Teaching quality evaluation method and system based on classroom student behavior analysis

Abstract

The invention relates to the technical field of teaching analysis, in particular to a teaching quality evaluation method and system based on classroom student behavior analysis, and the method comprises the steps: judging whether classroom video data has multi-view shooting features, determining a key behavior frame sequence of the classroom video data based on a judgment result, carrying out the behavior segment extraction of the key behavior frame sequence, and obtaining a behavior segment extraction result; obtaining a behavior fragment set; identifying student behaviors in the behavior fragment set to obtain behavior type labels, and classifying the behavior type labels to obtain classified behavior labels; and performing time point sampling on the behavior fragments corresponding to the classified behavior tags to obtain time sampling points. By analyzing the classroom behaviors of the students and extracting the key behavior frame sequence, the real performance of the students in the classroom can be more objectively and accurately captured, and the key behavior mode influencing the teaching quality can be accurately identified through detailed time chain construction, behavior fragment extraction and behavior frequency analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of teaching analysis, in particular to a teaching quality evaluation method and system based on classroom student behavior analysis. BACKGROUND

[0002] With the rapid development of information technology, the education field gradually moves towards informatization and intelligentization, especially the progress of big data, artificial intelligence and video analysis technology, which provides a new perspective and tool for teaching quality evaluation. Under this background, the traditional teaching evaluation method has been unable to meet the demand of precision and individualization of modern education.

[0003] At present, the traditional teaching evaluation often relies on the subjective judgment of teachers or evaluators, which may be affected by personal bias, emotion or evaluation standard, resulting in low accuracy and objectivity of the evaluation results. Moreover, it is usually based on classroom observation, student self-reporting or questionnaire survey, etc. These means are limited in observing and capturing student behavior, and it is difficult to fully reflect the dynamic behavior of students in the classroom.

[0004] In addition, the traditional evaluation method is static and cannot accurately track the trend of student behavior changes. There is no dynamic analysis based on time chain, and it is difficult to reveal the evolution process of student behavior and its specific impact on teaching effectiveness. Moreover, there is a lack of efficient data analysis means, and the data processing of classroom behavior is usually rough, which cannot start from the details to dig out the deep information behind the student behavior, for example, lacking accurate analysis of behavior frequency, time point correlation and other data. SUMMARY

[0005] To achieve the above purpose, the present application provides the following technical scheme: a teaching quality evaluation method based on classroom student behavior analysis, comprising: determining whether the classroom video data has a multi-view shooting feature, determining the key behavior frame sequence of the classroom video data based on the determination result, extracting behavior segments from the key behavior frame sequence to obtain a behavior segment set; identifying the student behavior in the behavior segment set to obtain a behavior type label, classifying the behavior type label to obtain a classified behavior label; time point sampling of the behavior segment corresponding to the classified behavior label to obtain each time sampling point, correlation analysis of each time sampling point to obtain a correlation analysis result, and time chain construction of adjacent time sampling points in the correlation analysis result that meet a preset correlation degree threshold to obtain a corresponding time chain; feature vector normalization and similarity operation on the behavior feature vector corresponding to the time sampling point in the time chain, filtering of the time sampling point according to the similarity value to obtain a filtered time sampling point; The target time sampling points are obtained by expanding the screened time sampling points corresponding to the local behavior type satisfying the preset behavior frequency change condition in the classified behavior label, and the target behavior mode is obtained by performing mode mining on the target time sampling points.

[0006] Preferably, whether the classroom video data has a multi-view shooting feature is judged, and a key behavior frame sequence of the classroom video data is determined based on the judgment result, including: whether the classroom video data has a multi-view shooting feature is judged; In response to the classroom video data having a multi-view shooting feature, the multi-view classroom video data is subjected to view fusion processing, and the key behavior frame sequence is extracted from the fused video data; In response to the classroom video data not having a multi-view shooting feature, the key behavior frame sequence is directly extracted from the classroom video data.

[0007] Preferably, the key behavior frame sequence is subjected to behavior segment extraction to obtain a behavior segment set, including: whether the key behavior frame sequence satisfies a preset inter-frame change condition is judged; In response to the key behavior frame sequence satisfying the preset inter-frame change condition, a local dynamic inter-frame difference analysis algorithm is used to extract behavior segments from the key behavior frame sequence to obtain a behavior segment set; In response to the key behavior frame sequence not satisfying the preset inter-frame change condition, a global dynamic inter-frame difference analysis algorithm is used to extract behavior segments from the key behavior frame sequence to obtain a behavior segment set.

[0008] Preferably, the key behavior frame sequence is subjected to behavior segment extraction using a local dynamic inter-frame difference analysis algorithm to obtain a behavior segment set, including: a local region dynamic change value between adjacent frames in the key behavior frame sequence is calculated; whether the local region dynamic change value satisfies a preset inter-frame change condition is judged; If yes, the adjacent frames before and after satisfying the condition are taken as boundary frames of the behavior segment; the key behavior frame sequence is divided based on the boundary frames to obtain a plurality of behavior segments; the plurality of behavior segments are grouped into a behavior segment set.

