Eye movement behavior intention prediction method

Abstract

The invention belongs to the technical field of eye movement tracking, and particularly relates to an eye movement behavior intention prediction method. The method comprises the following steps: acquiring eye movement data of a user at the current moment, extracting eye movement behavior characteristics of the current moment from the eye movement data, inputting the characteristics into an intention prediction model, and predicting to obtain eye movement behavior intention of the user at the current moment; the eye movement behavior characteristics comprise a gazing point dispersion degree; the gazing point dispersion degree at the t moment comprises a distance mean value from each gazing point to the center of the gazing point in the [t-omega + 1, t] time period in the [t-omega + 1, t] time period and a proportion that the distance from the gazing point to the center of the gazing point in the [t-omega + 1, t] time period is smaller than the distance mean value, and the center of the gazing point in the [t-omega + 1, t] time period is the position mean value of the gazing point in the [t-omega + 1, t] time period; omega is the size of the set time window. The eye movement behavior characteristics are extracted, information in the eye movement behaviors is fully mined, the intention of the eye movement behaviors is accurately predicted, and the man-machine interaction task driven by the eye movement behaviors can be more efficiently completed.

Description

Technical Field

[0001] The present invention belongs to the technical field of eye movement tracking, and particularly relates to a method for predicting eye movement behavior intention. Background Art

[0002] With the development of technology, great progress has been made in infrared-based pupillary corneal reflection method and non-contact eye movement tracking technology based on image analysis. Non-contact tracking avoids the discomfort and deviation that may be brought by traditional contact methods, improves the effectiveness and reliability of data, and provides powerful experimental tools and data support for multiple disciplinary fields such as human-computer interaction and user experience research. In eye movement tracking research, common eye movement behavior intentions include fixation, saccade, and smooth pursuit, etc. Among them, a common method is to fixate on a target on the operation interface for a long time until it is selected. Using eye movement behavior for interaction can be hands-free, which is particularly helpful for people with limited upper limb activities. By capturing and analyzing eye movement data and extracting more eye movement behaviors available for interaction, it provides powerful support for realizing a more intelligent and efficient human-computer interaction control method.

[0003] For example, a Chinese patent application for invention with publication number CN111308144A and publication date of June 19, 2020 discloses a method, device, and storage medium for identifying fixation behavior in a three-dimensional space. It is to identify the fixation behavior in the three-dimensional space, and the means used for identification are: selecting eye movement data for fixation behavior identification, calculating the eye movement angular velocity at multiple sampling time points in the selected data, and determining the fixation point data corresponding to the fixation behavior based on the calculated eye movement angular velocity and the eye movement angular velocity threshold corresponding to the fixation behavior. Merging all the fixation point data that meet the requirements to obtain the fixation duration and the coordinates of the fixation point, and then identifying the fixation behavior. That is to say, it mainly uses the eye movement angular velocity to identify the fixation behavior in the three-dimensional space. However, in addition to the fixation behavior, there are other intentions in human eye movement behavior. Just through one fixation behavior, it is not enough to determine the target that the subject is interested in or to achieve an effective human-computer interaction control behavior (such as driving a robot to move). Other eye movement behavior intentions should also be combined to accurately judge the target that the subject is interested in, such as saccade and intentional fixation. Among them, saccade is an eye movement behavior that inevitably exists during the transfer of the fixation target, and intentional fixation is a collection of a series of fixation behaviors within a certain period of time. Obviously, it is impossible to accurately identify all these eye movement behavior intentions only by using the eye movement angular velocity, so it is impossible to accurately identify the target that a person is interested in and the true intention, which is not efficient enough for a human-computer interaction task that needs to control the movement and operation target of a robot at the same time. Summary of the Invention

[0004] The object of the present invention is to provide a method for predicting eye movement behavior intention, so as to solve the problem in the prior art that the eye movement behavior intention other than fixation (especially the intentional eye movement behavior intention) cannot be accurately recognized, resulting in the inability to recognize the true intention of a person.

