A network training software platform based on behavior characteristic recognition

Abstract

The application relates to a network training software platform based on behavior feature recognition, which comprises a training node management module for storage, a course presentation module for presenting a course interface of a current training node, a node anchor point coding module for generating a current challenge seed and generating a corresponding visual anchor point event sequence, an image acquisition module for acquiring continuous image frames of students, a behavior feature recognition module for extracting human body key point motion trajectories and outputting student behavior sub-action sequences, a coupling behavior checking module for coupling matching and outputting a node coupling checking result, a node continuity checking module for switching time intervals and outputting a node continuity checking result, a flow control module for judging whether to open a next training node or to re-encode a challenge, and a node evidence module for generating a node evidence summary of the current training node. The application realizes node-level coupling checking of interface presentation events of a current training node and student behavior responses.

Description

Technical Field

[0001] This invention relates to the fields of online training and computer vision technology, and in particular to an online training software platform based on behavioral feature recognition. Background Technology

[0002] With the increasing application scenarios of online training and learning, training platforms typically need to acquire learners' learning status information during course implementation to manage the delivery of training content and subsequent processes. Existing technologies already include solutions that analyze learners' emotional state, cognitive state, and comprehension level by collecting information such as facial expressions, eye movements, gaze points, and head postures during the learning process. These solutions then select subsequent task data after the completion of the current task data set, enabling the delivery and management of the learning and training plan. For example, CN108428202A discloses "A Method and System for Delivering and Managing Learning and Training Plans," which falls under this technical approach. This existing technology can already provide a basis for organizing and delivering subsequent training content based on the sensor data collected during the learning phase.

[0003] However, the key technical focus of the aforementioned existing technologies lies primarily in assessing the learner's emotional state, cognitive state, or comprehension rate based on information collected during the learning process, such as facial expressions, eye movements, fixation points, and head postures, and adjusting subsequent task data accordingly. While this approach can reflect the learner's overall learning state in online training scenarios, it lacks a technical mechanism that can bind and couple the interface presentation events of the current training node with the learner's behavioral responses at the node level. In other words, existing technologies focus more on judging general learning states rather than verifying "what is presented on the interface and whether the learner has responded accordingly" at a specific training node.

[0004] In this context, relying solely on general gaze, facial expressions, or head posture analysis is often insufficient to directly and reliably determine whether trainees have provided genuine and effective behavioral responses around the current training stage. It also fails to provide a stable basis for controlling the release, retry, and documenting of training stages. These shortcomings become even more pronounced in application scenarios that demand higher standards of authenticity, continuity, and traceability in the training process.

[0005] Therefore, it is necessary to provide a new network training software platform based on behavioral feature recognition to at least solve the problem that the existing technology lacks a technical mechanism for node-level coupling and verification of the interface presentation events of the current training node and the student's behavioral response. Summary of the Invention

[0006] To overcome the aforementioned technical deficiencies, the present invention aims to provide a network training software platform based on behavioral feature recognition. The present invention generates a visual anchor point event sequence corresponding to the current training node in the course interface, acquires continuous image frames of the trainee and extracts the motion trajectory of human key points and the trainee's behavioral sub-action sequence, and then performs coupling matching based on the spatiotemporal relative relationship between the human key point motion trajectory and the visual anchor point event sequence, and compares it with a node behavior script template. This achieves node-level coupling verification of the interface presentation events of the current training node and the trainee's behavioral responses.

[0007] This invention discloses a network training software platform based on behavioral feature recognition, comprising: The training node management module is used to store multiple training nodes and the node behavior script templates corresponding to each training node; The course presentation module is used to present the course interface of the current training node to the student's terminal; The node anchor point encoding module is used to generate the current challenge seed based on the node identifier of the current training node and the current terminal session identifier, and to generate the visual anchor point event sequence corresponding to the current training node in the course interface based on the current training node and the current challenge seed. The image acquisition module is used to acquire continuous image frames of trainees within the node response window corresponding to the visual anchor point event sequence; The behavior feature recognition module is used to perform behavior feature recognition on continuous image frames to extract the motion trajectory of human key points and output the trainee's behavior sub-action sequence; The coupling behavior verification module is used to perform coupling matching based on the spatiotemporal relative relationship between the human body key point motion trajectory and the visual anchor point event sequence, and output the node coupling verification result of the current training node. The node continuity verification module is used to output the node continuity verification result of the current training node based on the termination behavior status of the previous training node, the initial behavior status of the current training node, and the node switching time interval. The process control module is used to open the next training node when both the node coupling verification result and the node continuity verification result are passed, and to trigger the recoding challenge of the current training node when either verification result fails. The node evidence storage module is used to generate a node evidence summary for the current training node based on the current challenge seed, the sequence of visual anchor events, the sequence of trainee behavior sub-actions, the node coupling verification result, the node continuity verification result, and the node evidence summary of the previous training node.

[0008] Preferably, the node anchor point encoding module is used to determine at least two anchor point carriers from the knowledge point title area, progress indicator area, subtitle area and operation prompt icon area of ​​the course interface, and configure the display position, display order, display duration and display level of each anchor point carrier based on the current challenge seed, so as to generate a visual anchor point event sequence.

[0009] Preferably, the node behavior script template includes a directed acyclic graph consisting of multiple behavior sub-action nodes and multiple allowed transfer edges. Each behavior sub-action node stores a specified body part, a corresponding anchor point carrier, a target relative relationship type, and a hold duration threshold. Each allowed transfer edge stores a maximum transfer duration threshold. The coupled behavior verification module is used to perform graph matching verification on the trainee's behavior sub-action sequence based on the directed acyclic graph.

[0010] Preferably, the target relative relationship type includes at least the relative direction relationship, relative distance interval relationship, and relative dwelling area relationship of the specified body part relative to the corresponding anchor point carrier; the coupling behavior verification module determines that the corresponding behavior sub-action is successfully matched only when the specified body part continuously satisfies the relative direction relationship, relative distance interval relationship, and relative dwelling area relationship to the duration threshold.

[0011] Preferably, the training node management module is also used to store one-to-one corresponding prompt voice segments for each behavior sub-action node; the course presentation module is used to synchronously play the prompt voice segments when presenting the course interface of the current training node; the coupled behavior verification module performs graph matching verification on the corresponding behavior sub-action node only within the effective reaction interval formed by the start and end time of the corresponding prompt voice segment playback.

[0012] Preferably, the behavior feature recognition module includes a key point trajectory recognition unit and a gaze trajectory recognition unit; The key point trajectory recognition unit is used to extract the motion trajectory of human key points of a specified body part from continuous image frames. The gaze trajectory recognition unit is used to extract the sequence of trainees' gaze resting areas from consecutive image frames; The behavior feature recognition module outputs the corresponding sub-action only when the movement trajectory of the human body's key points and the sequence of gaze resting areas both correspond to the same anchor point carrier, and the time offset does not exceed the preset upper limit.

[0013] Preferably, the image acquisition module opens the node response window when the first anchor carrier in the visual anchor event sequence switches from a hidden state to a displayed state, and caches a preset number of preparatory image frames before opening; and closes the node response window when the last anchor carrier in the visual anchor event sequence switches from a displayed state to a hidden state, and caches a preset number of subsequent image frames after closing.

