A system for real-time comparison and deviation identification of motion postures based on image analysis

Abstract

The present application relates to the technical field of image processing, in particular to a motion posture real-time comparison and deviation identification system based on image analysis, which comprises a motion posture feature identification module, a key point correlation deviation analysis module, a posture deviation situation evaluation module, a trajectory evolution modeling module and a motion compliance identification verification module.In the present application, the position identification sequence in the image sequence is extracted and the three-dimensional space dynamic change mode is analyzed, the adjacent frame feature point Euclidean distance change rate is calculated and the motion vector change mode is extracted, the abnormal jump point and the synchronous deviation are accurately identified, the dynamic change rate is compared with the preset behavior interval, the noise is effectively filtered and the real posture deviation is separated, the trajectory evolution model is established to enhance the identification ability for complex nonlinear motion, the robust tracking in the variable speed environment is ensured, and the real-time accuracy and adaptability of the motion compliance verification are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a real-time motion posture comparison and deviation recognition system based on image analysis. Background Technology

[0002] Image processing technology refers to the field of computer-based image analysis, processing, and recognition. Its core aspects include image acquisition, processing, analysis, and recognition, encompassing technologies such as image enhancement, image segmentation, feature extraction, image registration, and object detection and tracking. With the continuous development of computer vision technology, image processing has been widely applied in fields such as medical imaging, remote sensing monitoring, industrial inspection, and autonomous driving. Image processing technology not only includes traditional image analysis methods but also encompasses the application of modern artificial intelligence technologies in image analysis, such as the combination of deep learning and machine learning methods, continuously expanding the accuracy and application scope of image processing technology.

[0003] Traditional image analysis-based real-time motion posture comparison and deviation recognition systems analyze image data during motion to monitor and compare deviations in motion posture in real time. Traditional solutions employ image registration and feature matching methods, calculating the position information and motion trajectory of differential key points in the image to identify whether posture deviations exist during motion. Traditional methods typically calibrate the position and angle of the moving object and combine image processing techniques to compare motion postures, thereby detecting motion deviations. Traditional technologies often rely on relatively simple image comparison algorithms and have weak adaptability to complex motions or changing environments.

[0004] Traditional techniques rely on fixed-position angle calibration combined with basic geometric feature matching, which makes it difficult to capture highly dynamic changes in complex environments. Over-reliance on static reference systems leads to significant delays when processing continuous motion sequences. Simple comparison algorithms cannot effectively distinguish between background noise and actual posture differences, resulting in high false detection rates in multi-target scenarios. The lack of in-depth correlation analysis of spatiotemporal features makes it impossible to accurately identify minor joint linkage anomalies or trajectory drift. Existing mechanisms cannot dynamically adapt to nonlinear velocity changes or occlusion, thus limiting the improvement of deviation recognition accuracy in real-time applications. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide a real-time motion posture comparison and deviation recognition system based on image analysis. The technical solution is as follows: On the one hand, a real-time motion posture comparison and deviation recognition system based on image analysis is provided, the system comprising: The motion posture feature recognition module obtains the position identifier sequence of the feature points of the moving target based on the runtime sequence features of the moving target in a continuous image sequence, divides the position identifier sequence according to a preset time window, analyzes the dynamic change pattern of each sequence in the two-dimensional spatial coordinate system, and outputs a set of motion feature identifiers. The key point association deviation analysis module calculates the Euclidean distance change rate of feature points in adjacent frames based on the motion feature identifier set, extracts motion vector change patterns, identifies abnormal jump points, analyzes the spatiotemporal correlation between the differential feature points of the motion vector change patterns and the abnormal jump points, and marks the attitude synchronization deviation group. The attitude deviation situation assessment module extracts the dynamic change rate of the position identifier sequence and the key point association features based on the attitude synchronization deviation group, and compares them with the preset deviation behavior interval to obtain a list of motion attitude deviation information. Based on the motion posture deviation information list, the trajectory evolution modeling module collects real-time image response signals of the moving target, extracts posture evolution features for the corresponding time period, performs dimensionality reduction processing on the feature space, and obtains the motion posture trajectory evolution deviation pattern recognition result.

[0006] As a further embodiment of the present invention, the motion feature identifier set includes angular velocity identifier, displacement vector identifier, and key point confidence identifier; the attitude synchronization deviation group includes multi-joint jump position, joint linkage jump time, and jump combination vector; the motion attitude deviation information list includes deviation pattern category, jump correlation strength, and deviation occurrence time interval; and the motion attitude trajectory evolution deviation pattern recognition result includes trajectory boundary movement vector, key feature point distribution density change trend, and motion envelope expansion amplitude.

[0007] As a further aspect of the present invention, the motion posture feature recognition module includes: The spatiotemporal feature extraction submodule extracts the key point coordinates, texture features, and motion optical flow features of moving targets in the image sequence, and inputs them into the preset data buffer channel in frame order for normalization processing to generate runtime sequence feature datasets. The spatiotemporal window segmentation submodule sets a time window with a predetermined number of frames, performs windowed slicing on each type of feature in the runtime sequence feature dataset, extracts the motion trend pattern of each type of feature sequence within the corresponding time window and combines them to generate a runtime sequence identifier sequence set. The dynamic pattern calculation submodule analyzes the derivative distribution of each identifier sequence in the time domain based on the runtime sequence identifier sequence set, organizes the motion pattern sequences of all features, and obtains the motion feature identifier set.