[0009] Preferably, the key behavior frame sequence is subjected to behavior segment extraction using a global dynamic inter-frame difference analysis algorithm to obtain a behavior segment set, including: a global dynamic difference value between all adjacent frames in the key behavior frame sequence is calculated; whether the global dynamic difference value does not satisfy a preset inter-frame change condition is judged; If not, the key behavior frame sequence is taken as a potential behavior segment; Performing internal dynamic change analysis on the potential behavior segment to determine a sub-interval with internal dynamic change less than a preset frame difference threshold value; Taking the video frame sequence corresponding to the sub-interval as a final behavior segment; Grouping a plurality of behavior segments to form a behavior segment set.

[0010] Preferably, the behavior type labels are classified to obtain classified behavior labels, including: Classifying the behavior type labels by a preset behavior label classification rule to obtain classified behavior labels; The preset behavior label classification rule is to filter labels in the behavior type labels that do not meet a preset positive behavior condition, and to filter labels in the behavior type labels that do not meet a preset negative behavior condition.

[0011] Preferably, a time chain is constructed by using adjacent time sampling points in the correlation analysis result that meet a preset correlation threshold value to obtain a corresponding time chain, including: Calculating the correlation analysis result corresponding to the adjacent two time sampling points to obtain a correlation value; If the correlation value is greater than the preset correlation threshold value, the adjacent two time sampling points are retained, and a time chain is constructed by using the retained time sampling points to obtain a corresponding time chain.

[0012] Preferably, the behavior feature vectors corresponding to the time sampling points in the time chain are normalized and similarity operations are performed on the behavior feature vectors, and the time sampling points are screened according to the similarity values to obtain screened time sampling points, including: Based on a first time sampling point and a second time sampling point of the time chain, a first behavior feature vector is obtained, and based on the second time sampling point and a third time sampling point, a second behavior feature vector is obtained; The first behavior feature vector and the second behavior feature vector are normalized and similarity operations are performed to obtain a similarity value; If the similarity value meets a preset similarity condition, the second time sampling point is screened out to obtain screened time sampling points.

[0013] Preferably, the screened time sampling points corresponding to a local behavior type that meets a preset behavior frequency change condition in the classified behavior labels are expanded to obtain target time sampling points, and the target time sampling points are pattern mined to obtain a target behavior pattern, including: statistically obtain a behavior frequency value, and determine whether the behavior frequency value is greater than a preset frequency threshold value; If the behavior frequency value is greater than the preset frequency threshold value, the post-screening time sampling points corresponding to the local behavior type are expanded to obtain target time sampling points, and a sequence pattern mining algorithm is used to mine the target time sampling points to obtain a target behavior pattern.

[0014] A teaching quality evaluation system based on classroom student behavior analysis, which is applicable to the teaching quality evaluation method based on classroom student behavior analysis, comprises: A segment extraction unit is configured to determine whether the classroom video data has a multi-view shooting feature, determine a key behavior frame sequence of the classroom video data based on the determination result, extract behavior segments from the key behavior frame sequence, and obtain a behavior segment set. A label classification unit is configured to identify student behavior in the behavior segment set to obtain a behavior type label, and classify the behavior type label to obtain a post-classification behavior label. An association analysis unit is configured to sample time points from behavior segments corresponding to the post-classification behavior label to obtain each time sampling point, analyze the association of each time sampling point to obtain an association analysis result, and construct a time chain from adjacent time sampling points in the association analysis result that meet a preset association degree threshold to obtain a corresponding time chain. A sampling and screening unit is configured to perform feature vector normalization and similarity calculation on a behavior feature vector corresponding to a time sampling point in the time chain, and screen the time sampling points according to the similarity value to obtain post-screening time sampling points. A pattern mining unit is configured to expand the post-screening time sampling points corresponding to a local behavior type in the post-classification behavior label that meets a preset behavior frequency change condition to obtain target time sampling points, and mine the target time sampling points to obtain a target behavior pattern.

[0015] Compared with the prior art, the present application has the following advantages: (1) By analyzing the classroom behavior of students and extracting a key behavior frame sequence, the present application can more objectively and accurately capture the real performance of students in the classroom. Through detailed time chain construction, behavior segment extraction, and behavior frequency analysis, the present application can accurately identify the key behavior patterns that affect the teaching quality, avoiding the overly subjective or limited evaluation methods in traditional teaching evaluation. Moreover, the present application can process classroom video data shot from multiple angles, fully utilizing information fusion under different angles, thereby improving the comprehensiveness and depth of student behavior analysis. Even if the classroom video is not shot from multiple angles, the system can still rely on a single angle for effective behavior analysis, ensuring effective implementation in different teaching environments. (2) The application can more accurately extract student behavior segments through local dynamic and global dynamic inter-frame difference analysis algorithm, especially in identifying and classifying different types of student behavior, by setting preset behavior label classification rules, the clear distinction between positive and negative behaviors is ensured, which helps teachers to discover the classroom participation status of students in time and make targeted adjustments according to needs, and through the correlation analysis of the time points of the key behavior frames, the time chain change of the student behavior can be accurately captured, which not only can reveal the attention change of the students in the classroom, but also can help teachers to identify the specific influence of certain behavior patterns on the teaching effect, for example, the patterns of efficient learning or inefficient learning can be mined through the high-frequency behavior segments in the time chain, and the teaching strategy can be improved accordingly. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is a step flow diagram of the overall method in an embodiment of the application. Figure 2 It is a system architecture diagram of the overall system in an embodiment of the application.