[0014] Preferably, the behavior feature recognition module further includes an optical flow recognition unit; the key point trajectory recognition unit, the gaze trajectory recognition unit, and the optical flow recognition unit each output candidate behavior sub-actions; the coupled behavior verification module adopts the candidate behavior sub-action as a valid behavior sub-action only when the outputs of the key point trajectory recognition unit and the optical flow recognition unit for the same candidate behavior sub-action are consistent, and the gaze trajectory recognition unit gives a dwell result for the anchor point carrier corresponding to the same candidate behavior sub-action.

[0015] Preferably, the node anchor encoding module is also used to update the current challenge seed based on the node evidence summary of the previous training node, the node identifier of the current training node, and the current terminal session identifier; the process control module calls the node anchor encoding module when any verification result fails, and regenerates a re-encoded visual anchor event sequence that is different from the original visual anchor event sequence based on the updated current challenge seed.

[0016] Preferably, after the node coupling verification result fails and a recoding challenge is triggered, the process control module inserts a correction node corresponding to the behavior sub-action that caused the node coupling verification result to fail before the current training node, and only restarts the current training node after the correction node passes.

[0017] Preferably, the behavior feature recognition module further includes a failure sample suppression unit, which is used to retrieve the nearest failure trajectory cluster in the failure sample library according to the node identifier of the current training node and the current challenge seed, and perform suppression output on the candidate behavior sub-actions that match the nearest failure trajectory cluster, so as to distinguish the behavior feature recognition result corresponding to the recoding challenge from the failure behavior pattern corresponding to the nearest failure trajectory cluster.

[0018] Preferably, the node continuity verification module is used to generate a termination behavior state signature by extracting the human body orientation, head posture and preset body part spatial intervals from the last consecutive frames of the previous training node, and to generate an initial behavior state signature by extracting the human body orientation, head posture and preset body part spatial intervals from the first consecutive frames of the current training node; the node continuity verification result is output only when the difference between the termination behavior state signature and the initial behavior state signature does not exceed a first threshold and the node switching time interval does not exceed a second threshold.

[0019] Preferably, the node evidence storage module is used to divide the continuous image frames and the visual anchor point event sequence into multiple corresponding segments according to the time sequence, generate a segment summary for each corresponding segment, and then generate the node root summary of the current training node based on the multiple segment summaries.

[0020] Preferably, the node evidence storage module is further used to generate a first source summary and a second source summary for the visual anchor event sequence and the trainee behavior sub-action sequence, respectively, and to cascade the first source summary, the second source summary, the node root summary, and the node evidence summary of the previous training node to generate a chain evidence summary of the current training node, which serves as the node evidence summary of the current training node.

[0021] Preferably, the course presentation module is used to display pose normalization markers in the boundary area of ​​the course interface; the behavior feature recognition module is used to perform camera pose normalization processing on the motion trajectory of human key points based on the pose normalization marker positions and face region positions in consecutive image frames; and the coupling behavior verification module is used to perform coupling matching based on the normalized human key point motion trajectory.

[0022] Compared with existing technologies, the above technical solution has the following advantages: 1. Existing technologies typically rely on general information such as facial expressions, eye movements, gaze points, or head posture to determine a learner's learning status, making it difficult to establish a stable and clear correspondence between the interface content presented at the current training node and the learner's actual behavioral responses. This invention generates a corresponding visual anchor point event sequence at the current training node and couples it with the movement trajectories of key human points, the sequence of areas where the learner's gaze lingers, and node behavior script templates. This creates a node-level behavior verification mechanism around the current training node, significantly improving the accuracy of learner behavior response determination.

[0023] 2. In existing technologies, even if the trainee's focused state or general interaction state can be identified, it is difficult to effectively determine whether the trainee has actually completed the corresponding behavioral response around the current training node. In particular, it is difficult to identify abnormal situations such as a substitute operator taking over briefly, leaving their seat during node switching, or misalignment of action sequence. This invention verifies the behavioral connection between the previous training node and the current training node through a node continuity verification module, and combines the node coupling verification results to jointly control the training process, thereby significantly improving the ability to verify the authenticity and continuity of the training process.

[0024] 3. Existing technologies are prone to misjudgment or omission when faced with repetitive error paths, deliberate avoidance behaviors, or mechanical repetitive responses based on memory. This invention introduces a recoding challenge mechanism. When a verification fails, a recoded visual anchor event sequence, different from the original visual anchor event sequence, is regenerated based on the updated current challenge seed. Furthermore, a failure sample suppression unit suppresses the output of nearest-neighbor failure trajectory clusters, thereby effectively reducing the probability of trainees repeatedly failing along the same error path and improving the system's ability to identify and suppress avoidance behaviors and repeated failure patterns.

[0025] 4. Existing technologies typically only allow for rough adjustments to subsequent training content based on the overall learning status, lacking fine-grained process control capabilities oriented towards node-level behavioral outcomes. This invention uses both node coupling verification results and node continuity verification results as the basis for process control. When verification passes, the next training node is opened; when verification fails, a recoding challenge is triggered. Corrective nodes can be inserted based on failed behavior sub-actions, thereby achieving more targeted node release control, retry control, and corrective control, ultimately improving the reliability of training process control.

[0026] 5. Existing technologies typically only repeat the original content after a trainee's response fails, lacking a targeted correction mechanism based on the cause of the failure. This invention, when the node coupling verification result fails, can insert a correction node corresponding to the failed behavior sub-action, and reopen the current training node after the correction node passes. Simultaneously, failure sample suppression and recoding challenges prevent repeated repetition of the same failure pattern, thereby improving the correction efficiency after trainee retries and the node recovery success rate.

[0027] 6. In existing technologies, the behavior recognition results of a single path are easily affected by gaze drift, partial occlusion, camera installation angle deviation, or occasional motion interference. This invention, through multi-channel output from a key point trajectory recognition unit, a gaze trajectory recognition unit, and an optical flow recognition unit, and by performing consistency judgment on candidate behavior sub-actions, can reduce the impact of single-path recognition errors on the final judgment result and improve the stability of behavior recognition.

[0028] 7. Differences in camera position, display angle, and installation height often exist between different student terminals, which can easily affect the consistency of behavior recognition results. This invention reduces geometric errors between different terminal devices and improves the consistency of behavior verification results under different terminal conditions by setting posture normalization markers in the boundary area of ​​the course interface and performing camera posture normalization processing on the motion trajectory of human key points based on the position of the posture normalization markers and the position of the face region.

[0029] 8. Existing technologies typically focus on process or result recording, lacking data structures capable of supporting node-level review and link verification. This invention divides continuous image frames and visual anchor event sequences into multiple corresponding segments and generates segment summaries, node root summaries, first source summaries, second source summaries, and chain evidence summaries. This enables current training nodes to form a traceable and verifiable chain of node evidence summaries, thereby significantly improving node-level review efficiency, disputed node verification capabilities, and training audit convenience.