[0008] As a further aspect of the present invention, the key point correlation deviation analysis module includes: The vector jump calculation submodule analyzes the magnitude change of the feature vector between adjacent motion patterns within a continuous time window based on each type of identifier sequence in the motion feature identifier set, and uses this as the instantaneous change trend of each type of identifier sequence to calculate the mode jump variable and generate a jump trend data table of the identifier sequence. The abnormal jump location submodule locates the moment when the jump feature exceeds a preset change threshold on the time axis based on the jump trend data table of the identified sequence and marks it as an abnormal jump point. It then extracts the timestamp and corresponding feature point category of the abnormal jump point to generate an abnormal jump point set. The synchronization deviation marking submodule calculates the overlap probability of the differential feature point jump points in the abnormal jump point set on the time axis, filters the jump combinations that meet the preset synchronization threshold within the same time window, and generates the attitude synchronization deviation group.

[0009] As a further aspect of the present invention, the attitude deviation situation assessment module includes: The feature index construction submodule establishes a multi-dimensional mapping index for the two types of indicators, motion amplitude and motion frequency, according to the feature point type, based on the dynamic change pattern and jump characteristics of the corresponding identifier sequence in each group of the attitude synchronization deviation group, and generates a set of deviation behavior parameters. The deviation interval comparison submodule performs calculations based on the dynamic change patterns and jump characteristics in the deviation behavior parameter set and the preset deviation threshold range of the motion deviation, determines the interval belonging status of the feature parameters, and generates a deviation behavior matching sequence. The deviation information mapping submodule filters the identifier groups in the abnormal interval of the deviation behavior matching sequence, retains the original image frame index information, performs logical mapping between the identifier sequence and the abnormal motion state, and generates a list of motion posture deviation information.

[0010] As a further aspect of the present invention, the trajectory evolution modeling module includes: The image response signal acquisition submodule retrieves the original pixel data of the corresponding time period in the image sequence according to the motion posture deviation information list, converts the pixel grayscale changes and gradient flow into a posture response map, extracts the main motion axis and response intensity distribution value in the map, and generates a state response map dataset. The evolutionary feature analysis submodule locates the initial behavior boundary, key feature point cluster center and trajectory width in the state response map dataset, identifies the geometric range of continuous response distribution in the posture interval, and generates a behavior evolution feature sequence. The boundary drift screening submodule determines whether there is a continuous displacement of the trajectory boundary based on the slope of the change in the number of key feature points in the behavior evolution feature sequence and the normal expansion direction of the trajectory boundary. It also compares the trend of feature point spatial density change with the preset boundary evolution threshold to obtain the motion posture trajectory evolution deviation pattern recognition result.

[0011] As a further aspect of the present invention, determining whether there is a continuous displacement of the trajectory boundary means that if the trend of the number of key feature points changes monotonically within a predetermined number of continuous monitoring periods, and the sum of the offsets of the normal extension direction of the trajectory boundary within a predetermined number of continuous monitoring periods exceeds a preset displacement threshold, then it is determined that there is a continuous displacement of the trajectory boundary.

[0012] As a further aspect of the present invention, the system also includes a motion compliance identification and verification module: The motion compliance identification and verification module identifies the current posture comparison result of the moving target based on the motion posture trajectory evolution deviation pattern recognition result, obtains the reference trajectory information associated with the moving target, verifies the motion state with the preset standard posture sequence, performs posture deviation judgment, and obtains the motion posture comparison verification conclusion.

[0013] As a further aspect of the present invention, the motion posture comparison and verification conclusion includes the quantified value of posture deviation, the category of motion trajectory, and the direction of posture evolution trend.

[0014] As a further aspect of the present invention, the motion compliance identification and verification module includes: The running identifier construction submodule, based on the motion posture trajectory evolution deviation pattern recognition result and combined with the motion target temporal behavior data extracted from the image sequence, locates the motion mode switching segment on the time axis, performs feature marking on the posture transition trend in the signal, and generates a motion transition state identifier sequence. The compliance judgment submodule retrieves the number of deviation behaviors, the duration of deviation, and the trajectory boundary offset vector within the current time slice based on the motion transition state identifier sequence, performs multi-dimensional joint interval fitting, identifies the curvature change amplitude of the current sampling point of the fitted curve, and obtains the attitude comparison result. The comparison conclusion output submodule extracts the deviation quantization score and the current motion transition state identifier based on the attitude comparison results, and outputs them through the real-time data transmission channel to obtain the motion attitude comparison verification conclusion.

