Real-time Eye-gaze Tracing System Based On CPU+GPU

Gaze tracing technology based on mechanical, electronic, optical method is moreand more widely used in the areas of human-computer interaction, helping the old andthe disabled, the vehicle assistance, virtual reality, military and so on. The gazetracing system created in this paper is based on digital video analysis as well as imageprocessing technology. In our system, images are read in from a single camera, anddisposed to locate the pupils and the inner corner of the eyes. Then the characteristicpoints are substituted into the sight estimation model to judge the point where thepeople watching at and display the gaze point on the screen.The system contains two main phases of the parameter initialization andreal-tracing process. In the phase of parameter initialization, nine images are acquiredwhen people watching different locations on screen and then the images are detectedto locate faces and eyes’ characteristics. In the phrase of real-time tracing, each framecontinuously acquired in are detected to get eyes’ characteristics which are used toestimate gaze point by applying gaze estimation model, then the gaze points aredisplayed on the screen.In order to achieve real-time tracing on the platform of desktop system, the entiresystem is built on the platform of CPU+GPU and a reasonable tast partitioning andscheduling strategy is designed based on the features of system task. CPU isresponsible for constructing system and task scheduling, and the large amount ofcalculation part which is to detect faces and eye’s features on images is allocated onGPU. In this paper, we make an intensive study of the parallelism of Adaboost facedetection algorithm and partial eye’s features detection algorithm and design effectiveparallel algorithm based on the parallel architecture of GPU. Good detection resultsare verified by experiments and real-time eye’s features detection is achieved. On thisbasis, the system effectively estimates the gaze points of eye on real-time, and realizesthe gaze tracing result.