[0030] 9. This invention is applicable not only to knowledge-transmission training scenarios such as cybersecurity compliance training, but also to vocational skills training scenarios such as equipment inspection standard training. Furthermore, this invention is applicable to training nodes that include hand interaction actions, as well as purely visual training nodes that complete responses based on head orientation and the sequence of trainees' gaze areas. Therefore, it has strong scenario expansion capabilities and implementation flexibility. Attached Figure Description

[0031] Figure 1 A schematic diagram of the overall structure of the online training software platform; Figure 2 A schematic diagram of node coupling verification and node behavior script template; Figure 3 A schematic diagram of the node continuity verification and recoding challenge process; Figure 4 A schematic diagram for node evidence storage and chain evidence digest generation; Figure 5 A bar chart comparing key performance indicators; Figure 6 A graph showing the variation of node continuity difference as a function of node switching time interval; Figure 7 A graph showing the change in repeated failure rate before and after the recoding challenge; Figure 8 This is a schematic diagram illustrating the temporal correspondence between the visual anchor event sequence and the behavioral response. Figure 9 This is a schematic diagram illustrating multi-channel consistency determination and the generation of valid sub-actions. Figure 10 This is a schematic diagram illustrating the suppression of failed samples and the updating of nearest-neighbor failed trajectory clusters. Figure 11 A schematic diagram illustrating the generation and verification of chained evidence summaries; Figure 12 This is a schematic diagram of pose normalization labeling and camera pose normalization processing. Detailed Implementation

[0032] The present invention will now be described in further detail with reference to the accompanying drawings. It should be noted that the following embodiments are only used to illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention. Where there is no conflict, the technical features in the following embodiments can be combined with each other.

[0033] To facilitate understanding by those skilled in the art, a general overview is provided first: The structure and collaboration of the training node management module, course presentation module, node anchor encoding module, image acquisition module, behavior feature recognition module, coupled behavior verification module, node continuity verification module, process control module, and node evidence storage module mainly correspond to the overall platform solution of this invention; the descriptions of visual anchor event sequences, node behavior script templates, target relative relationship types, prompt voice segments, effective response intervals, and node response windows are mainly used to support node-level coupled verification related solutions; the descriptions of recoding challenges, failure sample suppression units, nearest neighbor failure trajectory clusters, and correction nodes are mainly used to support anti-repeated failure and retry optimization related solutions; the descriptions of fragment summaries, node root summaries, first source summaries, second source summaries, chain evidence summaries, and background review call scenarios are mainly used to support node evidence storage related solutions; the descriptions of pose normalization marking, camera pose normalization processing, and geometric normalization under different terminal conditions are mainly used to support cross-terminal consistency processing related solutions.

[0034] Example 1: Combination Figure 1The network training software platform in this embodiment includes a training node management module, a course presentation module, a node anchor point encoding module, an image acquisition module, a behavior feature recognition module, a coupled behavior verification module, a node continuity verification module, a flow control module, and a node evidence storage module. The training node management module stores multiple training nodes and their corresponding node behavior script templates, prompt voice clips, correction node content, failure sample suppression parameters, and re-encoding parameters. The course presentation module presents the course interface of the current training node to the student's terminal. The node anchor point encoding module generates a current challenge seed based on the node identifier and current terminal session identifier of the current training node, and generates a visual anchor point event sequence corresponding to the current training node in the course interface based on the current challenge seed. The image acquisition module acquires continuous image frames of the student within the node response window corresponding to the visual anchor point event sequence. The behavior feature recognition module performs behavior feature recognition on the continuous image frames to extract the motion trajectory of key human points and the student's gaze area. The system includes a domain sequence and trainee behavior sub-action sequences; a coupled behavior verification module is used to perform coupling matching based on the spatiotemporal relative relationship between the human body keypoint motion trajectory and the visual anchor point event sequence, and to compare with the node behavior script template; a node continuity verification module is used to output the node continuity verification result based on the termination behavior state of the previous training node, the initial behavior state of the current training node, and the node switching time interval; a flow control module is used to control the release of the next training node, the retry of the current training node, and the insertion of correction nodes based on the node coupling verification result and the node continuity verification result; and a node evidence storage module is used to perform fragmented summarization, node root summary generation, and chain evidence summary generation on continuous image frames, visual anchor point event sequences, trainee behavior sub-action sequences, and verification results.

[0035] In this embodiment, a "Cybersecurity Compliance Training Course" is used as an example. The entire course is divided into multiple training nodes, such as a "Login Environment Confirmation Node," a "Sensitive Information Identification Node," an "External Approval Operation Node," and a "Graduation Confirmation Node." Each training node corresponds to a minimum verification unit, which includes not only the presentation of course content but also the visual anchor event sequence bound to the course content and the behavioral responses that the trainee should complete. Taking the "External Approval Operation Node" as an example, the course interface of this training node includes a knowledge point title area, a progress indicator area, a subtitle area, and an operation prompt icon area. The node anchor encoding module generates the visual anchor event sequence corresponding to the current training node in the above different areas. The visual anchor event sequence is not fixed but dynamically determines the display position, display order, display duration, and display level of the anchor carrier based on the current challenge seed. Therefore, the visual anchor event sequence of the trainee can change in different terminal sessions, thereby forming a stronger node-level binding relationship between the trainee's behavioral responses and the visual anchor event sequence in the current training node.

[0036] In a preferred implementation, the display position of the anchor carrier can be selected from a preset set of positions within the corresponding area. The number of candidates in the preset set can be 3 to 12, preferably 4 to 8. The display order of the anchor carriers can be arranged from 2 to 6 event nodes, preferably 3 to 5 event nodes. The display duration of a single anchor carrier can be 0.3s to 2.5s, preferably 0.6s to 1.8s. The display layer can be any one of the foreground layer, middle layer, and background layer, or a combination thereof. In this embodiment, the first anchor carrier in the knowledge point title area is displayed for 1.2s, the second anchor carrier in the operation prompt icon area is displayed for 1.0s, and the third anchor carrier in the subtitle area is displayed for 1.5s.

[0037] The training node management module pre-stores node behavior script templates. These templates can be pre-configured by the course creator based on training content, interface elements, and expected behavioral responses. Alternatively, administrators can generate the templates by entering the behavior sub-action nodes, allowed transition edges, target relative relationship types, hold duration thresholds, maximum transition duration thresholds, and corresponding anchor point carrier information for each training node through the backend configuration interface. Combined with... Figure 2 The node behavior script template can adopt a directed acyclic graph structure. Taking the "external approval operation node" as an example, the node behavior script template includes multiple sub-action nodes and multiple allowed transition edges. The specified body part corresponding to sub-action node A is the head, the corresponding anchor carrier is the first anchor carrier in the knowledge point title area, the target relative relationship type is the relative direction relationship of the head relative to the first anchor carrier, and the holding duration threshold is set to 0.6s; the specified body part corresponding to sub-action node B is the right hand, the corresponding anchor carrier is the second anchor carrier in the operation prompt icon area, the target relative relationship type is the relative distance interval relationship of the right hand relative to the second anchor carrier, and the holding duration threshold is set to 0.4s; the specified body part corresponding to sub-action node C is the gaze, the corresponding anchor carrier is the third anchor carrier in the subtitle area, the target relative relationship type is the relative dwell area relationship of the gaze relative to the third anchor carrier, and the holding duration threshold is set to 0.8s. Transition edges are allowed to constrain the maximum transition time between adjacent action nodes. For example, the maximum transition time from action node A to action node B is set to 1.5 seconds, and the maximum transition time from action node B to action node C is set to 1.8 seconds. The coupled behavior verification module performs graph matching verification on the trainee's action sequence based on the node's behavior script template, thereby determining whether the trainee has completed a true and effective behavioral response around the visible anchor event sequence of the current training node.