[0005] To solve the above technical problems, the present invention provides a method for predicting eye movement behavior intention, and the method includes:

[0006] Obtain the eye movement data of the user at the current moment, extract the eye movement behavior characteristics at the current moment from it and input them into the intention prediction model, and predict the eye movement behavior intention of the user at the current moment;

[0007] The eye movement behavior characteristics include the degree of dispersion of the fixation points; among them, the degree of dispersion of the fixation points at time t includes the average distance from each fixation point in the time period [t - ω + 1, t] to the center of the fixation points in the time period [t - ω + 1, t] and the proportion of the fixation points in the time period [t - ω + 1, t] whose distance to the center of the fixation points is less than the average distance. The center of the fixation points in the time period [t - ω + 1, t] is the average value of the fixation point positions in the time period [t - ω + 1, t]; ω is the set size of the time window;

[0008] The intention prediction model is obtained by training a machine learning model using a data set, and the data set contains historical eye movement behavior characteristic data and corresponding eye movement behavior classification labels.

[0009] Further, the eye movement behavior characteristics further include the average eye movement speed; the average eye movement speed at time t is the average value of the real-time eye movement speeds in the time period [t - ω + 1, t].

[0010] Further, the average eye movement speed at time t includes the average eye movement speed vector at time t and the average eye movement speed scalar at time t.

[0011] Further, the eye movement behavior characteristics further include the real-time eye movement speed; the real-time eye movement speed at time t includes the real-time eye movement speed vector at time t and the real-time eye movement speed scalar at time t.

[0012] Further, the eye movement behavior intention includes intentional fixation and intentional action. Intentional fixation refers to a set of fixation behaviors within a certain time period, and intentional action refers to the behavior of the eyes quickly saccading to complete an action or trajectory.

[0013] Further, the eye movement behavior intention further includes fixation and saccade.

[0014] Further, the machine learning model is a decision tree, a random forest or a KNN.

[0015] The present invention is an improved invention, and its beneficial effects are as follows: In order to be able to identify eye movement behaviors other than fixation, the present invention uses eye movement behavior characteristics different from the prior art for eye movement behavior intention recognition, and uses a machine learning model to analyze the internal relationship between eye movement behavior characteristics and eye movement behavior intentions, so as to achieve accurate human-computer interaction behavior recognition. Considering that in fixation-like eye movement behaviors, the fixation points will fluctuate within a small range and the fixation points are relatively concentrated, that is, the degree of dispersion is low, while in saccade-like eye movement behaviors, the fixation points are distributed along the entire saccade path and the overall distribution is relatively dispersed. Therefore, the degree of dispersion of fixation points is helpful for identifying both fixation and saccade behaviors. Therefore, the eye movement behavior characteristics different from those in the prior art in the present invention include the degree of dispersion of fixation points, and the degree of dispersion of fixation points includes the average distance of each fixation point to the center of the fixation points and the proportion of the distance from the fixation point to the center of the fixation points that is less than the average distance. Using these two parameters to comprehensively reflect the action amplitude of eye movement behavior within a period of time has a significant impact on the recognition of various eye movement behaviors including fixation and saccade, and moreover, this feature can produce better recognition effects for all eye movement behaviors, especially for eye movement behaviors such as intentional fixation and intentional actions. In summary, the eye movement behavior characteristics extracted by the present invention fully exploit the information in eye movement behaviors, realize the accurate prediction of the intentions of eye movement behaviors, and can more efficiently complete the human-computer interaction tasks driven by eye movement behaviors. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is the overall framework diagram of the eye movement behavior intention prediction method of the present invention;

[0017] Figure 2a 、 Figure 2b are schematic diagrams of two different data acquisition experiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The overall idea of the present invention is to obtain the eye movement data of the user at the current moment, extract the eye movement behavior characteristics at the current moment from it and input them into the intention prediction model, and predict the eye movement behavior intention of the user at the current moment. The eye movement behavior characteristics include the degree of fixation point dispersion; the degree of fixation point dispersion at time t includes the average distance from each fixation point in the time period [t - ω + 1, t] to the center of the fixation points in the time period [t - ω + 1, t] and the proportion of the fixation points in the time period [t - ω + 1, t] whose distance to the center of the fixation points is less than the average distance, and the center of the fixation points in the time period [t - ω + 1, t] is the average value of the fixation point positions in the time period [t - ω + 1, t]; ω is the set time window. The intention prediction model is obtained by training a machine learning model using a data set, and the data set contains historical eye movement behavior characteristic data and corresponding eye movement behavior classification labels. The degree of fixation point dispersion of the present invention includes two parameters: the average distance from each fixation point to the center of the fixation points and the proportion of the fixation points whose distance to the center of the fixation points is less than the average distance. Using these two parameters can comprehensively reflect the action amplitude of the eye movement behavior within a period of time, and the intention can accurately identify the eye movement behavior characteristics other than fixation, and more efficiently complete the human-computer interaction task driven by eye movement behavior.

