Our algorithm is designed to minimize false-negative detections as much as possible.Īlso, we propose a new ROI selection technique that suits to mobile environments. Also, false-negative detections may let advisories bypass blinking detection based antispoofing mechanisms in face-recognition systems. For example, a CVS advisory system based on an eye-blinking detection cannot generate proper advices when the system over-estimates the blinking counts with false-negative detections. We see that, in eye-blinking detection algorithms, false-negative detections, that is, reporting of a detection of eye blinking when there is no actual eye blinking, are problematic. In other words, we use a SVM detector as a region proposal method in object detection. Thus, in our scheme, the two are conjoined to achieve efficiency and accuracy in the eye-blinking detection. SVM classifiers are fast but less accurate deep CNN models are accurate but slow. To address the problem, we take a hybrid approach conjoining two machine learning techniques, a SVM (support vector machine) classifier with the histogram of oriented gradients (HOG) features and a deep CNN model called LeNet-5 such that the eye-blinking detection can be performed efficiently and reliably on resource-limited smartphones. Especially when it comes to eye-blinking detection, the applicability of the CNN technique becomes more restricted since the eye-blinking detection problem imposes a stricter real-time requirement, that is, a processing rate over 10 frames per second (fps), than eye tracking and thus must be more computationally efficient. Although deep CNN models have been successfully applied to solve various computer vision problems in a past few years, running deep CNN models on smartphones are still considered to be challenging due to the computational complexity of deep CNN models. Recently, a few deep convolutional neural network- (CNN-) based eye tracking algorithms for smartphones are proposed, for example. Some scheme, for example,, utilizes built-in sensors such as accelerometers to effectively predict the eyes’ positions in the input video frames and thus achieving real-time eye tracking in dynamic mobile environments.
Eye tracking for android how to#
In a ROI-based solution, how to set ROI is the key to the system’s performance, and the problem gets much harder in mobile environments. Locating eyes in the entire region of every video frame is time and energy consuming and thus may not suitable for real-time processing on smartphones. Instead of searching for the eyes in the entire region of each video frame, a small region called ROI is examined. One way of tackling these problems is to utilize region of interest (ROI), for example. There exist a number of studies on eye-tracking and eye-blinking detection algorithms based on video input, for example, however, schemes developed for desktop environments may not work properly in mobile/smartphone environments due to the limitations of computational resources of smartphones and/or the frequent change of the positions of the eyes in input video frames caused by user movements. Also, computer vision syndrome that many smartphone users are experiencing can be mitigated when eye-blinking detection systems running in background give advices to smartphone users regarding their eye-blinking habits. For example, it can be used as a countermeasure against spoofing in face recognition systems. IntroductionĮye-tracking and/or -blinking detection algorithms have various applications in smartphone platforms. Experimental results on commodity smartphones show that our approach achieves a precision of 94.4% and a processing rate of 22 frames per second. To tackle the problem, we take a hybrid approach combining two machine learning techniques: SVM (support vector machine) and CNN (convolutional neural network) such that the eye-blinking detection can be performed efficiently and reliably on resource-limited smartphones. In resource limited smartphone environments, one of the key issues of the eye-blinking detection problem is its computational efficiency. Eye-blinking detection or eye-tracking algorithms have various applications in mobile environments, for example, a countermeasure against spoofing in face recognition systems. We propose an efficient method that can be used for eye-blinking detection or eye tracking on smartphone platforms in this paper.