[0038] It should be noted that, in this embodiment, the holding time threshold can be set to 0.3s to 1.5s, preferably 0.4s to 1.0s, and in this embodiment, it is taken as 0.6s, 0.4s and 0.8s respectively; the maximum transfer time can be set to 0.8s to 3.0s, preferably 1.2s to 2.0s, and in this embodiment, it is taken as 1.5s and 1.8s.

[0039] In a preferred embodiment, the target relative relationship types include relative direction relationship, relative distance interval relationship, and relative dwelling area relationship. The relative direction relationship describes whether the orientation of a specified body part points to the corresponding anchor point carrier; the relative distance interval relationship describes whether the distance between the specified body part and the corresponding anchor point carrier is within the allowed range; and the relative dwelling area relationship describes whether the trainee's gaze continuously remains within the valid window of the corresponding anchor point carrier. The coupled behavior verification module determines that the behavior sub-action node is successfully matched only when the specified body part continuously satisfies the corresponding relationship and reaches the holding time threshold. This excludes short-term random actions, occasional gaze drift, and meaningless hand movements from valid behavior responses.

[0040] The training node management module also configures a one-to-one corresponding prompt voice clip for each behavioral sub-action node. When the course presentation module presents the corresponding anchor carrier in the course interface, it synchronously plays the prompt voice clip corresponding to that behavioral sub-action node. The coupled behavior verification module performs graph matching verification on the corresponding behavioral sub-action node only within the effective reaction interval formed by the start and end times of the prompt voice clip playback. The length of the effective reaction interval can be 0.5s to 3.0s, preferably 0.8s to 2.0s. In this embodiment, when the prompt voice clip starts playing at 3.0s and ends at 4.2s, the effective reaction interval can be set to 3.0s to 5.0s. By binding the voice prompts with the behavior verification time window, the interference of students' early or late actions on the verification results can be reduced.

[0041] The image acquisition module opens the node response window when the first anchor carrier in the visual anchor event sequence switches from a hidden state to a visible state, and caches a preset number of preparatory image frames before opening. When the last anchor carrier in the visual anchor event sequence switches from a visible state to a hidden state, the node response window closes, and caches a preset number of subsequent image frames after closing. The number of preparatory image frames can be 5 to 30 frames, preferably 10 to 20 frames; the number of subsequent image frames can be 5 to 25 frames, preferably 8 to 18 frames. In this embodiment, the number of preparatory image frames is 15 frames, and the number of subsequent image frames is 10 frames. This covers both the preparatory actions before the student responds and the concluding actions after the response, facilitating subsequent node continuity verification and fragmented node evidence storage.

[0042] The behavioral feature recognition module includes a keypoint trajectory recognition unit, a gaze trajectory recognition unit, and an optical flow recognition unit. The keypoint trajectory recognition unit extracts the motion trajectories of key points for specified body parts from consecutive image frames; the gaze trajectory recognition unit outputs a sequence of gaze-holding regions, which is the result sequence output by the gaze trajectory recognition unit; the optical flow recognition unit extracts the motion trends of local areas to help confirm the direction and degree of completion of actions. The sequence of gaze-holding regions can be obtained in various ways, such as using an eye-tracking estimation model to obtain the gaze point and then mapping it chronologically to form the sequence; or using a gaze regression model, taking eye images and head posture as input, and outputting a gaze region sequence corresponding to the screen reference coordinate system; or using a geometric mapping method based on facial keypoints and head posture to recover the gaze-holding region relative to the screen reference plane.

[0043] The behavior feature recognition module only outputs the corresponding sub-action when the movement trajectory of the human body's key points and the sequence of the trainee's gaze placement areas both correspond to the same anchor point carrier, and the time offset does not exceed a preset upper limit. In scenarios requiring improved robustness, the coupled behavior verification module further requires that the key point trajectory recognition unit and the optical flow recognition unit output the same candidate sub-action consistently, and that the gaze trajectory recognition unit only recognizes the candidate sub-action as valid when it provides a placement result for the anchor point carrier corresponding to that candidate sub-action. The upper limit of the time offset can be from 0.05s to 0.50s, preferably from 0.10s to 0.30s, and in this embodiment, it is set to 0.25s. This significantly reduces misjudgments caused by single-path recognition errors.

[0044] See Figure 9 As shown, The keypoint trajectory recognition unit, gaze trajectory recognition unit, and optical flow recognition unit each output a set of candidate behavior sub-actions for consecutive image frames within the same time window. For each candidate behavior sub-action, the action type, start and end times, corresponding anchor point carrier, and action intensity score can be recorded. The consistency judgment rules in the coupled behavior verification module can include at least four items: First, at least two of the three recognition paths output the same action type; second, the difference in duration estimates for this action type between at least two paths does not exceed a preset duration deviation threshold; third, at least two paths make consistent judgments about the anchor point carrier corresponding to this action type; fourth, the time offset between the action start time output by the keypoint trajectory recognition unit and the dwell start time output by the gaze trajectory recognition unit does not exceed a preset time offset upper limit. When the above conditions are met, the corresponding candidate behavior sub-action is retained as a valid behavior sub-action; when the above conditions are not met, the candidate behavior sub-action is marked as a low-confidence candidate behavior sub-action and not included in subsequent graph matching verification. Figure 9 It is evident that the multi-channel consistency determination is not simply a matter of "taking the majority," but rather a joint determination based on action type, consistency duration, corresponding anchor point carrier, and time offset, thereby providing a clear basis for the generation of effective sub-actions.

[0045] In one specific implementation, the multi-channel consistency score can be calculated using the following formula: in, Indicates the multi-channel consistency score. Indicates the consistency score of action types. Indicates the consistency score for duration. This indicates the consistency score of the corresponding anchor point carrier. This indicates the intensity score of the movement. If... If the value is greater than or equal to a preset consistency threshold, the corresponding candidate behavior sub-action is determined as a valid behavior sub-action. For example, when , , , Then: When the preset consistency threshold is 0.80, because Therefore, the candidate action can be retained as a valid action.

[0046] To accommodate differences in camera position and display plane between different student terminals, the course presentation module can also display pose normalization markers in the boundary areas of the course interface. The behavior feature recognition module performs camera pose normalization processing on the motion trajectory of human key points based on the pose normalization marker positions and face region positions in consecutive image frames, and then provides the normalized human key point motion trajectory to the coupled behavior verification module. This maintains consistency in the judgment of relative direction relationships, relative distance interval relationships, and relative dwelling area relationships under different terminal device conditions.

[0047] Combination Figure 12 The pose normalization markers can be set at the four corners of the course interface to determine the display reference plane. The behavior feature recognition module first recovers the display reference plane based on the positions of the four pose normalization markers in consecutive image frames. Then, it estimates the pitch, yaw, and roll deviations of the camera relative to the display reference plane by combining the face region position. Finally, it performs coordinate transformation on the motion trajectory of the human key points, mapping it to a unified normalized screen reference coordinate system. In an optional implementation, the normalization process can be expressed as: in, This represents the original human body key point coordinate vector. This represents the camera pose normalization transformation matrix determined based on the pose normalization marker positions and the face region positions. This represents the normalized coordinate vector of human keypoints. After this processing, the original human keypoint motion trajectories from different terminal devices can be transformed to a unified reference plane, thereby reducing geometric errors caused by different camera installation angles and terminal placement methods. Figure 12 It is evident that the posture normalization process is not an isolated step, but rather forms a complete closed loop with the posture normalization markers, behavior feature recognition module, and coupled behavior verification module in the boundary area of ​​the course interface.