[0019] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0020] An eye movement behavior intention prediction method of the present invention, its overall flow block diagram is as Figure 1 shown, and the specific process is as follows:

[0021] Step 1, construct multiple eye movement behavior data sets.

[0022] Generally, human eye movement behaviors can be divided into two categories: fixation and saccade. Fixation is the behavior in which the gaze coordinates are concentrated in a certain limited area within a certain period of time. Saccade is an eye movement behavior that necessarily exists during the transfer of the fixation target. It is not enough to determine the target of interest of the subject only through one fixation behavior. Only when the subject fixates on a target for a long time can it be determined that the subject is interested in the target, and this behavior is called intentional fixation. Intentional fixation can be considered as a collection of a series of fixation behaviors within a certain period of time. When the eyes complete a specific action or trajectory with a relatively fast saccade speed, this process is called an intentional action, and this process will include saccade data and fixation data. The existing data sets lack the data of these two behaviors, intentional fixation and intentional action, and it is necessary to construct a data set containing four eye movement behaviors: fixation, saccade, intentional fixation and intentional action.

[0023] Experimental equipment: This experiment was conducted in a closed and quiet laboratory using an aSeePro desktop eye tracker. This device is based on a three-dimensional human eye model and uses a telemetry tracking algorithm. It does not require head fixation and can accurately calculate the fixation point coordinates within the range of 55 cm - 80 cm, with an accuracy of 0.5°, and records data at a frequency of 250 Hz.

[0024] Experimental procedure: Each experimenter needed to calibrate the eye tracker through the five-point calibration method before the experiment and start data collection based on the calibration score (a score of 85 or above is considered valid data). During the experiment, the corresponding target objects were presented on the display, and the experimenter completed data collection according to the specified tasks. The experiment was divided into four parts, collecting four types of eye movement behaviors: fixation, saccade, intentional fixation, and intentional movement. Five images were set for each part of the experiment, and each person repeated the experiment 10 times.

[0025] Experiment 1: Fixation experiment. As Figure 2b shown, the experiment set a starting position (the green dot in the figure) and multiple target objects (the black dots in the figure). The experimenter quickly and randomly fixated on several targets. To avoid repetition with the intentional fixation behavior during the process, after completing one experiment, the experimenter needed to return to the initial position and then conduct the next experiment. The i-VT algorithm was used to extract the fixation behavior among them.

[0026] Experiment 2: Saccade data. The experimenter fixated on the starting position during the experiment, and then the fixation point continuously switched between two target objects. In this experiment, the fixation problem was not considered, and long-term fixation on one target would not affect the data. The saccade data between the two target objects was retained.

[0027] Experiment 3: Intentional fixation data. The experimenter first fixated on the starting point, and then fixated on a randomly selected object as the target of this fixation. When it was considered an intentional fixation, the marking button was pressed, and all subsequent fixations on the target points were regarded as intentional fixations.

[0028] Experiment 4: Intentional movement data. To adapt to controlling the movement of the robot, the intentional movement was a quick upward, leftward, rightward, and downward saccade. One target was set. The experiment started with fixation on the target. When about to make an intentional movement, the marking button was pressed, and the button was released after the movement was completed. The data during the button press was regarded as the intentional movement. As Figure 2a shown.

[0029] Step 2: Extract the eye movement behavior characteristics from the multiple eye movement behavior data sets obtained in Step 1.

[0030] In the experimental design, the number of target points is limited and the data acquisition process is centered around the target points. Machine learning models often easily learn the relationship between the fixation target points and behaviors, which is obviously not the desired result. In addition, during the data acquisition process, it is inevitable to be affected by natural behaviors such as head jitter and blinking, resulting in the original gaze data containing outliers. These outliers will interfere with the accurate learning of eye movement behaviors by machine learning models. To solve these problems, the present invention adopts a processing method based on data characteristics, abandoning the method of directly using the fixation point coordinates as input, and instead focusing on extracting and constructing a series of non-position-dependent characteristics from the fixation behaviors. These characteristics are calculated based on the distribution, duration, dynamic changes, etc. of the fixation points, aiming to comprehensively and deeply analyze the essence of eye movement behaviors, thereby guiding the machine learning model to learn a more generalized and robust model.