[0017] In the figure: 1, segment extraction unit; 2, label classification unit; 3, correlation analysis unit; 4, sampling screening unit; 5, pattern mining unit. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0019] Embodiment one, please refer to Figure 1 The application provides a technical solution: a teaching quality evaluation method based on classroom student behavior analysis, comprising: S1, judging whether the classroom video data has multi-view shooting characteristics, determining the key behavior frame sequence of the classroom video data based on the judgment result, performing behavior segment extraction on the key behavior frame sequence to obtain a behavior segment set; S2, identifying the student behavior in the behavior segment set to obtain a behavior type label, classifying the behavior type label to obtain a classified behavior label; S3, time point sampling is performed on the behavior segments corresponding to the classified behavior labels to obtain each time sampling point, correlation analysis is performed on each time sampling point to obtain a correlation analysis result, and a time chain is constructed by performing time chain construction on adjacent time sampling points in the correlation analysis result that meet a preset correlation degree threshold, to obtain a corresponding time chain; S4, feature vector normalization and similarity operation are performed on the behavior feature vectors corresponding to the time sampling points in the time chain, and the time sampling points are filtered according to the similarity values to obtain filtered time sampling points; S5, the filtered time sampling points corresponding to the local behavior types that meet the preset behavior frequency change condition in the classified behavior labels are expanded to obtain target time sampling points, and the target time sampling points are subjected to pattern mining to obtain a target behavior pattern.

[0020] It should be noted that it is necessary to determine whether the classroom video data has the shooting feature of multiple perspectives; this is usually achieved by capturing the classroom scene through cameras at different angles; if there are multiple perspectives, each perspective may capture the behavior characteristics of different students; through this step, it can be determined which video frames are key frames, and these key frames represent the moments of key behaviors of students in the classroom; for example: assuming that there is a classroom video of a math class, multiple camera angles are used in the video to capture multiple perspectives such as teacher explanation, student answering, student interaction, etc.; it is necessary to determine which video frames are key frames, such as the moment when a student raises his hand to answer a question or the moment when a student bends down to write homework, as the basis for subsequent analysis; Once the behavior segments (such as students raising their hands, students speaking, students focusing on listening, etc.) are extracted from the video, it is necessary to perform behavior recognition on these segments; for example, using image recognition technology or other behavior analysis algorithms to classify each behavior segment and label the specific behavior type (such as "raising hands", "speaking", "communicating with classmates"); for example: assuming that there is a segment in which a student raises his hand; in this segment, through analysis, the system will identify it as a "raising hands" behavior label; other segments may be labeled as "student writing homework", "student answering questions", etc.; The classified behavior labels are sampled according to time points to obtain behavior data at each key moment; these time points will help subsequent correlation analysis to find the temporal relationship between behaviors, for example, whether a student will answer a question within a short period of time after raising his hand; for example: assuming that at 10:00, a student raises his hand, and time point sampling will record the exact time (such as 10:03) when the "raising hands" behavior occurs; if the student starts answering the question at 10:04, the system will record this time point as the basis for subsequent analysis; After determining the behavior sampling points, the next step is to normalize the behavior features of these sampling points, so that different behavior data can be compared; for example, for different students' hand-raising behavior, the sampled feature values may be different due to different conditions such as angle and brightness; through normalization processing, these differences can be eliminated; then, the similarity between different time points is calculated to find similar behavior patterns; for example: assuming that student A and student B raise their hands at different times, after normalization processing, the system will calculate the similarity of the two students' hand-raising behavior, and judge the similarity of the two behaviors; if their behavior similarity is high, it may indicate that the two hand-raising behaviors are triggered by similar situations; After normalizing the behavior and calculating the similarity, the system will find behaviors whose frequency changes, for example, the participation frequency of some students in the classroom suddenly increases or decreases; by mining the patterns of these behaviors, it can help to identify changes in students' learning state, such as changes in their concentration or participation; for example: assuming that student A often raises his hand at the beginning of the class, but his hand-raising frequency suddenly decreases in a certain period of time; this change may indicate that student A's attention concentration decreases in this period of time; through pattern mining, the system can identify this change and further analyze the reasons, such as whether the teacher's explanation changes, or the student's emotional fluctuations, etc.

[0021] In an optional embodiment, it is judged whether the classroom video data has a multi-view shooting feature, and the key behavior frame sequence of the classroom video data is determined based on the judgment result, comprising: judging whether the classroom video data has a multi-view shooting feature; In response to the classroom video data having a multi-view shooting feature, performing view fusion processing on the multi-view shot classroom video data, and extracting the key behavior frame sequence from the fused video data; In response to the classroom video data not having a multi-view shooting feature, directly extracting the key behavior frame sequence from the classroom video data.

