| This paper describes the development of a vehicle driver drowsiness warning system using multi-layer perceptron neural network as a classifier for driver's different orientations of gazes. After long hours of driving or in a less than alert kind of mental state, drivers in fatigue exhibit certain visual behaviors easily observable from changes in their facial features like the eyes, head, and face. Typical visual characteristics observable from the image of a person with reduced alertness level include slow eyelid movement, smaller degree of eye openness, frequent nodding, yawning, gaze, etc, which create risks for accidents. The principle of the proposed system is based on a non-invasive approach for monitoring drive's gaze. Facial images of the driver are taken by a CCD camera, which is installed on the dashboard in front of the driver. Techniques of wavelet transform and entropy analysis are introduced to location face/eyes. Then, MLP is used to classify driver's gaze and assess driver's vigilance level. If it was necessary, the system should warn the driver accordingly.In this project entropy analysis integrated particle filters model is investigated and extended in the field of eyes tracking. This method could location eyes fast and exactly. Research focuses on the following topics:(1) Designing system: The system has two phases-ROI detection and drowsiness analysis. Face detection starts with a pre-processing to remove external illumination interference, followed by a global search of the whole image to locate the first frame of face in the video. If failure, algorithm will detection ROI again. Eyes tracking locally searches the eyes based on the eyes' positions in prior frames.(2) Face detection approach: We present an Adaboost classifier approach based on entropy analysis. After study from training set, we select the features that best represent the pattern class of faces, and setup decision tree classifier models. These models are applied by the input of likelihood of features. Experiments show that this approach could locate face in spite of any poses and time expenditure is small.(3) Eyes tracking algorithm: within the region of detected face, eyes are tracked basedon SPF (Superstate Particle Filter) in real time. Recently particle filters or the condensation algorithm has shown to be very suitable to perform real-time tracking in cluttered environments. Unlike Kalman filters, which are limited to Gaussian probability distributions, particle filters are able to represent multimodal distributions. A discrete set of particles represents the object-state and evolves over time driven by the means of "survival of the fittest". The states are estimated by post density. In this paper, Auto Regression model is used to predict eyes states. Entropy analysis algorithm is used to identify eyes so that subpopulations are able to evolve independently. Meanwhile, various techniques to deal with overlapping, eyes entering and eyes leaving scenes are discovered. To reduce the number of particle and the complexity of computing, we perform pyramid algorithm on the video sequences. The tracking algorithm is performed on the coarse image level.(4) After recognition of eyes the system use MLP as a classifier of eyeball gaze. At last, we sum some rules to detect driver's drowsiness by the image of recognized eyes and the output of MLP.
|
PDF Full Text Request