[0048] The node continuity verification module is used to verify whether the behavioral transitions between adjacent training nodes are genuine and continuous. Combined with... Figure 3 The node continuity verification module extracts the human body orientation, head posture, and preset body part spatial intervals from the last consecutive frames of the previous training node to generate a termination behavior state signature; then, it extracts the human body orientation, head posture, and the same preset body part spatial intervals from the first consecutive frames of the current training node to generate an initial behavior state signature. When the difference between the termination behavior state signature and the initial behavior state signature does not exceed a first threshold, and the node switching time interval does not exceed a second threshold, a passed node continuity verification result is output; otherwise, a failed node continuity verification result is output. The first threshold can be 0.10 to 0.30, preferably 0.15 to 0.22, and in this embodiment, it is 0.18; the second threshold can be 0.5s to 2.0s, preferably 0.8s to 1.5s, and in this embodiment, it is 1.2s. In this way, abnormal situations such as leaving the seat, short-term takeover by a substitute operator, camera interruption, or repositioning during node switching intervals can be effectively identified.

[0049] Figure 8 This illustrates the temporal correspondence between the visual anchor point event sequence, prompting voice segments, node response windows, human keypoint motion trajectory sampling intervals, and the sequence of student gaze placement areas. Figure 8 As can be seen, the first, second, and third anchor carriers in the visual anchor event sequence appear sequentially, corresponding to the prompt voice segments A, B, and C respectively; the node response window opens slightly earlier than the first anchor carrier and closes later than the last anchor carrier; the human keypoint motion trajectory sampling intervals A, B, and C respectively cover the sampling time windows of the corresponding behavior sub-action nodes; the dwell intervals A, B, and C in the student's gaze dwell area sequence maintain a correspondence with the human keypoint motion trajectory sampling intervals within a preset time offset upper limit. Figure 8The structure shown can more intuitively illustrate how the effective response range, the upper limit of time offset, the maximum transfer duration, and the start and end boundaries of the node response window work together.

[0050] The process control module simultaneously receives node coupling verification results and node continuity verification results. When both pass, the next training node is opened; when either result fails, a recoding challenge is triggered for the current training node. The node anchor encoding module updates the current challenge seed based on the node evidence summary of the previous training node, the node identifier of the current training node, and the current terminal session identifier. It then regenerates a recoded visual anchor event sequence different from the original visual anchor event sequence based on the updated current challenge seed. If the node coupling verification result fails, the process control module can also insert a correction node corresponding to the sub-action that caused the failure before the current training node, allowing the trainee to perform correction training before returning to the current training node. This improves the retargeting of retry opportunities and avoids invalid repetition.

[0051] In a further embodiment, the behavior feature recognition module also includes a failure sample suppression unit. The failure sample suppression unit retrieves nearest-neighbor failure trajectory clusters from the failure sample library according to the node identifier of the current training node and the current challenge seed, and performs suppression output on candidate behavior sub-actions matching the nearest-neighbor failure trajectory clusters, thereby distinguishing the behavior feature recognition result corresponding to the re-encoded challenge from the failure behavior patterns corresponding to the nearest-neighbor failure trajectory clusters. To make this part more feasible, this embodiment further explains the formation method, update mechanism, and nearest-neighbor determination method of the failure sample library. Specifically, in the initial deployment phase of the system, an initial failure sample library can be established using historical training samples, trial operation samples, or manually constructed failure action sample sets; in the system operation phase, the process control module can archive failure samples whose node coupling verification results fail and are confirmed by the background review into the failure sample library in the combination of "node identifier of the current training node—current challenge seed—failure behavior pattern label"; when the number of similar failure samples reaches a preset number, such as 20, 30, or 50, trajectory clustering can be performed on these failure samples to generate or update the corresponding nearest-neighbor failure trajectory clusters. The failure behavior pattern labels can include "brief stay and then leave", "hand passes by but does not complete the click", "head is facing the correct direction but the line of sight does not enter the effective stay area", "incorrect action sequence", etc.

[0052] Combination Figure 10The complete process of failure sample suppression includes: first, the failure sample database retrieves nearest-neighbor failure trajectory clusters according to the node identifier of the current training node and the current challenge seed; second, the current candidate action sub-action is compared with the retrieved nearest-neighbor failure trajectory clusters in terms of comprehensive proximity; third, the failure sample suppression unit performs suppression output or retains output on the candidate action sub-action based on the comparison results; finally, the real failure samples confirmed by the background are archived, and failure sample archiving and clustering update are triggered after the number of samples reaches a threshold. Figure 10 The structure shown makes it clearer that failure sample suppression is not simply "deleting failure results," but rather forming an implementable data closed-loop path through retrieval, nearest neighbor determination, suppression decision, background review, and clustering update.

[0053] In one specific implementation, the cluster of failed nearest neighbor trajectories can be determined by a comprehensive nearest neighbor score. The comprehensive nearest neighbor score can be determined based on the trajectory shape distance, time offset distance, and overlap of the dwell areas, and can be calculated, for example, by the following formula: in, Represents the comprehensive nearest neighbor degree. Represents the distance of the normalized trajectory shape. Indicates the normalized time offset distance. Indicates the degree of overlap in the dwelling areas. When When the value is greater than or equal to a preset nearest neighbor threshold, the candidate sample and the target failed trajectory cluster can be determined to be neighbors. For example, when the nearest neighbor threshold is set to 0.75, and a candidate sample's... , , Then: because Therefore, it can be determined that the candidate sample and the failed trajectory cluster are neighbors. By adding this path, the failed sample suppression unit is no longer a simple functional conceptual module, but an implementable module with data sources, archiving logic, neighbor determination methods, and update mechanisms.

[0054] The node evidence storage module is used to form a node-level process traceability chain. Combined with... Figure 4The node evidence storage module first divides the continuous image frames and visual anchor event sequences into multiple corresponding segments according to time sequence, and generates a segment summary for each corresponding segment. Then, it generates the node root summary for the current training node based on the multiple segment summaries. Simultaneously, the node evidence storage module generates a first source summary and a second source summary for the visual anchor event sequence and the trainee behavior sub-action sequence, respectively. Then, it cascades the first source summary, the second source summary, the node root summary, and the node evidence summary from the previous training node to generate a chain-like evidence summary for the current training node, which serves as the node evidence summary for the current training node. To further enhance the support of this part of the specification, this embodiment supplements the description of the storage location and calling scenarios of the chain-like evidence summary. Specifically, the chain-like evidence summary can be stored as the node evidence summary for the current training node in the node evidence table or node evidence index library within the node evidence storage module. It can be called by the process control module before the next training node is opened, and can also be called by the background review module, audit module, or administrator terminal during node anomaly review, training dispute review, and compliance audit.

[0055] Combination Figure 11 The chain evidence digest generation and verification process can be further explained as follows: The first source digest, the second source digest, the node root digest, and the node evidence digest of the previous training node are input into the cascading processing unit to generate the chain evidence digest of the current training node. After being stored as the node evidence digest of the current training node, the chain evidence digest can be called by the process control module, the background review module, and the audit module. When the process control module calls the chain evidence digest, it can be used to verify the completeness of the node-level processing chain of the current training node before the next training node is opened. When the background review module calls the chain evidence digest, it can be used to perform link consistency verification on disputed nodes. When the audit module calls the chain evidence digest, it can be used to perform cross-node link auditing.