[0022] It needs to be explained that whether there is a feature of multiple camera angles in the classroom video data is analyzed; for example, multiple cameras may be used in the classroom to capture the behavior of students and teachers from different angles; multi-view video can provide more comprehensive information, for example, one camera can focus on the facial expressions of students, while another camera can capture the interaction between students and teachers; for example: suppose there is a classroom video with two cameras shooting: one camera captures the overall reaction of students from one side of the classroom, including raising hands, talking, participating, etc.; the other camera focuses on the podium, capturing the teacher's actions and the content on the blackboard; the system needs to determine whether such multi-view features exist; if there are multiple perspectives, perspective fusion processing is needed; if there is only one camera or no obvious multi-view features, key frames under a single perspective are directly extracted; When the classroom video data indeed has multi-view shooting features, the next step is to fuse the video data from different perspectives; this process is called perspective fusion processing, which aims to integrate image information from multiple cameras into a unified view, making the presentation of each behavior frame more comprehensive and accurate; the fused video data can provide richer behavior features, helping to better identify key behaviors of students; for example: suppose there are two video streams: perspective 1: camera A captures the moment when a student raises his hand, clearly showing the student's facial expressions and body language; perspective 2: camera B captures the change in the student's sitting position from a panoramic perspective, showing whether the student is talking to others; after perspective fusion processing, the system may combine the data from the two perspectives into a video frame, capturing the student's hand-raising action and simultaneously displaying the student's facial expressions and surrounding environmental information; this fused perspective can provide richer behavior information, helping to more accurately identify and analyze behavior; Once the perspective fusion is complete, the system can extract a sequence of key behavior frames from the fused data; key behavior frames refer to those moments that represent important behaviors of students in the classroom; for example, moments when students raise their hands, answer questions, communicate with classmates, etc.; extracting these frame sequences will provide a data basis for subsequent behavior analysis; for example: suppose the fused video shows that student A raises his hand at 10:02, starts answering questions at 10:03, and continues speaking at 10:04; the system will extract frames of these key moments (e.g. 10:02, 10:03, 10:04) from the video as a sequence of key behavior frames; If the classroom video data does not have the multi-view shooting feature, the system will directly extract the key behavior frame sequence from the existing single-view video data; this means that there is no need for fusion steps, and the system will directly identify important behavior moments according to the student behavior performance in the video data and extract the corresponding frames; for example: assuming that there is a classroom video shot by only one camera, the camera is facing the interaction between the teacher and the students; if in the video, student A raises his hand at 10:00, answers the question at 10:01, and stops speaking at 10:02; the system will directly extract the key behavior frame sequence from these time points, without involving any multi-view fusion.

[0023] In an optional embodiment, the key behavior frame sequence is subjected to behavior segment extraction to obtain a behavior segment set, including: determining whether the key behavior frame sequence meets a preset frame change condition; in response to the key behavior frame sequence meeting the preset frame change condition, using a local dynamic inter-frame difference analysis algorithm to extract behavior segments from the key behavior frame sequence to obtain a behavior segment set; in response to the key behavior frame sequence not meeting the preset frame change condition, using a global dynamic inter-frame difference analysis algorithm to extract behavior segments from the key behavior frame sequence to obtain a behavior segment set.

[0024] It should be noted that it is determined whether the extracted key behavior frame sequence meets the preset “frame change condition”; this condition can be based on the change amplitude between adjacent frames, time interval, and other factors; in simple terms, if the change between adjacent frames is large or the change speed is fast, it can be indicated that the behavior is very active or prominent, and the local dynamic inter-frame difference algorithm is suitable for analysis; on the contrary, if the change is small or the action is slow, a global dynamic inter-frame difference algorithm can be used for more extensive analysis.

[0025] In an optional embodiment, the key behavior frame sequence is subjected to behavior segment extraction using a local dynamic inter-frame difference analysis algorithm to obtain a behavior segment set, including: calculating a local region dynamic change value between adjacent frames in the key behavior frame sequence; determining whether the local region dynamic change value meets a preset frame change condition; if it meets, the adjacent frames before and after meeting the condition are taken as boundary frames of the behavior segment; dividing the key behavior frame sequence based on the boundary frames to obtain a plurality of behavior segments; grouping the plurality of behavior segments into a behavior segment set.

[0026] It should be noted that when the key behavior frame sequence meets the preset inter-frame change condition (for example, the action change is very fast, and the inter-frame difference is large), the system will use the local dynamic inter-frame difference analysis algorithm to extract the behavior segment; the local dynamic difference refers to the rapid change between frames within a certain time window; for example, when a student raises his hand and answers a question, the action change frequency is high, and the system will analyze these rapidly changing parts and extract them into a complete behavior segment.