[0015] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: By extracting position marker sequences from continuous image sequences and analyzing dynamic change patterns in three-dimensional space, the reliance on static calibration is overcome. The Euclidean distance change rate of feature points in adjacent frames is calculated and motion vector change patterns are extracted, enabling accurate identification of abnormal jump points and synchronization deviations. The dynamic change rate is matched and compared with preset behavior intervals, effectively filtering noise and separating true posture deviations. Dimensionality reduction processing is performed on the feature space and a trajectory evolution model is established, enhancing the recognition capability for complex nonlinear motions. This logic ensures robust tracking of joint linkage anomalies and trajectory drift under speed changes or occlusion, significantly improving the real-time accuracy and adaptability of motion compliance verification, and providing multi-dimensional data support for posture correction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of a real-time motion posture comparison and deviation recognition system based on image analysis provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention; Figure 3 This is a flowchart of the motion posture feature recognition module in this invention; Figure 4 This is a flowchart of the key point correlation deviation analysis module in this invention; Figure 5 This is a flowchart of the attitude deviation situation assessment module in this invention; Figure 6 This is a flowchart of the trajectory evolution modeling module in this invention; Figure 7 This is a flowchart of the motion compliance identification and verification module in this invention. Detailed Implementation

[0018] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0019] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0020] This invention provides a real-time motion posture comparison and deviation recognition system based on image analysis, such as... Figure 1-2 The diagram shown illustrates a real-time motion posture comparison and deviation recognition system based on image analysis. This system includes: The motion posture feature recognition module identifies dynamic posture parameters based on the runtime sequence features of the moving target in a continuous image sequence, obtains the position identifier sequence of the moving target feature points, divides the position identifier sequence according to a preset time window, analyzes the dynamic change pattern of each segment in the two-dimensional spatial coordinate system, and outputs a set of motion feature identifiers. The key point association deviation analysis module extracts the motion vector change pattern of each position identifier sequence based on the motion feature identifier set. By calculating the Euclidean distance change rate of feature points between adjacent frames, it identifies abnormal jump points in motion behavior, compares the spatiotemporal correlation between differential feature points and jump points, and marks the attitude synchronization deviation group. The attitude deviation situation assessment module is based on the attitude synchronization deviation group. It extracts the dynamic change rate of the position identifier sequence and the key point association features, matches and compares them with the preset deviation behavior interval, and filters to obtain a list of motion attitude deviation information. The trajectory evolution modeling module is based on the list of motion posture deviation information. It collects the real-time image response signal of the moving target, extracts the posture evolution features within the time period corresponding to the position identifier sequence, and obtains the motion posture trajectory evolution deviation pattern recognition result by performing dimensionality reduction processing on the feature space. The motion compliance identification and verification module identifies the current posture comparison result of the moving target based on the motion posture trajectory evolution deviation pattern recognition result, obtains the reference trajectory information associated with the moving target, verifies the motion state with the preset standard posture sequence, judges the posture deviation, and obtains the motion posture comparison and verification conclusion. The motion feature identifier set includes angular velocity identifiers, displacement vector identifiers, and key point confidence identifiers. The attitude synchronization deviation group includes multi-joint jump positions, joint linkage jump times, and jump combination vectors. The motion attitude deviation information list includes deviation pattern categories, jump correlation strength, and deviation occurrence time intervals. The motion attitude trajectory evolution deviation pattern recognition results include trajectory boundary movement vectors, key feature point distribution density change trends, and motion envelope expansion amplitude. The motion attitude comparison and verification conclusions include attitude deviation quantification values, motion trajectory categories, and attitude evolution trend directions.

[0021] Specifically, such as Figure 2 , 3 As shown, the motion posture feature recognition module includes: The spatiotemporal feature extraction submodule extracts the key point coordinates, texture features, and motion optical flow features of moving targets in the image sequence, and inputs them into the preset data buffer channel in frame order for normalization processing to generate runtime sequence feature datasets. A high-definition camera acquisition terminal was connected to acquire a raw image sequence with a resolution of 1920×1080 pixels at a sampling rate of 60 frames per second. Color space conversion was performed on the image sequence, converting the RGB color space pixel data to grayscale data using a weighted average method. The weights for the red channel were set to 0.299, the green channel to 0.587, and the blue channel to 0.114. Based on this grayscale image, a multi-layer convolutional neural network architecture was used to locate key points on the human body. The first convolutional layer used a 3×3 kernel to perform a scanning operation with a stride of 1 on the input image to extract edge features. Negative values ​​were set to 0 by modifying the linear unit activation function to enhance the non-linear feature expression. A max-pooling layer was used to select the maximum pixel value within a 2×2 region for dimensionality reduction. The dimensionality-reduced feature data was then fed into subsequent cascaded deep convolutional layers to extract high-dimensional semantic features. Coordinate prediction was performed using a fully connected regression layer, outputting two-dimensional coordinate data containing 18 skeletal key points of the human body, including the horizontal and vertical coordinates of the head, neck, shoulders, elbows, wrists, hips, knees, and ankles. For texture feature extraction, a local binary mode algorithm is used. The grayscale value of the center pixel is compared with that of its eight neighboring pixels. Pixels with a grayscale value greater than the center pixel are marked as 1, and those less than are marked as 0. An 8-bit binary code is generated and converted to decimal as the texture feature value. For motion optical flow features, the optical flow method is used to calculate the displacement vector of corresponding pixels in two adjacent frames. The image grayscale constraint equation is expanded using Taylor series, and the horizontal and vertical velocity components are solved by simultaneously solving for the spatial and temporal gradients. The extracted keypoint coordinates, texture feature values, and optical flow velocity components are accessed in frame order via a high-speed bus to the image processing platform's data buffer channel. Within the buffer channel, Z-score normalization is performed. The arithmetic mean and standard deviation of each type of feature data in the current buffer queue are calculated. The original feature value is subtracted from the arithmetic mean and then divided by the standard deviation to eliminate the influence of dimensions. For example, if the original horizontal coordinate of a keypoint is 500 pixels, and the mean of the buffer queue coordinates within that time period is 400 pixels with a standard deviation of 50 pixels, the normalized feature value is 2. After normalization, the data is stored in a structured manner, generating a runtime sequence feature dataset.

