Gaze Tracking System With Free Head Motion

Eye tracking technology can be used in many aspects such as medical rehabilitation, military operations, vehicle driving, etc. Using the invariance property of cross-ratios for gaze estimation has been research hot spot for the eye tracking. The main problem of this method is the low accuracy which caused by adapting static calibration parameters, this paper presented a method to establish a dynamic virtual plane based on dynamic calibration parameters, the main contributions are as follows:An eye tracking system under free head motion was built. For different users, the calibration is needed before using the invariance of the cross-ratio. The system could not only satisfied the large range of head motion during gazing estimation but also had good portability. the IR LED placed on an adjustable shelf making the system applied for different size of screen.In view of the shortage of traditional gaze estimation method which used static parameters, An gaze estimation algorithm which based on gaze point was proposed. By introducing a dynamic virtual plane model, established the relationship between the gaze point and the parameters in the off-line parameter learning. The method was divided into two basic steps: estimation and optimization steps. Estimated the Rough location with the traditional method; Optimized the parameters for final gaze estimation. This method got higher precision compared with the traditional method.In addition,this paper proposed cross-ratio based dynamic parameters modeling and calibration method. We have improved the calibration method. An iterating method was used to optimized calibration parameters.Combined with the traditional parameters for gaze estimation, This method detailed the characterization of the calibration parameters, which can effectively improve the estimation accuracy.We have finished the experiments using two methods above descripted at the same time, we found that the estimation precision based on static parameters is 2.6°, while the estimation precision based on dynamic parameters is 1.1 °. The experimental results show that the proposed method based on dynamic parameters can signi?cantly improve the gaze estimation performance.