[0056] In the background review scenario, when the node coupling verification result of a training node fails twice consecutively, or the node continuity verification result triggers anomalies consecutively, the process control module can mark the training node as a node to be reviewed. Administrators or reviewers can retrieve the visual anchor point event sequence log, trainee behavior sub-action sequence log, fragment summary, node root summary, and chain evidence summary of the current training node through the background review module, and perform review in conjunction with the fragment index of consecutive image frames. If the reviewer confirms that the failed sample is a genuine failed sample, it can be archived to the failed sample library for subsequent updates of the nearest neighbor failed trajectory cluster; if the reviewer confirms that it is an occasional misidentification, it can be marked as an invalid failed sample and not included in the failed sample library. Through this closed-loop review process, the sample quality and failure sample suppression effect during system operation can be further improved.

[0057] To enable those skilled in the art to more clearly understand the key determination process in this invention, several specific calculation examples are given below.

[0058] In an example concerning relative distance intervals, the center coordinates of the second anchor point carrier in the operation prompt icon area are set to... The coordinates of the student's right fingertip in the normalized screen reference coordinate system are set to The screen projection distance from the right hand to the center of the second anchor point carrier can be calculated using the following formula: in, This represents the screen projection distance from the right hand to the center of the second anchor point carrier. Indicates the coordinates of the right fingertip. This indicates the coordinates of the center of the second anchor point carrier.

[0059] In a specific sample of this embodiment, the coordinates of the right fingertip are: The center coordinates of the second anchor point carrier are Then we have: When the allowed range for the relative distance interval is set to 20 to 80 pixels, since 39.85 falls within this range, the right hand satisfies the relative distance interval relationship in this frame. If the interval relationship is satisfied in consecutive sampled frames and the hold duration threshold is reached, then the behavior sub-action node B is successfully matched.

[0060] In an example regarding the relationship between relative dwell areas and the dwell time threshold, the effective dwell window for the subtitle area is defined as a rectangular area 10 pixels inward from the subtitle boundary. Assuming the student's gaze falls within this effective dwell window for 25 consecutive frames, and the sampling frame rate is 25 frames per second, the continuous dwell time is calculated using the following formula: in, Indicates the duration of continuous stay. This represents the number of frames that continuously meet the dwell condition. This represents the sampling frame rate. Substituting the example values, we get: When the duration threshold for the behavior sub-action node C is set to 0.8s, due to Therefore, it can be determined that the trainee's line of sight relative to the third anchor point carrier satisfies the relative dwelling area relationship and reaches the holding time threshold.

[0061] In an example concerning relative directional relationships, the head's principal orientation vector can be denoted as... The target vector pointing from the center of the head to the center of the first anchor point carrier is denoted as... The directional deviation angle of the head relative to the first anchor point carrier can be calculated using the following formula: in, Indicates the directional deviation angle. Represents the vector dot product. and These represent the vector magnitudes. When the direction deviation angle threshold is set to 15°, if the calculated... This indicates that the head orientation satisfies the relative directional relationship.

[0062] In an example of node continuity verification, the combined difference between the termination behavior state signature and the initial behavior state signature can be calculated using the following formula: in, Indicates the overall degree of difference. This represents the absolute value of the head posture deviation angle. This represents the absolute value of the angle of deviation of the human body's orientation. This indicates the overlap rate of the preset body part spatial intervals.

[0063] In a specific sample, the head pose of the last consecutive frames of the previous training node is: The head pose of the initial consecutive frames of the current training node is ,but The orientation of the human body in the last consecutive frames of the previous training node is... The initial consecutive frames of the current training node show the human body orientation as follows: ,but If the presupposes an overlap rate of 0.92 between body parts, then: When the first threshold is set to 0.18, the node switching time interval is 0.8s, and the second threshold is set to 1.2s, because... ,and Therefore, the output of this sample is a successful node continuity check result. If the node switching time interval increases to 1.4s, or the overall difference increases to 0.21, then a failed node continuity check result will be output. Figure 6 The curves showing the overall difference between normal and abnormal handover samples as a function of node handover time intervals are presented. From... Figure 6It can be intuitively seen that when the first threshold is 0.18, normal switching samples are still below the threshold within a short node switching time interval, while abnormal switching samples cross the first threshold earlier, indicating that the node continuity verification method can effectively distinguish between normal node switching and abnormal node switching.

[0064] In an example of node notarization, the node notarization module uses the SHA-256 hash algorithm. To avoid interpreting the hash algorithm as the only implementation, this embodiment also states that the hash algorithm can also be SHA-3, SM3, or other one-way hash algorithms; this embodiment only uses the SHA-256 hash algorithm as an example. Assume the current training node is divided into three corresponding segments, and the segment hashes of these three segments are denoted as Segment Hash 1, Segment Hash 2, and Segment Hash 3, respectively. Their specific hash values ​​can be: Segment Summary 1: 768cb05785a766ca9158ef534b75b3d2c20f22c6554f0403fc6f8ed852c7f6de; Segment Summary 2: e6984569c01835447ca0ae51590a4a00b59f48968346a6c8d34a0a88195c26c8; Segment Summary 3: 0f82032855aed7484d8a9094b6f4243f9e7addcf6c9f670c152d74c92fb9d812 The root digest of a node can be obtained by concatenating the three fragment digests in sequence and then inputting them into the SHA-256 digest algorithm. Correspondingly, the root digest of a node can be represented as: dd2d8a9f8fbf500262e09045b81cbe2116a523bcca71e0c742df970a330d7065 If the first source summary of the visual anchor event sequence is: 93b43fc2010b47782b5810bfffbdcaf3cbe08074c81e54a44fd849ef69971713 The second source summary of the trainee's behavioral sub-action sequence is: dffc6c4fb5e049e4da355e7fbe1069f9ce039a1bb0b17a0ae602cf21b411ab22 The node evidence summary for the previous training node is as follows: d7262d8789e93cade616a71b8baf4679a0ecfc97a74d21246da1eacd1dc8bc58 The first source digest, the second source digest, the root node digest, and the node evidence digest of the previous training node are concatenated in sequence and then input into the SHA-256 digest algorithm to obtain the chained evidence digest of the current training node. For example, in this embodiment, the following can be obtained: 31b517f4d9f3f3012e2c98298c6dcd83a18ae989e2958fbf957dbb347d9b130e The chain-like evidence digest is stored as the node evidence digest of the current training node in the node evidence table of the node evidence storage module, and can be called by the process control module, the background review module, and the audit module. Through the above method, it can be seen that the node evidence storage module does not simply save the original image, but forms a summary chain that supports node-level traceability through a three-level structure of fragment-level digests, node root digests, and chain-like evidence digests.

[0065] To more intuitively illustrate the technical effects of the present invention, experimental examples are provided below. Comparative Example 1 employs a comparative system constructed based on existing online training status recognition methods. Specifically, during training, facial expressions, eye movements, fixation points, and head postures are collected. Subsequent training content is adjusted according to the trainee's emotional state, cognitive state, and level of comprehension. However, no visual anchor event sequence is set, no node behavior script templates are set, no node continuity verification is performed, and no node-level chain-based evidence storage is implemented. The embodiments of the present invention adopt the above-described complete technical solution.