[0022] The spatiotemporal window partitioning submodule sets a time window with a predetermined number of frames, performs windowed slicing on each type of feature in the runtime sequence feature dataset, extracts the motion trend pattern of each type of feature sequence within the corresponding time window and combines them to generate a runtime sequence identifier sequence set. Based on the average periodicity of motion, a predetermined time window length of 30 frames is set, and a sliding step size of 10 frames is set, meaning there is a 20-frame overlap between adjacent windows to ensure continuous feature extraction. The runtime sequence feature dataset is traversed, and slicing operations are performed on keypoint coordinates, texture features, and optical flow features according to the set window length, generating a series of local time window data blocks. Within each time window, the motion trend pattern of each type of feature sequence is extracted, and represented by the mean and variance of the slope of the feature data changing over time within the window. For example, for the ordinate sequence of a keypoint within 30 frames, the coordinate difference between each adjacent two frames is calculated, summed, and divided by the total number of frame intervals to obtain the average movement speed. If the average movement speed is positive and the variance is less than the preset stability threshold of 0.5, the motion trend pattern of this feature sequence is marked as stable upward. The spatiotemporal window segmentation submodule performs combined encoding on the extracted keypoint trends, texture change trends, and optical flow distribution trends within the same time window to generate a unique runtime sequence identifier corresponding to that window. All window identifiers are arranged in chronological order to generate a runtime sequence identifier sequence set.

[0023] The dynamic pattern calculation submodule analyzes the derivative distribution of each identifier sequence in the time domain based on the runtime sequence identifier sequence set, organizes the motion pattern sequences of all features, and obtains the motion feature identifier set. The system receives a set of runtime sequence identifiers. For each identifier sequence, it analyzes the distribution of its derivatives in the time domain. The second-order central difference method is used to calculate the first-order derivative velocity and second-order derivative acceleration of the characteristic values ​​in the identifier sequence, thereby constructing a dynamic change curve. This curve is traversed to identify zero-crossing points where the derivative sign flips, defining these as action peaks or troughs. The time span and amplitude difference between adjacent peaks and troughs are statistically analyzed. Based on the statistical results, the current derivative distribution features are matched against a preset action primitive library. If the first-order derivative of a sequence remains positive for five consecutive time units and the second-order derivative changes from positive to negative, it is matched as an acceleration-deceleration motion pattern. All features, including 18 keypoints, texture, and optical flow, are summarized and organized for motion patterns in each time window. A structured information system containing action type, start time, duration, and intensity level is constructed, resulting in a motion feature identifier set, as shown in Table 1, which displays some feature normalization processing and intermediate data for pattern calculation. Table 1: Runtime Timing Feature Data and Pattern Calculation Table As shown in Table 1, after the coordinate changes of the right elbow key point in consecutive frames are normalized, the dynamic mode in this time period is deduced by combining the optical flow component, which gradually changes from accelerated rise to a more stable state. This result is incorporated into the motion feature identifier set.

[0024] Specifically, such as Figure 2 , 4 As shown, the key point correlation deviation analysis module includes: The vector jump calculation submodule analyzes the magnitude change of the feature vector between adjacent motion patterns within a continuous time window based on each type of identifier sequence in the motion feature identifier set. This serves as the instantaneous change trend of each type of identifier sequence. The module then calculates the pattern jump variable and generates a jump trend data table for the identifier sequence. The system reads and stores each type of identifier sequence from motion feature identifiers. For each feature vector within a time window, it performs magnitude change analysis. First, it constructs the feature vector by normalizing the x-coordinate and y-coordinate of the same key point in the current frame as vector components. Using the Pythagorean theorem, it calculates the square root of the sum of the squares of the x-coordinate and y-coordinate to obtain the magnitude of the current feature vector. Then, it calculates the difference in magnitude between two adjacent moments within a continuous time window, taking the absolute value of this difference as the instantaneous change trend. For example, if the feature vector magnitude at time t1 is 10 and at time t2 is 12, the calculated magnitude change is 2. Further, it calculates the mode jump variable by comparing the difference between the current instantaneous change trend and the previous instantaneous change trend. If the change at the previous moment was 0.5 and the change at the current moment is 2, the mode jump variable is 1.5. This calculation is performed on all identifier sequences one by one. The calculated mode jump variables are then entered into the database according to timestamp and feature category, generating an identifier sequence jump trend data table.