[0031] The single-frame data sampling time is very short, making it difficult to fully capture the complete eye movement behavior. Eye movement data needs to put ω frames together to form a time window, and each feature is calculated based on a unified length ω, more effectively extracting the key information in the eye movement behavior features.

[0032] (1) Eye movement real-time speed F vt : During the process of changing from fixation behavior to saccade behavior, the eye movement speed will increase rapidly. Due to the fast response characteristic of the real-time speed, this is beneficial to the identification of saccade behaviors. The real-time speed of the eye movement fixation point includes the eye movement real-time speed vector V t (V x (t), V y (t)) and the magnitude of the eye movement real-time speed |V t | (i.e., the eye movement real-time speed scalar), and its calculation method is as follows:

[0033] V x (t) = x(t) - x(t - ω + 1)

[0034] V y (t) = y(t) - y(t - ω + 1)

[0035]

[0036] In the formula, x(t) and x(t - ω + 1) respectively represent the fixation point coordinates in the horizontal direction at times t and t - ω + 1, and y(t) and y(t - ω + 1) represent the fixation point coordinates in the vertical direction at times t and t - ω + 1; V x (t) represents the eye movement real-time speed in the horizontal direction at time t, V y(t) represents the real-time vertical eye movement velocity at time t. If the fixation points between two consecutive frames are used to calculate the velocity, the velocity value may be very small, making it difficult to reflect the actual changes in fixation behavior. Moreover, outliers are inevitable during data acquisition, and random perturbations will affect the data. Therefore, the difference between the coordinates of two fixation points within the time window length is used to calculate the velocity, that is, ω > 1 is required.

[0037] (2) Average eye movement velocity F vt : During the transition from saccade behavior to fixation behavior, the velocity will first decrease. After transitioning to fixation behavior, the fixation point will fluctuate slightly within a certain range. Compared with the real-time velocity, the average velocity can effectively eliminate these fluctuations and reduce the influence of accidentally generated outliers on the data. Therefore, this feature is more conducive to the judgment of fixation behavior. The average eye movement velocity also includes the average eye movement velocity vector and the magnitude of the average eye movement velocity (i.e., the average eye movement velocity scalar), and the calculation method is as follows:

[0038]

[0039] That is:

[0040]

[0041] In the formula, V i represents the velocity vector at time i, including two directions, x and y.

[0042] (3) Fixation point dispersion degree F dis : During fixation behavior, the fixation point will fluctuate within a small range, and the fixation points are relatively concentrated, that is, the dispersion degree is low; during saccade behavior, the fixation points are distributed along the entire saccade path, and the overall distribution is relatively dispersed. The dispersion degree is helpful to a certain extent in identifying fixation and saccade behaviors. It is quantified as the mean value of the fixation point set within a time window [t - ω + 1, t] as the center of the fixation points within this time window, and the mean distance from each fixation point to the center is denoted as The proportion of the distance from the fixation points to the center of the fixation points within the time window less than is denoted as d per , and the magnitude of d per can reflect the movement amplitude of the eye movement behavior within the time window to a certain extent, which is the quantified value of the fixation point dispersion degree. The specific calculation method is as follows:

[0043]

[0044] In the formula, represents the mean value of the fixation coordinates in the horizontal direction within the time window [t - ω + 1, t]; It represents the mean value of the fixation coordinates in the vertical direction within the time window [t - ω + 1, t]. It represents the center of the fixation points within the time window [t - ω + 1, t]. It represents the mean value of the distances from the saccade points to the center of the saccade within a time window; the set of distances from each point to the center within the time window is D; d per It represents within the time window, the proportion of those whose distance to the center is less than of the total.

[0045] Step 3: Integrate the data obtained in Step 1 and Step 2, and divide the obtained data set into a training set and a test set.

[0046] Step 4: Build a machine learning model, and use the training set and test set obtained in Step 3 to train and test the machine learning model respectively to obtain the final intention prediction model.

[0047] Extract eye movement behavior features, use the machine learning model to learn the internal relationship between eye movement behavior features and eye movement behavior, and accurately identify four eye movement behavior intentions: fixation, saccade, intentional fixation, and intentional movement. In the robot control scenario, fixation and saccade behaviors are uniformly regarded as null operation instructions and do not trigger specific actions; intentional fixation is used as the judgment benchmark for accurately selecting targets; while intentional eye behaviors are directly converted into directional instructions for robot movement.

[0048] Among them, the specific machine learning model can adopt Random Forest (RF), and its effect is the best. As other implementation manners, KNN (K-Nearest Neighbors) or Decision Tree (DT) can also be selected.

