Background
Attention Deficit Hyperactivity Disorder (ADHD), commonly known as hyperactivity disorder, is the most common neurological disorder in childhood in teenagers, and its clinical manifestations are difficulty in focusing on, overactivity, impulsivity, instability of emotion, difficulty in learning, etc. The existing intelligent recognition technology of ADHD medical images is mainly based on research on pathological aspects such as brain functional nuclear magnetic resonance and electroencephalogram, or on observing the behavior characteristics of a patient solely based on eye movement or facial expression, etc.
At present, two main researches and technical routes of the eye movement characteristics of the children of ADHD exist, one is to track the eye movement condition and the gaze point position of the children through a head-mounted eye movement instrument, the other is to limit the positions of the body and the head of the children, and then the gaze point position of the children is calculated by using a formula with fixed parameters. The former has disadvantages in that the child needs to wear an eye tracker, which is somewhat inconvenient to operate, and the eye tracker is easily removed or changed in position by the child, resulting in inaccurate results. The latter has the disadvantage that the movement of the child needs to be limited, in which case part of the child's behavioral characteristics are difficult to characterize and the calculation of eye movements after the child has moved substantially also generates errors.
Disclosure of Invention
In order to overcome the defects of inconvenient operation and poor accuracy of the existing ADHD child eye movement characteristic determination mode, the invention provides the eye movement attention characteristic vector determination method for the children ADHD screening and evaluating system, which is convenient to operate and has good accuracy, and the eye movement attention is measured and calculated on the premise of not limiting and influencing the activities of the children with ADHD.
The technical scheme adopted for solving the technical problems is as follows:
an eye movement attention feature vector determination method for a children ADHD screening and evaluation system is characterized in that the input is acquired front video V Front And camera calibration matrix C r Output is the generated eye movement attention characteristic vector F g Comprising the following steps:
step 1 pupil position calculation: firstly, confirming the face position by an HOG-based algorithm, and detecting the key points of the face by a continuous conditional neural field model framework to calculate the pupil position e in the plane h The method comprises the steps of carrying out a first treatment on the surface of the Then, the EPnP algorithm is utilized to align the detected face with the average standard 3D face model F, and the rotation matrix R of the head part under the camera coordinate system is calculated r Translation vector t r Output is eye space position e r =t r +e h ;
Step 2, calculating the sight line direction, wherein the sight line direction characteristic is expressed as a two-dimensional vector g comprising a yaw angle and a pitch angle, the yaw angle is an included angle between the sight line direction and a vertical plane, and the pitch angle is an included angle between the sight line direction and a horizontal plane;
step 3, screen position conversion: determining the spatial position of the plane of the screen by an external calibration mode, and according to the sight direction characteristicsVector g and eye position e r The intersection point of the sight line and the screen plane can be calculated, namely the falling point p of the sight line on the screen s Obtaining an eye attention area r according to the screen structure;
step 4, combining the pupil position, the sight line direction and the screen attention to obtain a final eye movement attention feature vector, which is expressed as F g =[e r ,g,r]。
Preferably, in the step 2, normalization processing is performed on the eye image: multiplying the original image by the inverse of the camera projection matrix
Converting the head pose into a three-dimensional space; then the original posture of the head at the moment is multiplied by a transformation matrix M to fix the eye position, and finally the normalized eye posture is multiplied by a standard camera projection matrix C
n Obtaining a normalized two-dimensional eye image e;
to calculate the line of sight vector, the rotation matrix R is normalized n =M R r And converting the two-dimensional rotation vector h, inputting the obtained 2D head posture information h and the single-channel gray eye image e into a convolutional neural network model, and outputting a sight direction feature vector which is a two-dimensional vector g containing a yaw angle and a pitch angle in the step.