[0027] In an optional embodiment, the key behavior frame sequence is subjected to behavior segment extraction using a global dynamic inter-frame difference analysis algorithm to obtain a behavior segment set, including: calculating the global dynamic difference value between all adjacent frames in the key behavior frame sequence; determining whether the global dynamic difference value does not meet the preset inter-frame change condition; if not, the key behavior frame sequence is taken as a potential behavior segment; performing internal dynamic change analysis on the potential behavior segment to determine a sub-interval in which the internal dynamic change is less than a preset inter-frame difference threshold; taking the video frame sequence corresponding to the sub-interval as the final behavior segment; combining multiple behavior segments to form a behavior segment set.

[0028] It should be noted that if the key behavior frame sequence has no obvious inter-frame change (for example, the student's action is slow, and the difference between frames is small), the system will select to use the global dynamic inter-frame difference analysis algorithm; unlike the local difference analysis, the global difference algorithm focuses on the change pattern in the entire behavior process, rather than the details of a single local time window; in this case, the system will more extensively analyze the entire sequence in the video and extract the key behavior segment therefrom.

[0029] In an optional embodiment, the behavior type label is classified to obtain a classified behavior label, including: classifying the behavior type label by a preset behavior label classification rule to obtain a classified behavior label; wherein the preset behavior label classification rule is to filter labels in the behavior type label that do not meet a preset positive behavior condition, and to filter labels in the behavior type label that do not meet a preset negative behavior condition.

[0030] It should be noted that in the behavior analysis, some rules (preset behavior label classification rules) are first defined, which are used to classify different types of behavior labels; these rules can be based on the positivity, negativity and other characteristics of the behavior to filter and classify the labels; specifically, the system will: filter out labels that do not meet the positive behavior conditions: that is, identify and remove labels that do not meet the expected positive behavior; filter out labels that do not meet the negative behavior conditions: that is, identify and remove labels that do not meet the expected negative behavior; through these filtering rules, the system can divide the behavior labels into positive behavior labels and negative behavior labels; The preset rules usually set which behaviors belong to "positive behavior" or "negative behavior", and these rules can be defined based on the nature of the behavior itself; for example: positive behavior: obeying classroom discipline, raising hands to speak, actively participating in discussions, etc.; negative behavior: interrupting others, leaving the room, frequently looking down and playing with the phone, etc.; these preset conditions are used to determine whether each label belongs to positive or negative behavior; after filtering by the classification rules, the system will divide the behavior labels into different categories, for example: positive behavior labels: such as "raising hands to speak", "serious listening", etc.; negative behavior labels: such as "daydreaming", "late", etc.

[0031] In an optional embodiment, a time chain is constructed by analyzing the correlation between adjacent time sampling points that meet a preset correlation threshold, to obtain a corresponding time chain, including: The correlation analysis results corresponding to the adjacent two time sampling points are calculated to obtain a correlation value; If the correlation value is greater than the preset correlation threshold, the adjacent two time sampling points are retained, and the time chain is constructed using the retained time sampling points to obtain a corresponding time chain.

[0032] It should be noted that the correlation analysis is mainly through mathematical or statistical methods to calculate the correlation between different time points in a set of data; the correlation usually represents the relationship between the two variables; the greater the correlation value, the stronger the correlation between the two time points; for example, in behavior analysis, time series data can be used to observe the behavior patterns of a user or student at different time points; by calculating the correlation between behavior data at different times, the trend of behavior pattern change at different time points can be understood; the time chain construction refers to selecting the adjacent sampling points with strong correlation in multiple time sampling points to construct a continuous time period (time chain); the time chain is composed of adjacent time sampling points that meet the set correlation threshold; the pre-set correlation threshold is used to filter the correlation between adjacent time points; if the correlation value of two time points is greater than the pre-set threshold, it is considered that the correlation between the two time points is strong enough to be retained and used as part of the time chain; if the correlation value is lower than the threshold, it is considered that there is not enough correlation between the two time points, and it is not retained; in actual operation, the correlation between the two adjacent time points is calculated to obtain the correlation value between them; if the value exceeds the pre-set threshold, the two time points are retained, and a time chain is constructed using the retained time points; Specific example: suppose a student's learning behavior is being analyzed, and the data sampling points are recorded in minutes, recording the student's learning state every minute (for example, "concentrating", "daydreaming", "staring blankly", etc.); the following time points are collected (assuming each time point represents the student's behavior state): time point 1 (0 minutes): "concentrating", time point 2 (1 minute): "concentrating", time point 3 (2 minutes): "daydreaming", time point 4 (3 minutes): "concentrating", time point 5 (4 minutes): "daydreaming", time point 6 (5 minutes): "concentrating", calculate the correlation between adjacent time points; suppose a calculation method (such as correlation coefficient, cosine similarity, etc.) is used, and the following correlation values are obtained: correlation (time point 1, time point 2) = 0.95 (high correlation), correlation (time point 2, time point 3) = 0.2 (low correlation), correlation (time point 3, time point 4) = 0.8 (moderate correlation), correlation (time point 4, time point 5) = 0.3 (low correlation), correlation (time point 5, time point 6) = 0.9 (high correlation), set a preset correlation threshold of 0.7, i.e. only time point pairs with a correlation value greater than or equal to 0.7 can be retained; according to the set threshold 0.7, filter out the adjacent time point pairs with a correlation greater than 0.7: time point 1 and time point 2 (correlation 0.95, meet the threshold), time point 3 and time point 4 (correlation 0.8, meet the threshold), time point 5 and time point 6 (correlation 0.9, meet the threshold), so the final time chain is: time point 1->time point 2->time point 3->time point 4->time point 5->time point 6; after correlation filtering, a complete time chain is obtained, representing the student's learning state change in this period of time; specifically: from 0 to 1 minute: the student has been concentrating, showing a high degree of correlation; from 2 to 3 minutes: the student returned from daydreaming to concentrating, showing a strong correlation of behavior patterns; from 4 to 5 minutes: the student switches between daydreaming and concentrating, still a high correlation behavior.

