International Journal of Engineering Trends and Technology (IJETT) – Volume 29 Number 3 - November 2015 Driver Drowsiness Detection System Using Automatic Facial Gestures Regina Antony#1, G. S. Ajith*2 #P. G Scholar & Computer Science & Engineering & Amal Jyothi College of Engineering Kanjirapally, Erumely Rd, Koovapally, Kerala, 686518,India *Assistant Professor, Amal Jyothi College of Engineering, Kanjirappally, Erumely Rd, Koovapally, Kerala, 686518 India Abstract-Drowsy driverdetection and recognize drivers state with high performance is the objectiveof this work. Drowsy driving is one of the main reasons of traffic accidentsin which many people die or get injured. Drowsy driver detection methodsare divided into two main groups: methods focusing on drivers performanceand methods focusing on drivers state.Furthermore, methods focusing on drivers state are divided into two groups: first identify whether the driver is drowsy or not by processing the facialexpression of the driver and comparing the best classification method. Here, driver data are video segments captured by a camera. There are twomain states of a driver, those are alert and drowsy states. Video segmentscaptured are analysed by making use of image processing techniques.Thisuses various images of driver to detect drowsiness states using his/her eyesstates, mouth state and head poses. Keywords—Driver fatigue, Drowsiness detection, Invehicle monitoring, Driver warning system. I. INTRODUCTION The increasing growth of population leads to the increasing number of vehicles in the road. As a result the number of vehicle accidents also increases. Detailed studies shows that around half million accidents occur in a year, and thousands of people die in this accidents in India alone. Driver drowsiness is one of the main reasons for these traffic accidents. The reports of national sleep foundation shows that 60% of adult drivers drive while feeling drowsy and 37% have even actually fallen asleep during driving [1]. The following are the signs shown by a drowsy driver: • Driver cannot keep eyes open • Frequent yawning • Unable to keeping head up • Driver is hurry and impatient Different statistics in different countries were reported that accidents happened mainly due to driver fatigue and distraction. Driver drowsiness and lack of attention are the main reason of about 30% of crashes and 20% of deaths [1]. The only method for controlling these accidents is development of technologies for drowsiness detection and prevention. There are two main approaches for detecting driver drowsiness. First and the commonly used method are video recognition techniques using camera images. By using this approach the images captured by camera is ISSN: 2231-5381 analysed and detect the physical changes in drivers, such as eyelid Movement, eye gaze, yawning and head nodding. The second one is the measurement of biomedical signals. This can give significant information such as fatigue in addition to drowsiness of the driver. Monitoring the brain activity is the best method for measuring the drowsiness. But in this method, the brain Activity is measured using an electrode that is placed in the Head of the driver that makes this an intrusive approach. Another significant feature where symptoms of drowsiness appeared is eye. There is a close relation between drowsiness and the percentage of eyelid closure. This paper mainly focuses on the driver face monitoring system that investigates the driver physical condition based on the processing of drivers eye, head and mouth images. II. PREVIOUS WORKS Processing the face region is the best method for detecting whether a person is drowsy or not. There are many researches based on this particular method. The reason for this is that the first symptoms of drowsiness and distraction are appearing in the face of the driver. Face detection is the foremost part in the drowsiness detection system [2]. There are mainly two categories for face detection: (1) feature-based and (2) learning-based methods [2].By applying certain heuristic rules on features, the face in the image can be detected. This is the basic principle behind featurebased methods. The rules are based on colour of the face, shape of the face etc. In the case of learningbased face detection method, to learn the discriminative features it uses statistical learning methods and the training samples. This method has less error rates but it leads to more computational complexity. Viola-Jones [3] algorithm is the commonly used for object detection. Almost in all drowsiness detection system, because of the symptoms related to face, the face regions are always processed for detecting the state of the driver. Mainly, there are three face detection methods are available: (1) based on the imaging in the infrared spectrum, (2) featurebased methods, and (3) other methods. Imaging in the infrared spectrum is one of the fast and relatively accurate methods for face detection [4]. http://www.ijettjournal.org Page 123 International Journal of Engineering Trends and Technology (IJETT) – Volume 29 Number 3 - November 2015 The part selected for imaging includes lighting and camera. Lighting and camera selection is one of the important stages when designing the system because the system should work in all light conditions. The use of lighting devices not only provides enough light in environment, but also should not disturb driver’s vision. Thus, near infrared (IR) spectrum is usually used in lighting. Module Description Second method for face detection is feature-based approach. It includes methods such as: Image binarization and projection. Smith et al. [5] based on the skin colour, binarization of face region was performed which causes the needed parts appear black, while other parts of the face appear white. Then, the use of connected component analysis increases the accuracy of face detection. In projection also, assume that the region we need is darker than other skin region.. As a result, most of the driver drowsiness detection systems detect driver drowsiness and distraction based on the symptoms extracted from the face. This is the main focus of this paper. There are number of techniques available for detecting the drowsiness of aperson. They are i) Image processing based ii) EEG based iii) ECG