[0025] The abnormal jump location submodule locates the moment when the jump feature exceeds the preset change threshold on the time axis based on the jump trend data table of the identifier sequence and marks it as an abnormal jump point. It then extracts the timestamp and corresponding feature point category of the abnormal jump point to generate a set of abnormal jump points. The system accesses a data table of jump trends in the identifier sequence, setting a preset threshold of 3.0 for anomaly detection. This threshold is based on statistical analysis of the maximum speed of normal human movement. It shows that the normalized displacement of human joints rarely exceeds 3 standard units within 1 / 60th of a second of sampling time. Each jump feature value in the data table is iterated and compared with the preset threshold. If a key point's jump value is greater than 3.0 at a specific moment, an anomaly is identified. This moment is locked on the timeline and marked as an anomalous jump point. The system extracts the specific timestamp of this anomalous jump point (accurate to milliseconds), the corresponding feature point category (e.g., left knee joint), and the specific jump amplitude. This information is encapsulated into an anomaly record object, and a set of anomalous jump points is generated.

[0026] The synchronization deviation marking submodule calculates the overlap probability of differential feature points in the abnormal jump point set on the time axis, filters the jump combinations that meet the preset synchronization threshold within the same time window, and generates attitude synchronization deviation groups. Multi-point correlation analysis is performed on the data in the set of anomalous jump points. First, the probability of overlap on the time axis is calculated for differential feature points in the set, such as anomalous jump points of the left wrist and left elbow. Specifically, a synchronization time judgment window with a width of 50 milliseconds is set. If the absolute value of the difference in the timestamps of the anomalous jumps of two differential feature points is less than or equal to 50 milliseconds, the two jump points are determined to overlap in time. The overlap probability is calculated, which is the ratio of the number of anomalous point pairs that meet the time overlap condition to the total number of anomalous point pairs. Then, jump combinations that meet the preset synchronization threshold within the same time window are selected. The preset synchronization threshold is set to 0.8, which means that if two key point anomalous jumps occur simultaneously in 80% of cases, a linkage attitude deviation is considered to exist. The feature point combinations that meet the condition and their corresponding time periods are packaged to generate an attitude synchronization deviation group for subsequent situation assessment.

[0027] Specifically, such as Figure 2 , 5 As shown, the attitude deviation situation assessment module includes: The feature index construction submodule establishes a multi-dimensional mapping index for the two types of indicators, motion amplitude and motion frequency, based on the dynamic change pattern and jump characteristics of the corresponding identifier sequence in each group of attitude synchronization deviation groups, and generates a set of deviation behavior parameters. Based on the data provided by the attitude synchronization deviation group, dynamic change patterns such as oscillation, abrupt change, stiffness, and jump feature values ​​are extracted from each group of identifier sequences. A multi-dimensional mapping index mechanism is established, which includes two main dimensions: motion amplitude index and motion frequency index. For motion amplitude, the maximum displacement range of feature points within the deviation group during the abnormal time period is calculated. For motion frequency, the time-domain signal is converted into a frequency-domain signal through Fast Fourier Transform logic, and the frequency peaks of the main energy concentration are extracted. The calculated amplitude and frequency values ​​are used as index keys and mapped to specific feature point types. For example, for the right shoulder feature point, if its average oscillation amplitude during the abnormal segment is calculated to be 15 cm and the oscillation frequency is 6 Hz, a tuple containing the parameters "right shoulder", "15", and "6" will be generated. This operation is performed on all deviation groups to generate a deviation behavior parameter set containing multi-dimensional feature parameters.

[0028] The deviation interval comparison submodule performs calculations based on the dynamic change patterns and jump characteristics in the deviation behavior parameter set and the preset deviation threshold range of the motion deviation, determines the interval belonging status of the feature parameters, and generates a deviation behavior matching sequence. A pre-defined library of detailed motion deviation threshold ranges is used. This library categorizes deviation levels into three levels: slight disturbance, moderate deviation, and severe instability. The pre-defined threshold ranges are based on a human biomechanical balance model. For example, for upper limb joints, a frequency of 0 to 4 Hz and an amplitude of less than 10 cm is defined as the normal range; a frequency of 4 to 8 Hz or an amplitude of 10 to 20 cm is defined as the moderate deviation range; and a frequency greater than 8 Hz or an amplitude greater than 20 cm is defined as the severe instability range. Each set of parameters in the deviation behavior parameter set is read, and its amplitude and frequency values ​​are compared with the pre-defined deviation threshold ranges. If a parameter set has an amplitude of 15 cm (falling into the moderate range) and a frequency of 6 Hz (falling into the moderate range), the characteristic parameter range is determined to belong to the moderate deviation category. The determination results are then arranged in a time series to generate a deviation behavior matching sequence.

[0029] The deviation information mapping submodule filters the identifier groups in the abnormal interval of the deviation behavior matching sequence, retains the original image frame index information, performs logical mapping between the identifier sequence and the abnormal motion state, and generates a list of motion posture deviation information. The sequence matching the deviation behavior is filtered, removing segments judged as normal or slightly disturbed, retaining only the groups marked as moderate deviation and severe instability. The original image frame index information corresponding to these abnormal groups is read, recording the start and end frame numbers of the anomaly. Subsequently, a logical mapping operation is performed to associate abstract deviation states, such as severe instability, with specific motion anomaly semantics. For example, if both legs simultaneously exhibit severe instability for more than 2 seconds, the semantic mapping table is consulted to associate it with fall risk or gait abnormality. All mapping results are integrated to generate a list of motion posture deviation information containing frame indexes, involved key points, deviation levels, and corresponding anomaly semantic descriptions.