[0049] Step 5: Obtain the eye movement data of the user at the current moment, extract the eye movement behavior features at the current moment from it and input them into the intention prediction model obtained in Step 4 to predict the eye movement behavior intention of the user at the current moment.

[0050] The calculation method of the eye movement behavior features therein is the same as that in Step 2.

[0051] Thus, the entire method is completed. Next, experiments are conducted to prove the effect of the present invention.

[0052] 1. Eye movement behavior intention experiment.

[0053] Next, four machine learning models, KNN, SVM, DT, and RF, are used to conduct classification experiments on the data set in this embodiment to prove the effect of the present invention. The experimental results are shown in Tables 1 - 3. Among them, P refers to precision, representing the overall accuracy rate; R refers to recall, representing the recall rate; and F1 refers to the F1 score.

[0054] Table 1 Accuracy of Four Methods for S with a Time Window of 25

[0055]

[0056] Table 2 Accuracy of Four Methods with a Time Window of 50

[0057]

[0058] Table 3 Accuracy of Four Methods with a Time Window of 100

[0059]

[0060] By analyzing the data in Tables 1, 2, and 3, it is found that as the time window increases, the accuracy of the four algorithms has improved. To test the impact of the time window size on the computational cost during feature extraction, 7,500 data samples in 30 seconds were selected for feature extraction experiments on the Raspberry Pi 4B platform. According to the data in Table 4, when the time window becomes larger, the computational time also increases. Therefore, a time window size of 100 can be selected. This choice not only ensures a high algorithm accuracy but also meets the real-time requirements of data processing on low-computing-power platforms.

[0061] Table 4 Computational Time for Each Step of Three Time Windows

[0062]

[0063] Tables 1, 2, and 3 detail the classification results of four different machine model methods for eye movement behavior when the time windows are 25, 50, and 100 respectively. It can be seen that as the time window increases, the accuracy of the four methods has improved. And when the time window is 100, the accuracy is the highest. Subsequently, the time window of 100 is taken for further analysis. It can be seen from Table 3 that the classification accuracy of SVM is relatively low. KNN has a certain ability to handle data imbalance problems, so the accuracy is relatively high. However, when facing an imbalanced dataset, the classification effect of the minority class of DT is poor, resulting in a decrease in the overall accuracy. The overall accuracy of RF is the highest because RF reduces the overfitting problem that may occur in a single model by integrating multiple decision trees and has a certain anti-disturbance ability for imbalanced data. At the same time, RF can measure the importance of features for the classification result, providing guidance for further optimizing the model. Therefore, RF achieves the overall best result in the eye movement behavior classification task. The computational time of the four machine learning models was compared and tested. These four algorithms were all configured to run on a single core. A total of 509,300 data from the experimental collection were used, with 80% as the training set and 20% as the test set. The time window was set to 100 as a training parameter. The computational time experiment of the four machine learning models was carried out on the Raspberry Pi 4B platform.

[0064] Table 5 Time consumption of four methods

[0065]

[0066] According to Table 5, the training time of DT is the shortest, and the result can be output almost in real time. During the training of KNN, no explicit calculation is performed, so the training is fast, but the inference speed is relatively slow. SVM not only has slow training and inference speeds, but also has a low prediction accuracy. The training time of RF is relatively long, but the inference speed is relatively fast, which can meet the requirement of completing the calculation of one frame of data within an average of 1 ms. Therefore, on the premise of meeting the inference speed, the present invention can select RF with higher prediction accuracy as the eye movement behavior prediction algorithm model. This model not only ensures the real-time performance of inference, but also provides more reliable prediction results. And it can also meet the real-time requirement on the low-computing-power platform of Raspberry Pi.

[0067] Experimental analysis shows that the optimal time window length for eye movement behavior feature calculation is 100 frames, which can ensure that each frame of data contains sufficient features without affecting the system performance; through the comparison of four machine learning methods, RF is determined to be the most suitable classification strategy for converting eye movement behavior into robot control instructions; the overall method can effectively meet the system requirements.

[0068] 2. Comparative experiment on eye movement behavior intention prediction and other methods.

