Design and implementation of real-time eye detection and tracking system based on gpu

Non-Contact and remote camera-vision based eye detection and tracking systems has been widely used in many applications such as fatigue detection for the drivers, gaze detection for human-machine interfaces, iris recognition, and face normalization for automatic face detection systems. For these applications have the limitation of the real-time, how to accelerate the speed of the eye detection has always been an important topic. Until now, there has been several algorithm developed to decrease the computing time. Since the computational throughput still limits the types of algorithms that can be implemented in real time. However, the developing current Graphical Processing Units (GPU) offers algorithms in real time due to allow parallel programming and decreases the computing time. In this thesis, a real time algorithm is designed and implemented for eye detection and tracking system. An hybrid eye detection algorithm consist of feature based passive eye detection method and IR based active eye detection method is proposed. A new fast eye detection approach having 10 times fast performance with evaluated SCFace database on near infrared based eye detection has been proposed as a contribution. Also, the tracking methods such as Cam-Shift and Optical Flow methods are investigated with a created database including 30 people for the algorithm. The designed algorithm is implemented both CPU based embedded systems as Raspberry Pi, Banana Pi Pro, Beagle Bone Black, LeMaker Guitar and GPU based system Jetson TK1 for comparison. . The parallel computing advantages are presented for GPUs having higher almost ten times fast computation than CPU based systems.