[0030] Specifically, such as Figure 2 , 6 As shown, the trajectory evolution modeling module includes: The image response signal acquisition submodule retrieves the original pixel data of the corresponding time period in the image sequence according to the motion posture deviation information list, converts the pixel grayscale changes and gradient flow into a posture response map, extracts the main motion axis and response intensity distribution value in the map, and generates a state response map dataset. The system receives a list of motion attitude deviation information. Based on the start and end frame indices recorded in the list, it retrieves the pixel data of the original image sequence for the corresponding time period from the underlying memory. For each frame, it calculates the pixel grayscale gradient flow. Specifically, it applies the Sobel operator to calculate the grayscale gradient values ​​in the horizontal and vertical directions, and calculates the magnitude and direction of the gradient. Based on the gradient magnitude, it generates an attitude response map. Highlighted areas in the map represent areas with intense motion or clear edges. It further extracts the principal motion axis from the map and determines the first moment (center of gravity) and second moment (principal axis of inertia) of the image by calculating the spatial moments of the highlighted pixels. The principal motion axis is the direction of the principal axis of inertia. At the same time, it statistically analyzes the response intensity distribution along the principal motion axis, i.e., it accumulates the gradient magnitude in the neighborhood of the axis. It organizes the principal motion axis angle, center of gravity coordinates, and response intensity distribution values ​​of each frame in chronological order to generate a state response map dataset.

[0031] The evolutionary feature analysis submodule locates the initial behavioral boundary, key feature point cluster centers, and trajectory width in the state response map dataset, identifies the geometric range of continuous response distribution in the attitude interval, and generates a behavioral evolution feature sequence. In-depth processing is performed on the state response map dataset. First, the initial behavioral boundaries are located. Using the K-means clustering algorithm, pixels with response intensities greater than a preset background noise threshold (e.g., 20 pixels) are selected as data samples. The number of cluster centers is set to three, corresponding to the head, torso, and limbs. After multiple iterations and convergence, the coordinates of the key feature point cluster centers are obtained. Next, the trajectory width is calculated, i.e., the geometric span covering 95% of the high-response pixels along the normal direction of the main motion axis. The changes in the cluster centers and trajectory width are tracked over time to identify the geometric range of continuous response distribution within the posture interval, for example, forming a tubular trajectory envelope that extends over time. The sequences of cluster center trajectories and trajectory width changes are combined to generate a behavioral evolution feature sequence.

[0032] The boundary drift screening submodule determines whether there is a continuous displacement of the trajectory boundary based on the slope of the change in the number of key feature points in the behavior evolution feature sequence and the normal expansion direction of the trajectory boundary. It also compares the trend of feature point spatial density change with the preset boundary evolution threshold to obtain the motion posture trajectory evolution deviation pattern recognition result. Determining whether there is continuous displacement of the trajectory boundary means that if the trend of the number of key feature points changes monotonically within a predetermined number of continuous monitoring periods, and the sum of the offsets of the normal expansion direction of the trajectory boundary within a predetermined number of continuous monitoring periods exceeds a preset displacement threshold, then it is determined that there is continuous displacement of the trajectory boundary. First, the slope of the change in the number of key feature points in the behavioral evolution feature sequence is calculated. Ten consecutive monitoring periods are selected as a sliding window, and the trend of the number of feature points changing over time is linearly fitted using the least squares method to obtain the slope value. Simultaneously, the normal expansion direction of the trajectory boundary is calculated, i.e., the projection of the displacement vector of the trajectory width boundary point on the trajectory normal line at adjacent time points is calculated. A persistent displacement judgment logic is executed: if the slope of the change in the number of key feature points is greater than 0 and monotonically increasing or less than 0 and monotonically decreasing over ten consecutive monitoring periods, and the sum of the normal expansion offsets of the trajectory boundary exceeds a preset displacement threshold, then persistent displacement is determined to exist. The preset displacement threshold is set to 20% of the average trajectory width. For example, if the average trajectory width is 50 pixels and the threshold is 10 pixels, when the cumulative normal offset reaches 12 pixels over ten consecutive periods, boundary drift is determined. Combined with the feature point spatial density, i.e., the trend of the number of feature points per unit area, if the density decreases synchronously, trajectory divergence is further confirmed. The output includes the drift type, occurrence time, and severity of the motion posture trajectory evolution deviation pattern recognition result.

[0033] Specifically, such as Figure 2 , 7 As shown, the sports compliance identification and verification module includes: The running identifier construction submodule, based on the motion posture trajectory evolution deviation pattern recognition results and combined with the motion target temporal behavior data extracted from the image sequence, locates the motion mode switching segment on the time axis, performs feature marking on the posture transition trend in the signal, and generates a motion transition state identifier sequence. Based on the motion posture trajectory evolution deviation pattern recognition results, and combined with the temporal behavior data of the moving target extracted from the image sequence, such as the overall displacement velocity and direction, the motion mode switching segment is accurately located on the time axis. The inflection point of the velocity curve or the abrupt change point of the trajectory direction is identified as the mode switching boundary, such as the switch from stationary to moving, or from walking to running. Feature labeling is performed on the posture transition trend in the signal, and the acceleration peak and posture deformation rate parameters during the transition are attached as labels to the time period to generate a motion transition state identification sequence. This sequence clearly describes the dynamic process of the action switching between different modes.