[0069] Two datasets, namely GazeCom dataset and HMR dataset, are used. Four behaviors, namely fixation, saccade, smooth pursuit and invalid data, are included in both datasets. The method proposed in the present invention, which combines features with random forest machine learning, is used to conduct classification experiments on eye movement behaviors in the datasets. The sampling rates of the two datasets are almost the same as the data sampling rate in this experiment, so the time window length is set to 100. Comparative experiments are conducted with three methods, namely CNN-BiLSTM, CNN-LSTM and TCN, on the two datasets. Table 6 shows the classification performances of the four methods.

[0070] Table 6 Performance evaluation results of four models under two datasets

[0071]

[0072] Table 6 shows the overall accuracy, recall rate and F1 score of the four methods under the GazeCom and HMR datasets. The results show that the method proposed in the present invention has an accuracy improvement of 2.66%, a recall rate improvement of 2.07%, and an F1 score improvement of 1.81% on the GazeCom dataset. In the HMR dataset, the accuracy is improved by 6.12%, the recall rate is improved by 5.85%, and the F1 score is improved by 6.14%. This fully demonstrates the superiority and effectiveness of the method proposed in the present invention in eye movement behavior classification.

[0073] In summary, the present invention first constructs a dataset for annotating four eye movement behaviors, and extracts features that are conducive to the analysis of eye movement behaviors according to the characteristics of eye movement behaviors. At the same time, a machine learning algorithm is used to learn the relationship between eye movement features and eye movement behaviors, and the random forest algorithm achieves good results. Moreover, the real-time speed of eye movement has a greater impact on the determination of saccades and intentional actions, while the average speed of eye movement has a more significant impact on the recognition of the four eye movement behaviors. The fixation dispersion improves the recognition of all four behaviors, but has a more significant effect on intentional behaviors. Among them, intentional fixation and intentional action behaviors can be used as modalities for human-computer interaction. At the same time, the method proposed by the present invention is verified on the GazeCome and HMR datasets, and its accuracy exceeds that of the Temporal Convolutional Network (TCN). These results not only verify the effectiveness of the research method, but also demonstrate its generalization in the field of eye movement behavior analysis.

[0074] The specific implementation manners are given above, but the present invention is not limited to the described implementation manners. The basic idea of the present invention lies in the above basic solution. For those of ordinary skill in the art, according to the teachings of the present invention, it does not require creative labor to design various deformed models, formulas, and parameters. Changes, modifications, substitutions, and variations made to the implementation manners without departing from the principle and spirit of the present invention still fall within the protection scope of the present invention.

Claims

1. A method for predicting eye movement behavior intention, characterized in that: The method includes: Obtain the user's current eye movement data, extract the eye movement behavior features at the current moment and input them into the intention prediction model to predict the user's eye movement behavior intention at the current moment; The eye movement behavior characteristics include the degree of gaze point dispersion; wherein, the degree of gaze point dispersion at time t includes the average distance from each gaze point to the center of the gaze point in the time period [t-ω+1, t] and the proportion of the distance from the gaze point to the center of the gaze point in the time period [t-ω+1, t] that is less than the average distance; the center of the gaze point in the time period [t-ω+1, t] is the average position of the gaze point in the time period [t-ω+1, t]; ω is the set time window size; The intention prediction model is obtained by training a machine learning model using a data set, and the data set includes historical eye movement behavior feature data and corresponding eye movement behavior classification labels.

2. The eye movement behavior intention prediction method according to claim 1, characterized in that: The eye movement behavior characteristics also include an average eye movement speed; the average eye movement speed at time t is the average of the real-time eye movement speeds in the time period [t-ω+1, t].

3. The eye movement behavior intention prediction method according to claim 2, characterized in that: The average eye movement velocity at time t includes the average eye movement velocity vector at time t and the average eye movement velocity scalar at time t.

4. The eye movement behavior intention prediction method according to claim 1, characterized in that: The eye movement behavior characteristics also include real-time eye movement speed; the real-time eye movement speed at time t includes the real-time eye movement speed vector at time t and the real-time eye movement speed scalar at time t.

5. The eye movement behavior intention prediction method according to any one of claims 1 to 4, characterized in that: Eye movement behavior intention includes intentional gaze and intentional movement. Intentional gaze refers to the collection of a series of gaze behaviors within a certain period of time, and intentional movement refers to the behavior of scanning the eyes at a relatively fast speed to complete an action or trajectory.

6. The eye movement behavior intention prediction method according to claim 5, characterized in that: Eye movement behavior intention also includes fixations and saccades.

7. The eye movement behavior intention prediction method according to any one of claims 1 to 4, characterized in that: The machine learning model is a decision tree, random forest or KNN.

Images