.