[0034] The compliance judgment submodule retrieves the number of deviation behaviors, the duration of deviation, and the trajectory boundary offset vector within the current time slice based on the motion transition state identifier sequence, performs multi-dimensional joint interval fitting, identifies the curvature change amplitude of the current sampling point of the fitted curve, and obtains the attitude comparison result. Based on the motion transition state identifier sequence, the number of deviation behaviors (i.e., the number of abnormal points identified by the preceding module), the duration of deviation (i.e., the duration of the abnormal state), and the trajectory boundary offset vector are retrieved within the current time slice. A multidimensional joint interval fitting operation is performed to construct a polynomial function to fit the relationship between this parameter and time. The focus is on identifying the curvature change amplitude of the fitted curve at the current sampling point. The curvature calculation logic is to calculate the absolute value of the second derivative of the fitted function, divide it by 1, and add the square of the first derivative to the power of 1.5. If the curvature change amplitude exceeds the compliance benchmark value, such as 0.05, it means that the motion change is too drastic and does not meet the smooth transition compliance requirements. The calculated curvature value is compared with the compliant motion template in the standard library to obtain the posture comparison result. This result includes a compliance or violation judgment label and a specific difference value.

[0035] The comparison result output submodule extracts the deviation quantification score and the current motion transition state identifier based on the attitude comparison result, and outputs them through the real-time data transmission channel to obtain the motion attitude comparison verification result; Based on the posture comparison results, a quantitative scoring logic is executed. A base score of 100 points is set, and a deviation quantification score is extracted. This score is calculated by weighting the number of deviations, duration, and curvature difference, with the weights set as follows: deviation number weight 0.4, duration weight 0.3, and curvature difference weight 0.3. For example, if the deviation number results in a deduction of 10 points, duration a deduction of 5 points, and curvature difference a deduction of 5 points, the total deduction is 20 points, resulting in a final score of 80 points. This score is then packaged with the current motion transition state indicator, such as "running," and transmitted back to the monitoring terminal via a real-time data transmission channel, such as TCP / IP protocol. The monitoring terminal, based on the received data, presents a visual motion posture comparison verification conclusion on the display interface. If the movement is compliant, the score is 80; if the movement is non-compliant, there is excessive knee deviation. Table 2 shows the key parameters and conclusions in the compliance assessment process. As shown in Table 2, the evaluation is performed on the transition state within the differentiated time slice. Among them, the fast running state with ID T202303 has a score of only 45 points due to the long deviation time and large curvature change. The final output is a violation conclusion, which is directly used to guide subsequent correction or alarm operations.

[0036] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using differentiated programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. A real-time motion posture comparison and deviation recognition system based on image analysis, characterized in that, The system includes: The motion posture feature recognition module obtains the position identifier sequence of the feature points of the moving target based on the runtime sequence features of the moving target in a continuous image sequence, divides the position identifier sequence according to a preset time window, analyzes the dynamic change pattern of each sequence in the two-dimensional spatial coordinate system, and outputs a set of motion feature identifiers. The key point association deviation analysis module calculates the Euclidean distance change rate of feature points in adjacent frames based on the motion feature identifier set, extracts motion vector change patterns, identifies abnormal jump points, analyzes the spatiotemporal correlation between the differential feature points of the motion vector change patterns and the abnormal jump points, and marks the attitude synchronization deviation group. The attitude deviation situation assessment module extracts the dynamic change rate of the position identifier sequence and the key point association features based on the attitude synchronization deviation group, and compares them with the preset deviation behavior interval to obtain a list of motion attitude deviation information. Based on the motion posture deviation information list, the trajectory evolution modeling module collects real-time image response signals of the moving target, extracts posture evolution features for the corresponding time period, performs dimensionality reduction processing on the feature space, and obtains the motion posture trajectory evolution deviation pattern recognition result.

2. The real-time motion posture comparison and deviation recognition system based on image analysis according to claim 1, characterized in that: The motion feature identifier set includes angular velocity identifier, displacement vector identifier, and key point confidence identifier. The attitude synchronization deviation group includes multi-joint jump position, joint linkage jump time, and jump combination vector. The motion attitude deviation information list includes deviation pattern category, jump correlation strength, and deviation occurrence time interval. The motion attitude trajectory evolution deviation pattern recognition result includes trajectory boundary movement vector, key feature point distribution density change trend, and motion envelope expansion amplitude.

3. The real-time motion posture comparison and deviation recognition system based on image analysis according to claim 1, characterized in that: The motion posture feature recognition module includes: The spatiotemporal feature extraction submodule extracts the key point coordinates, texture features, and motion optical flow features of moving targets in the image sequence, and inputs them into the preset data buffer channel in frame order for normalization processing to generate runtime sequence feature datasets. The spatiotemporal window segmentation submodule sets a time window with a predetermined number of frames, performs windowed slicing on each type of feature in the runtime sequence feature dataset, extracts the motion trend pattern of each type of feature sequence within the corresponding time window and combines them to generate a runtime sequence identifier sequence set. The dynamic pattern calculation submodule analyzes the derivative distribution of each identifier sequence in the time domain based on the runtime sequence identifier sequence set, organizes the motion pattern sequences of all features, and obtains the motion feature identifier set.