basediv) ANN based techniques. In this the most widely used technique to detectthe drowsiness of a person is the image processing based technique. Thisuses various images of a drowsy person to identify changes in his/her facialexpression while driving. In this method, first the images of the driver while driving is extracted fromthe video using the webcam sensor placed in the dashboard of the vehicle.Face detection and tracking are important in the image processing basedtechnique. There are 2 modules in here which are detailed below: 1. Facial feature extraction 2. Drowsiness detection according to extracted features A. Facial feature extraction Face and feature detection III. PROPOSED SYSTEM In the proposed system, drivers face images are used for processing so thatone can find its states. From the face image one can see that driver is awakeor sleeping. Using same images, they can define drowsiness of driver becausein face image if driver is sleeping or dozing then his/her eyes are closed inimage.And also we can predict their drowsiness with the help of the symptomslike yawning, blinking and head position like nodding, tilting and shaking.And other symptoms of drowsiness can also detected from the face image. The first step in this process is to detect the face. The face is detectedusing Viola-Jones algorithm of Haar features with Adaboost learning. Thisis commonly used for face detection. The method proposed by Viola andJones is composed of weak classifiers cascaded and their output is a strongclassifier which detects the target object. For each stage in the cascade, aweak classifier is trained to reject a certain fraction of the non-target objectpatterns and not rejecting any part of the target object. The classifiers useHaar features in order to encode facial features. For each feature, the value is the difference between the sum of the pixels in black regions and the sumof the pixels in white regions. Fig. 2 Haar features Fig. 1 Overview of proposed system ISSN: 2231-5381 Rectangle features are computed very rapidly using the integral image whichis an intermediate representation. The integral image at location pixels (x, y)contains the sum of the upper left pixels of the original image, inclusively. Thevalue of the integral image at location 1 is the sum of the pixels in rectangleA, the value at location 2 is A+B, location 3 is A+C and at location 4is A+B+C+D. Then, the sum within D is computed as 4 + 1 2 3 thatmeans the sum http://www.ijettjournal.org Page 124 International Journal of Engineering Trends and Technology (IJETT) – Volume 29 Number 3 - November 2015 of the pixels within rectangle D is computed with four arrayreferences. Integral image provides the advantage of fast feature evaluation. Fig. 3 Integral image Adaboost is used to select a small set of features and train the classifiers. The learning algorithm for weak classifiers is designed to select the single rectangle feature best separating the positive and negative examples. For each feature, the weak classifier determines the optimal threshold. Track Points Using the Kanade Lucas Tomasi algorithm the feature points in the detected face are identified. The identified points in the detected face are tracked using vision point tracker. Point tracker tracks each point in the previous frame with the corresponding points in the current frame. Eye Detection Detecting the eye region from the detected face is the next step in this technique. Eye detection phase is a module in which detected face is the input and the eye regions are the outputs. The eye regions include both the images of left and right eye region. Image of the right eye region is the eye region cropped from the original image and the image of the left eye region is the eye region cropped from the original image. Before given the input to the artificial neural network for further training and classification, the eye regions cropped from the detected face is need to be modified. Yawning Detection Yawn is one of the symptoms of drowsiness. The yawn is assumed to be modelled with a large vertical mouth opening. Using face tracking and then mouth trackingone can detect yawn. The mouth is searched in the lower most part of the detectedface region, thus further decreasing the computation cost of the system.If themouth opened beyond a particular threshold for a long time, then we cansay that yawning and driver is about to sleep. Head pose is estimated by calculating optic flow of the facial features. Headmovement is one of the symptoms that shows the drowsiness. If the headis moving either towards left or right or his head is moving down, then wecan say that he is drowsy. When the head moves down or lean towards leftor right, the average distance between both the eyes going to decrease andif the distance reaches a particular threshold, then we can conclude that thedriver is going to sleep. B. Drowsiness Detection According To TheExtracted Features After the detection of all the features are completed and the selected featureimages are cropped from frames, we need to modify those images. The framesextracted from video segments are in RGB format. When image in RGB format and a gray-level are compared, RGB format image has negligible advantage in displaying the image. This is verified bytests on several networks. In addition, the size of RGB images are threetimes the size of gray-level images, that’s why we work with gray-level imagesinstead of images in RGB format. Training Neural Network The time and memory consumption for training neural network increases with the increasing number of neurons. Therefore, we need to minimize thenumber of neurons. Theneural networks we use are feed forward networks which suit the objective, deciding the state of an eye from eye region image. The number of theeye regions obtained from the feature extraction stage is given as input fortraining the neural network. The output obtained from the training regionis the state of the driver. Driver State Estimation Two cases are considered for detecting the state of the driver. First one is drowsiness and the second one is distraction. For drowsiness detection,closing and opening of the eyes, continuous yawning and head pose are considered.If the driver keeps the eye closed for a certain amount of time (2seconds in our