[0033] In an optional embodiment, the behavior feature vectors corresponding to the time sampling points in the time chain are normalized and similarity operations are performed on the behavior feature vectors, and the time sampling points are filtered according to the similarity values to obtain filtered time sampling points, including: Based on the first time sampling point and the second time sampling point in the time chain, a first behavior feature vector is obtained, and based on the second time sampling point and the third time sampling point, a second behavior feature vector is obtained; The first behavior feature vector and the second behavior feature vector are normalized and similarity operations are performed to obtain a similarity value, and it is determined whether the similarity value meets a preset similarity condition; If the similarity value meets the preset similarity condition, the second time sampling point is removed by screening to obtain a screened time sampling point.

[0034] It should be noted that each time sampling point corresponds to a behavior feature vector, which can be obtained by analyzing the behavior data (such as learning state, body posture, emotional change, etc.) at the time point; the feature vector contains the performance of the time point in a specific behavior, such as "concentrating attention", "daydreaming" or other behavior patterns in the learning state; feature vector normalization means uniform processing of each dimension of the feature vector, so that it has the same scale, avoiding the influence of some dimensions with too large or too small values on the overall analysis result; after normalization, the values of all feature vectors are usually compressed to a fixed range (such as between 0 and 1); this can ensure that the feature vectors of different time sampling points will not be misleading due to scale differences when calculating similarity; similarity calculation is to measure the similarity between two feature vectors, usually using cosine similarity, Euclidean distance or other similarity indicators to calculate; the result is a numerical value representing the similarity between the two feature vectors; according to the similarity value, a preset similarity condition is set, when the behavior feature similarity between two time sampling points exceeds this condition, their behavior is considered similar, and one of the time points can be regarded as redundant or unnecessary, and removed by screening; the specific steps are: first, extract the feature vector of the adjacent time sampling point based on the time chain; for example, assuming that there are two adjacent time sampling points (time point 1 and time point 2), calculate their feature vectors and normalize them; then, calculate the similarity of the two time points, if the similarity is high (i.e. their behavior is very similar), it can be considered that the two time points are redundant, and the second time sampling point does not need to be retained; if the similarity value is low, time point 2 is retained, and the feature vector of the next time point is calculated.

[0035] In an optional embodiment, the screened time sampling points corresponding to the local behavior type satisfying the preset behavior frequency change condition in the classified behavior label are expanded to obtain target time sampling points, and the target behavior mode is obtained by performing mode mining on the target time sampling points, including: The frequency of the screened time sampling points corresponding to the local behavior type satisfying the preset behavior frequency change condition in the classified behavior label is counted to obtain a behavior frequency value, and it is judged whether the behavior frequency value is greater than a preset frequency threshold; If the behavior frequency value is greater than the preset frequency threshold, the screened time sampling points corresponding to the local behavior type are expanded to obtain target time sampling points, and a sequence pattern mining algorithm is used to mine the target behavior mode from the target time sampling points.

[0036] It needs to be noted that the behavior data in a certain period of time is classified, and each time sampling point is assigned a behavior label (e.g., learning, daydreaming, resting, etc.); these behavior labels reflect the specific behavior exhibited by the individual at each time point; certain behavior types corresponding to time sampling points need to be filtered out according to preset behavior frequency change conditions; the behavior frequency change condition usually refers to the frequency of a certain behavior type appearing in a time period, such as frequent appearance, sudden increase or decrease, etc. For each behavior type label, the frequency in a certain time window needs to be calculated, i.e., the number of times a specific behavior occurs; for example, how many times does a certain behavior (such as "learning") occur in a period of time (such as a day, an hour); assuming that the behavior label classification has been completed, and the behavior frequency statistics are applied to the local behavior type; for example, assuming that the focus is on the behavior type of "intensive learning", in the time chain, the frequency of this behavior in each time period is counted; time period 1: intensive learning (occurs 3 times), time period 2: intensive learning (occurs 2 times), time period 3: intensive learning (occurs 5 times); Next, it is determined whether the behavior frequency meets the preset frequency threshold; for example, assuming that the frequency threshold set is that the behavior frequency must be greater than 4 times; then: the frequency of time period 1 is 3 times, which does not meet the threshold and needs to be filtered out; the frequency of time period 2 is 2 times, which does not meet the threshold and needs to be filtered out; the frequency of time period 3 is 5 times, which meets the threshold and can be retained; Once the time periods that meet the frequency threshold are filtered out, the time sampling points can be expanded, i.e., other time points related to the behavior type before and after the time sampling point are considered to form a more complete behavior pattern; for example, assuming that "time period 3" (intensive learning, frequency 5 times) is selected, and it is desired to expand the time range of the behavior pattern; it can be expanded in the following way: time period 3 expansion: the time points before and after "intensive learning" (such as time period 2 and time period 4) are also included in the expansion to construct a target time sampling point range; The expanded target time sampling point will contain behavior data in a series of time periods; next, a sequence pattern mining algorithm is used to discover rules and patterns from these data; sequence pattern mining is a technique used to discover frequently occurring patterns in time series data, usually based on the co-occurrence of certain events in a sequence; for example, assuming that the expanded target time sampling point is analyzed by a sequence pattern mining algorithm, the following target behavior pattern is obtained: target behavior pattern: the student regularly performs "intensive learning" (intensive learning lasts for 30 minutes, followed by a 10-minute rest, and then continues intensive learning) within a day; this pattern may appear multiple times in data sampling, reflecting the regularity of the student's learning.