4. The real-time motion posture comparison and deviation recognition system based on image analysis according to claim 3, characterized in that: The key point correlation deviation analysis module includes: The vector jump calculation submodule analyzes the magnitude change of the feature vector between adjacent motion patterns within a continuous time window based on each type of identifier sequence in the motion feature identifier set, and uses this as the instantaneous change trend of each type of identifier sequence to calculate the mode jump variable and generate a jump trend data table of the identifier sequence. The abnormal jump location submodule locates the moment when the jump feature exceeds a preset change threshold on the time axis based on the jump trend data table of the identified sequence and marks it as an abnormal jump point. It then extracts the timestamp and corresponding feature point category of the abnormal jump point to generate an abnormal jump point set. The synchronization deviation marking submodule calculates the overlap probability of the differential feature point jump points in the abnormal jump point set on the time axis, filters the jump combinations that meet the preset synchronization threshold within the same time window, and generates the attitude synchronization deviation group.

5. The real-time motion posture comparison and deviation recognition system based on image analysis according to claim 4, characterized in that: The attitude deviation situation assessment module includes: The feature index construction submodule establishes a multi-dimensional mapping index for the two types of indicators, motion amplitude and motion frequency, according to the feature point type, based on the dynamic change pattern and jump characteristics of the corresponding identifier sequence in each group of the attitude synchronization deviation group, and generates a set of deviation behavior parameters. The deviation interval comparison submodule performs calculations based on the dynamic change patterns and jump characteristics in the deviation behavior parameter set and the preset deviation threshold range of the motion deviation, determines the interval belonging status of the feature parameters, and generates a deviation behavior matching sequence. The deviation information mapping submodule filters the identifier groups in the abnormal interval of the deviation behavior matching sequence, retains the original image frame index information, performs logical mapping between the identifier sequence and the abnormal motion state, and generates a list of motion posture deviation information.

6. The real-time motion posture comparison and deviation recognition system based on image analysis according to claim 5, characterized in that: The trajectory evolution modeling module includes: The image response signal acquisition submodule retrieves the original pixel data of the corresponding time period in the image sequence according to the motion posture deviation information list, converts the pixel grayscale changes and gradient flow into a posture response map, extracts the main motion axis and response intensity distribution value in the map, and generates a state response map dataset. The evolutionary feature analysis submodule locates the initial behavior boundary, key feature point cluster center and trajectory width in the state response map dataset, identifies the geometric range of continuous response distribution in the posture interval, and generates a behavior evolution feature sequence. The boundary drift screening submodule determines whether there is a continuous displacement of the trajectory boundary based on the slope of the change in the number of key feature points in the behavior evolution feature sequence and the normal expansion direction of the trajectory boundary. It also compares the trend of feature point spatial density change with the preset boundary evolution threshold to obtain the motion posture trajectory evolution deviation pattern recognition result.

7. The real-time motion posture comparison and deviation recognition system based on image analysis according to claim 6, characterized in that: The determination of whether the trajectory boundary has a continuous displacement means that if the trend of the number of key feature points changes monotonically within a predetermined number of continuous monitoring periods, and the sum of the offsets of the normal expansion direction of the trajectory boundary within a predetermined number of continuous monitoring periods exceeds a preset displacement threshold, then the trajectory boundary is determined to have a continuous displacement.

8. The real-time motion posture comparison and deviation recognition system based on image analysis according to claim 1, characterized in that: The system also includes a sports compliance identification and verification module: The motion compliance identification and verification module identifies the current posture comparison result of the moving target based on the motion posture trajectory evolution deviation pattern recognition result, obtains the reference trajectory information associated with the moving target, verifies the motion state with the preset standard posture sequence, performs posture deviation judgment, and obtains the motion posture comparison verification conclusion.

9. The real-time motion posture comparison and deviation recognition system based on image analysis according to claim 8, characterized in that: The motion posture comparison and verification conclusions include the quantified value of posture deviation, the category of motion trajectory, and the direction of posture evolution trend.

10. The real-time motion posture comparison and deviation recognition system based on image analysis according to claim 8, characterized in that: The sports compliance identification and verification module includes: The running identifier construction submodule, based on the motion posture trajectory evolution deviation pattern recognition result and combined with the motion target temporal behavior data extracted from the image sequence, locates the motion mode switching segment on the time axis, performs feature marking on the posture transition trend in the signal, and generates a motion transition state identifier sequence. The compliance judgment submodule retrieves the number of deviation behaviors, the duration of deviation, and the trajectory boundary offset vector within the current time slice based on the motion transition state identifier sequence, performs multi-dimensional joint interval fitting, identifies the curvature change amplitude of the current sampling point of the fitted curve, and obtains the attitude comparison result. The comparison conclusion output submodule extracts the deviation quantization score and the current motion transition state identifier based on the attitude comparison results, and outputs them through the real-time data transmission channel to obtain the motion attitude comparison verification conclusion.

Images