[0066] The experiment recruited 60 participants, each completing 20 training nodes, resulting in 1200 node samples, including 960 normal node samples and 240 abnormal node samples. Abnormal node samples included premature actions, delayed actions, actions not corresponding to the current interface, node switching and leaving the seat, brief takeover by a substitute operator, and repeated retries along a recurring failure behavior pattern. In the experiment, the embodiments of the present invention and Comparative Example 1 used the same course content, the same group of participants, and the same data acquisition equipment, with a uniform sampling frame rate of 25 frames per second.

[0067] Both normal and abnormal node samples are jointly reviewed by two training managers and one technical annotator based on node video recordings, visual anchor point event sequence logs, node coupling verification results, and node continuity verification results. When the three parties agree on the sample labels, the final labels are determined. If discrepancies arise, the review is conducted by re-watching the video recordings and verifying the node evidence summary links until a consensus is reached. Abnormal node samples are manually annotated according to six categories: premature actions, delayed actions, actions not corresponding to the current interface, node switching / absence, short-term takeover by a substitute operator, and repeated failure behavior patterns.

[0068] To make the calculation of the experimental results clearer, this embodiment further supplements the explanation of the statistical caliber of the main indicators. The node-level coupling verification accuracy can be calculated using the following formula: in, This indicates the accuracy of node-level coupling verification. This represents the number of node samples that were correctly identified. This represents the total number of samples across all nodes.

[0069] The abnormal node detection rate can be calculated using the following formula: in, This indicates the abnormal node detection rate. This represents the number of correctly identified anomalous node samples. This represents the total number of abnormal node samples.

[0070] The success rate of correction after retry can be calculated using the following formula: in, This indicates the success rate of error correction after retrying. This represents the number of samples that successfully completed the current training node after passing the correction node. This represents the total number of samples that triggered the recoding challenge or correction node.

[0071] The traceability completion rate of a node-level process can be calculated using the following formula: in, This indicates the traceable completion rate of a node-level process. This represents the number of node samples that can be fully recovered from the node process and its preceding and following links.

[0072] Table 1 for setting test conditions Performance Comparison Results Table 2 Figure 5 This is a bar chart of the main performance indicators in Table 2. From... Figure 5As can be seen, the embodiments of the present invention are significantly superior to Comparative Example 1 in terms of node-level coupling verification accuracy, abnormal node detection rate, retry correction success rate, and node-level process traceability completion rate. This is because the embodiments of the present invention do not make an overall judgment based on a general learning state, but rather perform node-level coupling verification based on the visual anchor point event sequence corresponding to the current training node, the node behavior script template, the human body key point movement trajectory, the trainee's gaze dwell area sequence, and the node continuity verification results, thereby improving the pertinence and accuracy of behavior response judgment. Figure 6 The combined difference curves between normal handover samples and abnormal handover samples are shown. Figure 6 It is evident that when the node switching interval is short, the overall difference of normal switching samples is below the first threshold; while abnormal switching samples reach or exceed the first threshold within a short time interval. This result demonstrates that the node continuity verification mechanism constructed through termination behavior state signature and initial behavior state signature can provide effective evidence for identifying absence from the operator, short-term takeover by a substitute operator, and abnormal switching. Figure 7 The diagram shows the change in the re-encoding challenge recurrence failure rate with repeated challenge rounds before and after failure sample suppression. Figure 7 As can be seen, in the first round of the challenge, the repetition failure rate of the two groups of samples was the same; from the second round onwards, the repetition failure rate after adopting failure sample suppression decreased significantly and remained at a low level in subsequent rounds. This indicates that the failure sample suppression unit can effectively suppress the phenomenon of trainees repeatedly retrying along the same error path, thereby improving the correction efficiency of the recoding challenge.

[0073] Based on the above implementation methods, calculation examples, tabular data, and Figures 5 to 12 As can be seen, this invention generates a visual anchor point event sequence in the current training node, couples and matches the human body key point motion trajectory, the trainee's gaze lingering area sequence, and the trainee's behavioral sub-action sequence with the node behavior script template, and then combines node continuity verification, failure sample suppression, correction node insertion, and chain node evidence storage. This effectively improves the accuracy of node-level coupling verification, enhances the ability to detect abnormal nodes, increases the success rate of retry correction, and significantly enhances the authenticity verification capability and node-level traceability capability of the training process. Thus, it achieves more stable technical results in the authenticity verification and process control of online training.

[0074] Example 2: To further illustrate that this invention is not limited to the single scenario of "cybersecurity compliance training courses," a second embodiment is provided below. This embodiment provides a network training software platform based on behavioral feature recognition applied to vocational skills training scenarios. Its basic structure is the same as that of Embodiment 1, still including a training node management module, a course presentation module, a node anchor point encoding module, an image acquisition module, a behavioral feature recognition module, a coupled behavior verification module, a node continuity verification module, a process control module, and a node evidence storage module. The difference lies in the training node content, the configuration of behavioral sub-action nodes, and the selection method of the corresponding anchor point carrier.

[0075] In this embodiment, the "Equipment Inspection Specification Training Course" is used as an example course. The course can be divided into multiple training nodes such as "Tool Confirmation Node," "Checkpoint Identification Node," "Abnormal Item Marking Node," and "Result Confirmation Node." Taking the "Abnormal Item Marking Node" as an example, the course interface of this training node includes a process prompt area, a tool selection area, and a confirmation button area. The node anchor point encoding module generates a visual anchor point event sequence corresponding to the current training node in the process prompt area, tool selection area, and confirmation button area. In the node behavior script template, the specified body part corresponding to behavior sub-action node A is both eyes, and the corresponding anchor point carrier is the first anchor point carrier in the process prompt area, with the target relative relationship type being a relative dwelling area relationship; the specified body part corresponding to behavior sub-action node B is the right hand, and the corresponding anchor point carrier is the second anchor point carrier in the tool selection area, with the target relative relationship type being a relative distance interval relationship; the specified body part corresponding to behavior sub-action node C is the right hand, and the corresponding anchor point carrier is the third anchor point carrier in the confirmation button area, with the target relative relationship type being a combination of relative direction relationship and relative distance interval relationship. This configuration demonstrates that the present invention is applicable not only to content reading training nodes, but also to skills-based training nodes that include step confirmation, tool selection, and result confirmation.

[0076] In Example 2, the hold duration threshold can be set to 0.2s to 1.2s, preferably 0.3s to 0.8s, and can be adjusted according to the complexity of the actions in skills training; the maximum transfer duration can be set to 1.0s to 4.0s, preferably 1.5s to 2.5s; and the upper limit of time offset can be set to 0.08s to 0.40s, preferably 0.12s to 0.28s. Since the interaction of skills training nodes is often more complex than that of reading nodes, the node continuity verification in this example can further incorporate tool holding state information in addition to the termination behavior state signature and the initial behavior state signature, thereby improving the ability to identify short-term takeover and step-jumping operations by substitute operators. Therefore, the technical solution of this invention can be applied not only to knowledge-transferring online training but also to vocational skills demonstration and operational standard training, demonstrating good scenario adaptability.

