Method for stabilizing line-of-sight falling point
Disclosed is a line-of-sight detection method, including a step SA1 and a step SA2. The step SA1 performs a dynamic-threshold-to-switch action to correct the error of the data generated in the previous facial feature detection step. Furthermore, the step SA2 performs an eye-ROI-filtering action serving as a pretreatment step to filter a plurality of eye-ROI images and establish a sample set.
1 . A line-of-sight detection method, executed by a computer processor, including:
a step SA1 that performs a dynamic-threshold-to-switch action to correct errors of data generated in a previous facial landmark detection step, the step SA1 comprising calculating normalized root mean square error (abbreviated NRMSE) based on a scene image collected, employing the NRMSE less than a first threshold value as a basis for image grouping, calculating a standard deviation of multiple facial features based on images of each group, and selecting points of the facial features having a standard deviation less than a second threshold value, so as to set a dynamic threshold value, wherein the dynamic threshold value serves as criteria for calculating a reference point of a head posture feature when the reference point switches; and
a step SA2 that performs an eye-ROI-filtering action serving as a pretreatment step to filter a plurality of eye-ROI images and establish a sample set, comprising: extracting the plurality of eye-ROI images from a facial landmark generated in the previous facial landmark detection step; emphasizing a pupil image of the plurality of eye-ROI images by using Haar features, so as to improve difference between a positive sample and a negative sample; conducting a template matching for the plurality of eye-ROI images; calculating a similarity of each of the plurality of eye-ROI images, wherein the sample set comprises eye-ROI images having a similarity greater than or equal to a third threshold value, and eye-ROI images having a similarity less than the third threshold value are bypassed from subsequent gaze estimation processing.
2 . The line-of-sight detection method of claim 1 , wherein the step SA1 is performed by
extracting key points corresponding to different angles of a head.
3 . A line-of-sight detection method, executed by a computer processor, comprising:
a step S1, detecting a face;
a step S2, detecting a facial landmark;
a step SA1, executing a dynamic-threshold-to-switch action, comprising calculating normalized root mean square error (abbreviated NRMSE) based on a scene image collected, employing the NRMSE less than a first threshold value as a basis for image grouping, calculating a standard deviation of multiple facial features based on images of each group, and selecting points of the facial features having standard a deviation less than a second threshold value, so as to set a dynamic threshold value, wherein the dynamic threshold value serves as criteria for calculating a reference point of a head posture feature when the reference point switches;
a step SA2, performing an eye-ROI-filtering action serving as a pretreatment step to filter a plurality of eye-ROI images and establish a sample set, comprising: extracting the plurality of eye-ROI images from a facial landmark generated in the previous facial landmark detection step; emphasizing a pupil image of the plurality of eye-ROI images by using Haar features, so as to improve difference between a positive sample and a negative sample; conducting a template matching for the plurality of eye-ROI images; calculating a similarity of each of the plurality of eye-ROI images, wherein the sample set comprises eye-ROI images having a similarity greater than or equal to a third threshold value, and eye-ROI images having a similarity less than the third threshold value are bypassed from subsequent gaze estimation processing;
a step S3, estimating a head pose;
a step S4, detecting a pupil; and
a step S5, estimating a gaze.
4 . The line-of-sight detection method of claim 3 , wherein the step SA1 is performed by extracting key points corresponding to different angles of a head.