The beneficial effects of the invention are mainly shown in the following steps: no wearing equipment is needed, so that children can not feel abnormal; the activities of children do not need to be limited, so that the children can be more naturally represented; the measurement accuracy is higher, and the error angle can be controlled within 5 degrees.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
Referring to FIG. 1, an eye movement attention feature for a child ADHD screening assessment systemThe vector determination method is characterized in that the input is the collected front video V Front And camera calibration matrix C r Output is the generated eye movement attention characteristic vector F g Comprising the following steps:
step 1 pupil position calculation: first, confirming face position by HOG-based algorithm, and detecting face key point by Continuous Conditional Neural Field (CCNF) model frame, thereby calculating pupil position e in plane h The method comprises the steps of carrying out a first treatment on the surface of the Then, the EPnP algorithm is utilized to align the detected face with the average standard 3D face model F, and the rotation matrix R of the head part under the camera coordinate system is calculated r Translation vector t r The output of this step is the eye space position e r =t r +e h ;
Step 2, calculating the sight line direction, wherein the sight line direction characteristic can be expressed as a two-dimensional vector g comprising a yaw angle (yaw) and a pitch angle (pitch), the yaw angle is an included angle between the sight line direction and a vertical plane, and the pitch angle is an included angle between the sight line direction and a horizontal plane;
in order to acquire the sight line vector, normalization processing is performed on the eye image: multiplying the original image by the inverse of the camera projection matrix
Converting the head pose into a three-dimensional space; then the original posture of the head at the moment is multiplied by a transformation matrix M to fix the eye position, and finally the normalized eye posture is multiplied by a standard camera projection matrix C
n Obtaining a normalized two-dimensional eye image e; to calculate the line of sight vector, the rotation matrix R is normalized
n =M R
r Converting the two-dimensional rotation vector h, inputting the obtained 2D head gesture information h and the single-channel gray eye image e into a convolutional neural network model, and outputting a sight direction feature vector which is a two-dimensional vector g containing a yaw angle and a pitch angle;
step 3, screen position conversion: by means of external calibration, the spatial position of the plane of the screen can be determined, and the position of eyes and the position of the screen can be determined according to the sight direction characteristic vector gE is arranged r The intersection point of the sight line and the screen plane can be calculated, namely the falling point p of the sight line on the screen s Obtaining an eye attention area r according to the screen structure;
step 4, combining the pupil position, the sight line direction and the screen attention to obtain a final eye movement attention feature vector, which is expressed as F g =[e r ,g,r]。
Fig. 2 illustrates the relationship between a camera and a tested person in the children ADHD screening and evaluating system, and the camera No. 1 behind the computer display screen is used for capturing the front image of the tested person and collecting the eye movement and expression change information of the tested person. The 2 and 3 binocular depth camera modules positioned on the side face of the seat of the tester can be used for shooting the whole body of the tester and collecting three-dimensional body posture information of the tester. In this embodiment, a camera (camera 1) placed in front of the face of the child is used to capture the eye movement of the child, and finally a video V is obtained fr ont。
In this embodiment, the camera matrix C of the camera is calibrated by using the method of orthogonal constraint of the plane mirror-display r 。
The input of eye movement and attention characteristic vector calculation is the collected front video V front And camera calibration matrix C r Output is the generated eye movement attention characteristic vector F g . The concrete calculation mainly comprises three steps:
pupil position calculation: the face position is first confirmed by an HOG based algorithm. Detection of facial keypoints is achieved through a Continuous Conditional Neural Field (CCNF) model framework, whereby we can calculate pupil position e in the plane h . Then, the EPnP algorithm is utilized to align the detected face with the average standard 3D face model F, and the rotation matrix R of the head part under the camera coordinate system is calculated r Translation vector t r The output of this step is the eye space position e r =t r +e h 。
In order to acquire the sight line vector, normalization processing is performed on the eye image: multiplying the original image by the inverse of the camera projection matrix
Converting the head pose into a three-dimensional space; then the original posture of the head at the moment is multiplied by a transformation matrix M to fix the eye position, and finally the normalized eye posture is multiplied by a standard camera projection matrix C
n And obtaining a normalized two-dimensional eye image e. To calculate the line of sight vector, we normalize the rotation matrix R
n =M R
r Is transformed into a two-dimensional rotation vector h. The 2D head posture information h and the single-channel gray eye image e obtained in the previous step are input into a convolutional neural network model, and a sight direction feature vector is output in the step, wherein the feature vector is a two-dimensional vector g containing a yaw angle and a pitch angle.
Screen position conversion: by means of external calibration, we can determine the spatial position of the plane in which the screen is located according to the line of sight angle g and eye position e r The intersection point of the sight line and the screen plane can be calculated, namely the falling point p of the sight line on the screen s And obtaining an eye attention area r according to the screen structure.
The pupil position, the sight line direction and the screen attention are combined to obtain the final eye movement attention characteristic vector which is expressed as F g =[e r ,g,r]。
The embodiments described in this specification are merely illustrative of the manner in which the inventive concepts may be implemented. The scope of the present invention should not be construed as being limited to the specific forms set forth in the embodiments, but the scope of the present invention and the equivalents thereof as would occur to one skilled in the art based on the inventive concept.