system), the eye state will be considered as closed. If hekeeps opening his mouth continuously and leaning head forward for a certainamount of time, then also he is going to sleep. Then considering the distraction, if the eye of the driver keeps focusing either to left or to right again for a particular time (2 seconds in our system) thenwe can conclude that the driver is distracted. After considering these twoconditions we can conclude that the driver is drowsy and distracted. Head Movement detection ISSN: 2231-5381 http://www.ijettjournal.org Page 125 International Journal of Engineering Trends and Technology (IJETT) – Volume 29 Number 3 - November 2015 IV. EXPERIMENTAL RESULTS The proposed method was tested on different group of varying ages. The main purpose of this experiment is to acquire the date for analysis and to compare the validity and accuracy of the system. The web cam placed in the dashboard of the car captured the video of the driver while driving. From the video, frames are acquired and using that frames, the face of the driver is detected using Viola Jones algorithm. Then the eye and mouth regions are cropped and head node is calculated. Then it is given to the ANN for training. number of neurons. That is why; only 81 neurons in the input layerof neural networks are used. The neural network here used is feed forwardnetwork and which is reliably good for estimating the state of driver.Theneural networks here use are feed forward networks which suit the objective,deciding the state of the driver from various states. Tried to train neural networks by the back propagation methods. But the disadvantage is that, itsmemory requirement is more than the other methods. Since my problem canbe categorized as a nonlinear problem, here need to use nonlinear activationfunctions. Here, first the eye states are considered as open, half closed and closed andtrain each frames based on artificial neural network. Then head pose andmouth movements are considered and they are also trained. The graph showsthe prediction accuracy of actual and the predicted values of the featuresshowing the drowsiness. For that, eye movements, yawning and head movements are taken and given them to an ANN classifier for classification. Andit produces an ROC curve showing the accuracy of the drowsiness detectionsystem proposed here. Fig. 4 Detected face First, the video is input to the Frame Extractor module followed by the analysis of the output 150 frames one by one and then each frame is labelled according to eye states, head pose and mouth states. These valuesform the ground truth for all the features for each frame. After that theseare converted to gray level images. The intensity values of each pixel in theobtained images with size [10 16] are transferred into a matrix with size [814]. Reshaped feature image matrices with size [81 4] are concatenated and amatrix with size [81 150] is generated for a 30 second video segment. Fig. 6 Driver safety analysis graph A confusion matrix is also obtained from ANN after training. Confusion matrices for training, testing, and validation, and the three kinds of data combined and showed in the figure. The network outputs are very accurate, asyou can see by the high numbers of correct responses in the green squaresand the low numbers of incorrect responses in the red squares. The lowerright blue squares illustrate the overall accuracies. Fig. 5 Neural Network The time and memory consumption for training neural network increaseswith the increasing number of neurons. Therefore, there is a need to minimize the ISSN: 2231-5381 http://www.ijettjournal.org Page 126 International Journal of Engineering Trends and Technology (IJETT) – Volume 29 Number 3 - November 2015 he closes his eyes for long time which is giving information that thedriver might have slept. In this work, a method to locate and track a drivers mouth and eye usingcascade of classifiers training and detection of eyes and mouth. Training of eye movement and mouth images are done by using ANN. Finally, ANN isused to classify the mouth regions to detect yawning,eye region to detect eyemovement and head node then alert drowsiness. The experimental resultsshow that proposed method gives better results than methods using geometric features. The proposed method detects yawning alert fatigue earlier, andwill facilitate to make drive safer. ACKNOWLEDGMENT We thank computer science department of Amal Jyothi College of Engineering for providing us with relevant data. This work was supported as part of thesis project. Fig. 7 Confusion matrix REFERENCES [1] V. CONCLUSION Various Techniques for drowsiness detection has been reviewed with variousconditions. In this system, the eye movement of the driver, head movementand the yawning of the driver are used as various features used to review thedrowsiness. A non-intrusive visual based system is developed to locate eyesand mouth and determines the driver’s drowsiness level through horizontalaverage intensities of the eyes and mouth region at face. During monitoringthe system is able to detect when the eyes are closed, head movement andmouth open and simultaneously for too long and again and again in less period of time thus giving an alert to the driver. Also the system alerts thedriver if ISSN: 2231-5381 [2] [3] [4] [5] N. L. Haworth, T. J. Triggs and E. M. Grey, Driver Fatigue: Concepts, Measurement and Crash Countermeasures, Human Factors Group, Department of Psychology, Monash University, 1988 June. M. H. Yang, D. J. Kriegman and N. Ahuja, Detecting Faces in Images: A Survey, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, no. 1, 2002. P. a. M. J. Viola, Rapid Object Detection Using a Boosted Cascade of Simple Features, Proc. IEEE Conf. Computer Vision and Pattern Recognition, Kauai, HI, USA ,2001. R. Grace, V. E. Byme, D. M. Bierman, J. M. Legrand, D. Gricourt, R. K. Davis, J. J. Staszewski and B. Carnahan, A Drowsy Driver Detection System for Heavy Vehicles, Proceedings of 17th AIAA/IEEE/SAE Digital Avionics Systems Conference (DASC), Washington, USA , November, 1998. P. Smith, M. Shah and N. V. Lobo, Monitoring Head/Eye Motion for Driver Alertness with One Camera, Proceeding of 15th IEEE International Conference on Pattern Recognition, Barcelona, Spain, 2000. http://www.ijettjournal.org Page 127