[0077] To further avoid the misconception that this invention is only applicable to training nodes with "hand-click interaction," a modified implementation with no graphical clicks and pure gaze response is described. In this modified implementation, the visual anchor event sequence of the current training node is generated only in the knowledge point title area, subtitle area, and prompt box area. All behavioral sub-action nodes are represented by the head orientation and the sequence of the student's gaze placement area, without requiring right-hand interaction. In this case, the specified body parts in the node behavior script template can include only the head and gaze, with the corresponding anchor carriers being the first anchor carrier in the knowledge point title area, the second anchor carrier in the subtitle area, and the third anchor carrier in the prompt box area. The coupling behavior verification module can perform graph matching verification through relative direction relationships and relative placement area relationships. Even in a pure gaze response scenario, the flow control module can still control the current training node to proceed, retry, or trigger a recoding challenge based on the node coupling verification results and node continuity verification results, thereby maintaining the consistency of the main flow loop. Therefore, it can be seen that the "student behavior response" in this invention is not limited to clicks, drags, or confirmation button triggers; it can also be an observational, confirmation, or reading response based on the head orientation and the sequence of the student's gaze placement area.

[0078] In summary, Embodiment 1, Embodiment 2, and the above-described variations collectively demonstrate that the network training software platform based on behavioral feature recognition provided by this invention can achieve node-level coupling verification between the visual anchor event sequence and the student's behavioral response during network training through technical means such as the visual anchor event sequence corresponding to the current training node, node behavior script template, human body key point movement trajectory, student gaze lingering area sequence, node continuity verification, failure sample suppression, recoding challenge, correction nodes, and node-level chain storage. This further improves the training authenticity verification capability, abnormal behavior detection capability, retry correction capability, and node-level traceability capability.

[0079] It should be noted that the embodiments of the present invention have better implementability and are not intended to limit the present invention in any way. Any person skilled in the art may use the above-disclosed technical content to change or modify it into equivalent effective embodiments. However, any modifications or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A network training software platform based on behavioral feature recognition, characterized in that, include: The training node management module is used to store multiple training nodes and the node behavior script templates corresponding to each training node. The course presentation module is used to present the course interface of the current training node to the student's terminal; The node anchor point encoding module is used to generate the current challenge seed based on the node identifier and the current terminal session identifier of the current training node, and to generate the visual anchor point event sequence corresponding to the current training node in the course interface based on the current training node and the current challenge seed. The image acquisition module is used to acquire continuous image frames of trainees within the node response window corresponding to the visual anchor point event sequence. The behavior feature recognition module is used to perform behavior feature recognition on the continuous image frames to extract the motion trajectory of human key points and output the trainee's behavior sub-action sequence; The coupling behavior verification module is used to perform coupling matching by comparing the node behavior script template with the spatiotemporal relative relationship between the human body key point motion trajectory and the visual anchor point event sequence, and output the node coupling verification result of the current training node. The node continuity verification module is used to output the node continuity verification result of the current training node based on the termination behavior state of the previous training node, the initial behavior state of the current training node, and the node switching time interval. The process control module is used to open the next training node when both the node coupling verification result and the node continuity verification result are passed, and to trigger the recoding challenge of the current training node when either verification result fails. The node evidence storage module is used to generate a node evidence summary for the current training node based on the current challenge seed, the visual anchor event sequence, the trainee behavior sub-action sequence, the node coupling verification result, the node continuity verification result, and the node evidence summary of the previous training node.

2. The network training software platform based on behavioral feature recognition according to claim 1, characterized in that, The node anchor encoding module is used to determine at least two anchor carriers from the knowledge point title area, progress indicator area, subtitle area and operation prompt icon area of ​​the course interface, and configure the display position, display order, display duration and display level of each anchor carrier based on the current challenge seed, so as to generate the visual anchor event sequence.

3. The network training software platform based on behavioral feature recognition according to claim 2, characterized in that, The node behavior script template includes a directed acyclic graph consisting of multiple behavior sub-action nodes and multiple allowed transfer edges. Each behavior sub-action node stores a specified body part, a corresponding anchor point carrier, a target relative relationship type, and a hold duration threshold. Each allowed transfer edge stores a maximum transfer duration threshold. The coupled behavior verification module is used to perform graph matching verification on the trainee behavior sub-action sequence based on the directed acyclic graph.

4. The network training software platform based on behavioral feature recognition according to claim 3, characterized in that, The target relative relationship type includes at least the relative direction relationship, relative distance interval relationship, and relative dwelling area relationship of the specified body part relative to the corresponding anchor point carrier; the coupling behavior verification module determines that the corresponding behavior sub-action is successfully matched only when the specified body part continuously satisfies the relative direction relationship, the relative distance interval relationship, and the relative dwelling area relationship to the duration threshold.

5. The network training software platform based on behavioral feature recognition according to claim 3, characterized in that, The training node management module is also used to store a one-to-one corresponding prompt voice segment for each of the behavior sub-action nodes; the course presentation module is used to synchronously play the prompt voice segment when presenting the course interface of the current training node; the coupled behavior verification module performs graph matching verification on the corresponding behavior sub-action node only within the effective reaction interval formed by the start and end time of the playback of the corresponding prompt voice segment.

6. The network training software platform based on behavioral feature recognition according to claim 3, characterized in that, The behavior feature recognition module includes a key point trajectory recognition unit and a gaze trajectory recognition unit; The key point trajectory recognition unit is used to extract the motion trajectory of human key points of the specified body part from the continuous image frames; The gaze trajectory recognition unit is used to extract the sequence of the student's gaze resting areas from the continuous image frames; The behavior feature recognition module outputs the corresponding sub-action only when the movement trajectory of the human body key points and the sequence of gaze resting areas both correspond to the same anchor point carrier, and the time offset does not exceed a preset upper limit.

7. The network training software platform based on behavioral feature recognition according to claim 6, characterized in that, The image acquisition module opens the node response window when the first anchor carrier in the visual anchor event sequence switches from a hidden state to a displayed state, and caches a preset number of pre-prepared image frames before opening. When the last anchor carrier in the visible anchor event sequence switches from the display state to the hidden state, the node response window is closed, and subsequent image frames for a preset number of frames after closing are cached.

8. The network training software platform based on behavioral feature recognition according to claim 6, characterized in that, The behavior feature recognition module further includes an optical flow recognition unit; the key point trajectory recognition unit, the gaze trajectory recognition unit, and the optical flow recognition unit each output candidate behavior sub-actions; the coupled behavior verification module adopts the candidate behavior sub-action as a valid behavior sub-action only when the outputs of the key point trajectory recognition unit and the optical flow recognition unit for the same candidate behavior sub-action are consistent, and the gaze trajectory recognition unit gives a dwell result for the anchor point carrier corresponding to the same candidate behavior sub-action.

9. The network training software platform based on behavioral feature recognition according to claim 1, characterized in that, The node anchor encoding module is also used to update the current challenge seed based on the node evidence digest of the previous training node, the node identifier of the current training node, and the current terminal session identifier; When any verification result fails, the process control module calls the node anchor encoding module to regenerate a re-encoded visual anchor event sequence that is different from the original visual anchor event sequence based on the updated current challenge seed.

10. The network training software platform based on behavioral feature recognition according to claim 9, characterized in that, After the node coupling verification result fails and the recoding challenge is triggered, the process control module inserts a correction node before the current training node, corresponding to the behavior sub-action that caused the node coupling verification result to fail, and only restarts the current training node after the correction node passes.

Images