Drowsiness has been identiﬁed as one of the most important causes of traﬃc accidents, as it is pressent in the 20 % of them. Therefore, there is a growing interest in looking for ADAS (Advanced Driver Assistance Systems ), capable of detecting driver’s fatigue, to prevent potential accidents. So that, research in this area is vital.
In order to tactle drowsiness analysis while a user is driving, several worldwide research groups have been working on diﬀerent techniques. Computer Vision techniques are prominent among them, since it allows, by means of relatively cheap technology, to monitor driver’s state in a non intrusive way.
In this thesis a technique based on monocular image processing is introduced. This consists of detection, tracking and characterization of eye closure, able to deal with diﬀerent users and real driving conditions. Using this information and others acquired from the car, the driver behaviour is infered.
Driver-related signals have been obtained from a wide set of sequences, in which there are different people's faces, either simulation or real driving conditions, awake or sleepy users.
For the purpose of driver drowsiness evaluation is necessary to generate a ground truth, which can provide the real state of the user at a particular time. This signal is obtained by 3 experts, as a result of studying several parameters as: KSS Karolinska Sleepines Scale, facial features obtained from the recorded sequences and registered signals from the driving process. This ground truth usually has 2 levels, awakeness and drowsiness, although in this Thesis it has been extended, including a middle level: fatigue, to preciselly improve the classiﬁcation of the user’s state. After the analysis of the results, it has been proved that is better using the binary classiﬁer. This methodology is new because takes information of the KSS scale and some experts.
Once all the user’s face sequences were obtained, the developed computer vision techniques to obtain eye closure have been tested. Face detection is based on Viola & Jones algorithm, which is appearance-based, and eye detection is improved using clustering techniques and Kalman ﬁltering, as predictor. Eye closure is obtained applying, adaptive ﬁlters, projective integration and Gaussian modelling. All these image treatment algorithms makes the system robust against illumination variation and diﬀerent users, archieving real time operation. Once the closure is known, the parameter PERCLOS (PERcentaje of eye CLOSure ) is computed. This parameter is one of the most relevant in drowsiness detection.
Talking about driving signals, obtained from the car, which depends on the user, some signs ﬁltered using temporal window have been obtained like: standard deviation (STD), root mean square error (MSE) of the vehicle possition on the lane, number of lane excedances (Lanex), time to lane crossing (TLC), its standard deviation, fast changes on the steering wheel angle, yaw angle standard deviation, and generic variability indicators (GVI). In order to remove sign-user dependence, parameter optimization using genetic algorithms has been carried out, taking into account the ground truth.
In order to stablish the driver’s state, some indicators have been merged using a multilayer perceptron neural network, in which the number of neurons of the hidden layer is set using the ROC curve.
Some results are shown in simulation and real driving conditions. The classiﬁcation is performed by individual signals as well as fusing them, presenting the results using diagram error for 2 and 3 variables, and tables where the recall rate, the speciﬁcity, the sensitivity and the objective function are shown.
The results related to drowsiness detection demonstrates that PERCLOS is a fundamental parameter for the estimation of driver’s state, and merging it with other driving signals improves the overall recall rate. Indicators related to driving yield worse results than using PERCLOS because those signals are not only caused by drowsiness but also by real driving conditions
that are diﬃcult to estimate. Heading error has been tested only in simulation because in real conditions it has not been provided by the person in charge of driving signals. During the simulation, the best possible combinations are the following: the fusion of PERCLOS and the standard deviation of the heading angle; and PERCLOS and the optimized MSE indicator.
During real conditions, the best possible combination is the last one. If heading angle had had taken into account during real operation, the recall rate would have been improved. Therefore, the obtained results guarantee the methodology used, and can be easily extrapolated from the realistic simulator to real driving operation. Conclusions obtained using this methodology are valid for real conditions even though the detection rate is lower since the input signals are noisier.
The results are in line with other important works about this sub ject [Sandberg, 2011] except in the consideration of the PERCLOS, the best signal for us maybe due to we use our own vision system and not a commercial one. On the other hand, results are better than other important works of the state of the art [Friedrichs & Yang, 2010a, Caterpillar, 2008].