[0037] Embodiment two, please refer to Figure 2 The application provides a technical scheme: a teaching quality evaluation system based on classroom student behavior analysis, which is suitable for the above-mentioned teaching quality evaluation method based on classroom student behavior analysis, comprising: a segment extraction unit 1, which is used for judging whether the classroom video data has a multi-view shooting feature, determining the key behavior frame sequence of the classroom video data based on the judgment result, performing behavior segment extraction on the key behavior frame sequence, and obtaining a behavior segment set; A label classification unit 2 is used for identifying the student behavior in the behavior segment set to obtain a behavior type label, classifying the behavior type label, and obtaining a classified behavior label. An association analysis unit 3 is used for performing time point sampling on the behavior segment corresponding to the classified behavior label to obtain each time sampling point, performing association analysis on each time sampling point to obtain an association analysis result, and constructing a time chain by performing time chain construction on adjacent time sampling points in the association analysis result that meet a preset association degree threshold, so as to obtain a corresponding time chain. A sampling and screening unit 4 is used for performing feature vector normalization and similarity operation on the behavior feature vector corresponding to the time sampling point in the time chain, screening the time sampling points according to the similarity value, and obtaining screened time sampling points. A pattern mining unit 5 is used for expanding the screened time sampling points corresponding to the local behavior type in the classified behavior label that meets a preset behavior frequency change condition to obtain target time sampling points, performing pattern mining on the target time sampling points, and obtaining a target behavior mode.

[0038] The embodiments of the application are described in detail above in combination with the drawings, but the application is not limited thereto, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application.

Claims

1. A teaching quality evaluation method based on classroom student behavior analysis, characterized in that, The method comprises the following steps: determining whether the classroom video data has a multi-view shooting feature, determining a key behavior frame sequence of the classroom video data based on the determination result, extracting behavior segments from the key behavior frame sequence, and obtaining a behavior segment set; identifying student behavior in the behavior segment set to obtain behavior type labels, and classifying the behavior type labels to obtain classified behavior labels; sampling the behavior segments corresponding to the classified behavior labels at time points to obtain each time sampling point, performing correlation analysis on each time sampling point to obtain a correlation analysis result, constructing a time chain for adjacent time sampling points in the correlation analysis result that meet a preset correlation threshold to obtain a corresponding time chain; performing feature vector normalization and similarity calculation on behavior feature vectors corresponding to the time sampling points in the time chain, and filtering the time sampling points according to the similarity values to obtain filtered time sampling points; performing expansion on the filtered time sampling points corresponding to a local behavior type in the classified behavior labels that meets a preset behavior frequency change condition to obtain target time sampling points, and performing pattern mining on the target time sampling points to obtain a target behavior pattern. 2.The teaching quality evaluation method based on classroom student behavior analysis of claim 1, wherein, determining whether the classroom video data has a multi-view shooting feature, determining a key behavior frame sequence of the classroom video data based on the determination result, comprising: determining whether the classroom video data has a multi-view shooting feature; in response to the classroom video data having a multi-view shooting feature, performing view fusion processing on the multi-view classroom video data, and extracting a key behavior frame sequence from the fused video data; in response to the classroom video data not having a multi-view shooting feature, directly extracting a key behavior frame sequence from the classroom video data. 3.The teaching quality evaluation method based on classroom student behavior analysis of claim 2, wherein, extracting behavior segments from the key behavior frame sequence to obtain a behavior segment set, comprising: determining whether the key behavior frame sequence meets a preset inter-frame change condition; in response to the key behavior frame sequence meeting the preset inter-frame change condition, using a local dynamic inter-frame difference analysis algorithm to extract behavior segments from the key behavior frame sequence to obtain a behavior segment set; in response to the key behavior frame sequence not meeting the preset inter-frame change condition, using a global dynamic inter-frame difference analysis algorithm to extract behavior segments from the key behavior frame sequence to obtain a behavior segment set. 4.The teaching quality evaluation method based on classroom student behavior analysis of claim 3, wherein, using a local dynamic inter-frame difference analysis algorithm to extract behavior segments from the key behavior frame sequence to obtain a behavior segment set, comprising: calculating a local region dynamic change value between adjacent frames in the key behavior frame sequence; determining whether the local region dynamic change value meets a preset inter-frame change condition; if it meets, the adjacent frames before and after meeting the condition are taken as boundary frames of the behavior segments; dividing the key behavior frame sequence based on the boundary frames to obtain a plurality of behavior segments; the plurality of behavior segments are combined into a behavior segment set. 5.The teaching quality evaluation method based on classroom student behavior analysis of claim 4, wherein, using a global dynamic inter-frame difference analysis algorithm to extract behavior segments from the key behavior frame sequence to obtain a behavior segment set, comprising: calculating a global dynamic difference value between all adjacent frames in the key behavior frame sequence; determining whether the global dynamic difference value does not meet a preset inter-frame change condition; if it does not meet, the key behavior frame sequence is taken as a potential behavior segment; performing internal dynamic change analysis on the potential behavior segment to determine a sub-interval with internal dynamic change less than a preset inter-frame difference threshold; taking a video frame sequence corresponding to the sub-interval as a final behavior segment; grouping a plurality of the behavior segments into a behavior segment set. 6.The teaching quality evaluation method based on classroom student behavior analysis of claim 5, wherein, performing classification on the behavior type label to obtain a classified behavior label, including: performing classification on the behavior type label according to a preset behavior label classification rule to obtain the classified behavior label; wherein the preset behavior label classification rule comprises filtering a label in the behavior type label that does not satisfy a preset positive behavior condition, and filtering a label in the behavior type label that does not satisfy a preset negative behavior condition. 7.The teaching quality evaluation method based on classroom student behavior analysis of claim 6, wherein, performing time chain construction on adjacent time sampling points in the correlation analysis result that satisfy a preset correlation threshold to obtain a corresponding time chain, including: calculating a correlation value from the correlation analysis result of the adjacent two time sampling points; if the correlation value is greater than the preset correlation threshold, retaining the adjacent two time sampling points, and performing time chain construction using the retained time sampling points to obtain the corresponding time chain. 8.The teaching quality evaluation method based on classroom student behavior analysis of claim 7, wherein, performing feature vector normalization and similarity calculation on a behavior feature vector corresponding to a time sampling point in the time chain, and screening the time sampling point according to a similarity value to obtain a screened time sampling point, including: obtaining a first behavior feature vector based on a first time sampling point and a second time sampling point of the time chain, and obtaining a second behavior feature vector based on the second time sampling point and a third time sampling point; performing normalization and similarity calculation on the first behavior feature vector and the second behavior feature vector to obtain a similarity value; if the similarity value satisfies a preset similarity condition, screening and removing the second time sampling point to obtain the screened time sampling point. 9.The teaching quality evaluation method based on classroom student behavior analysis of claim 8, wherein, performing expansion on the screened time sampling point corresponding to a local behavior type in the classified behavior label that satisfies a preset behavior frequency change condition to obtain a target time sampling point, and performing pattern mining on the target time sampling point to obtain a target behavior pattern, including: performing frequency statistics on the screened time sampling point corresponding to the local behavior type to obtain a behavior frequency value, and determining whether the behavior frequency value is greater than a preset frequency threshold; if the behavior frequency value is greater than the preset frequency threshold, expanding the screened time sampling point corresponding to the local behavior type to obtain the target time sampling point, and performing pattern mining on the target time sampling point using a sequential pattern mining algorithm to obtain the target behavior pattern.

10. A teaching quality evaluation system based on classroom student behavior analysis, which is suitable for the teaching quality evaluation method based on classroom student behavior analysis according to any one of claims 1-9, characterized in that, including: a segment extraction unit configured to determine whether the classroom video data has a multi-view shooting feature, determine a key behavior frame sequence of the classroom video data based on a determination result, perform behavior segment extraction on the key behavior frame sequence, and obtain a behavior segment set; a label classification unit configured to identify student behavior in the behavior segment set to obtain a behavior type label, and perform classification on the behavior type label to obtain a classified behavior label; and The association analysis unit is configured to perform time point sampling on the behavior segments corresponding to the classified behavior labels to obtain time sampling points, perform association analysis on the time sampling points, obtain association analysis results, and perform time chain construction on adjacent time sampling points in the association analysis results that satisfy a preset association degree threshold to obtain corresponding time chains; The sampling screening unit is configured to perform feature vector normalization and similarity calculation on the behavior feature vectors corresponding to the time sampling points in the time chains, screen the time sampling points according to the similarity values, and obtain screened time sampling points; The pattern mining unit is configured to perform expansion on the screened time sampling points corresponding to local behavior types in the classified behavior labels that satisfy a preset behavior frequency change condition to obtain target time sampling points, perform pattern mining on the target time sampling points, and obtain a target behavior pattern.

Images