Behavior recognition-based human rhythm analysis methods, regulation methods, and analysis systems

Abstract

This invention discloses a method, regulation method, and analysis system for human rhythm analysis based on behavior recognition. The analysis method includes the following steps: acquiring activity images of the person to be analyzed to form an image sequence; recognizing the image sequence to obtain behavior recording data; processing the behavior recording data to obtain behavior curves; and inputting the behavior curves into a neural network model to calculate the activity-rest transition points and their time-series curves based on a neural network model that predicts activity-rest transition points. This invention acquires activity images using a vision system, performs behavior recognition on the activity images to form behavior recording data, and converts the behavior recording data into behavior curves suitable for deep learning. By optimizing the data processing algorithm and using a neural network model to predict the behavior curves, the human rhythm of the person to be analyzed is obtained, ensuring efficiency in processing large amounts of behavior recording data and calculating activity-rest transition points, while also exhibiting good accuracy.

Description

Technical Field

[0001] This invention relates to the field of human rhythm analysis technology, and in particular to human rhythm analysis methods, regulation methods and analysis systems based on behavior recognition. Background Technology

[0002] Studies have revealed that core genes and proteins, such as CLOCK and BMAL1, control the human circadian rhythm. These genes and proteins regulate the biological clock through a transcription-translation negative feedback loop. Light is considered the strongest external synchronizing factor, transmitting signals from the retina to the suprachiasmatic nucleus and influencing the biological clock. By changing the intensity and duration of light exposure, light therapy can help subjects reset their biological clock, control their circadian rhythm, and improve the quality of sleep and daily activities.

[0003] To regulate human circadian rhythms, it is essential to first record them. Currently, many methods exist, such as using wearable device data (e.g., wrist motion recorders and photometers) to predict the circadian rhythm misalignment of night shift workers. However, the accuracy of this method's circadian rhythm prediction model heavily relies on the data quality of the wearable devices. Due to factors such as wearing method, device calibration, and usage habits, data errors may exist, affecting the model's prediction accuracy. Furthermore, these errors are greater when predicting circadian rhythms over longer periods. Existing literature discusses methods for integrating wearable device data into circadian rhythm models, exploring the integration of data from wearable devices (e.g., activity trackers, smartwatches) into these models to better understand and predict changes in the human biological clock. However, the data quality and consistency of wearable devices can be affected by various factors, such as device wearing conditions and technical malfunctions, which may impact the accuracy and reliability of the data. Moreover, significant physiological and behavioral differences exist between individuals, and general models may perform poorly when adapting to different individuals, requiring extensive personalized adjustments.

[0004] Existing methods for analyzing human circadian rhythms utilize Fast Fourier Transform (FFT) to identify dominant periodicity in data and characterize daily rhythms. However, the accuracy of the Perfectly Matched Layer (PML) algorithm heavily relies on the quality and consistency of wearable device data, which can be affected by device wearing conditions and technical issues. Furthermore, the complexity of the PML algorithm may limit its application in real-time analysis or resource-constrained environments. Other analytical methods aim to leverage machine learning to integrate circadian rhythm models for a better understanding and prediction of changes in the human biological clock. Existing literature utilizes long-term continuous data collected from wearable devices to develop and apply various machine learning algorithms and statistical models to analyze and interpret this data, forming a detailed understanding of individual circadian rhythms. However, efficiently processing and analyzing this massive amount of data remains a challenge; the models presented in these studies are insufficient to establish long-term dependencies and cannot achieve long-term predictions.

[0005] As can be seen from the above, existing methods for analyzing human rhythms have significant limitations. They struggle to identify and record human behavior in real time and cannot be personalized based on individual behavioral data. This highlights the urgency and importance of developing a method that can record and analyze human rhythms accurately in real time, without relying on wearable devices. Summary of the Invention

[0006] The purpose of this invention is to overcome the above-mentioned shortcomings in the prior art and to provide a human rhythm analysis method, regulation method and analysis system based on behavior recognition.

[0007] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0008] The human rhythm analysis method based on behavior recognition includes the following steps:

[0009] Images of the activities of the individuals to be analyzed are collected to form an image sequence, and behavioral record data is obtained by identifying the image sequence.

[0010] Behavioral curves are obtained by processing behavioral record data;

[0011] Based on a neural network model that predicts the transition point between activity and rest states, the behavior curve is input into the neural network model for calculation, resulting in the activity-rest state transition point and the transition point time series curve.

[0012] In the technical solution of this invention, the state of the person to be analyzed is captured as activity images. These activity images are collected at different times and arranged into image sequences. Then, based on the image sequences, it is determined whether the person is active or resting, forming behavioral record data at different time points. This behavioral record data is then processed to obtain behavioral curves, which serve as the data basis for subsequent neural network model calculations. A neural network model suitable for predicting activity-rest transition points is employed. The input behavioral curves are calculated within the neural network model to predict the activity-rest transition points and their time-series curves, used to record the human rhythm of the person to be analyzed. This invention's method is based on a vision system to acquire images, performs behavioral recognition based on the images to generate behavioral record data, and transforms this data into behavioral curves that can be calculated using deep learning. By optimizing the data processing algorithm and then using a neural network model to predict the human rhythm of the person to be analyzed, efficiency is ensured when processing large amounts of behavioral record data and calculating activity-rest transition points. It also has good accuracy and solves the problem in existing technologies that rely on wearable devices and cannot identify and record human behavior in real time.

[0013] As a preferred embodiment of the present invention, the specific method for obtaining behavioral record data by recognizing image sequences is as follows:

[0014] S11. Real-time acquisition of activity images of the personnel to be analyzed to obtain image sequences;

[0015] S12. Assign identification codes (IDs) to the personnel to be analyzed based on the image sequence;

[0016] S13. For the personnel to be analyzed who have been assigned IDs, use a logistic regression model to determine whether the personnel to be analyzed are in a resting or active state at different time points.

[0017] S14. When it is determined that the person to be analyzed is in an active state, the activity behavior is identified and classified to obtain behavior record data.

[0018] As a more preferred embodiment of the present invention, the specific method for assigning an identity code ID to the person to be analyzed based on the image sequence is as follows: the image sequence is used to extract features using a convolutional network, the extracted feature map is passed through a region proposal network to generate candidate regions, the candidate regions are aligned with RoIs, and bounding boxes and category predictions are generated through a fully connected layer. A segmentation mask is generated through a fully convolutional layer, human features are extracted from the segmentation mask, and compared with the identity database to assign an identity code ID to the person to be analyzed.

[0019] As a more preferred embodiment of the present invention, in step S13, a logistic regression model is used to determine whether the person being analyzed is in a resting or active state, and the decision rule is as follows:

[0020]

[0021] In the formula Indicates the activity status. This represents the resting state, and τ is the decision threshold.

[0022] As a more preferred embodiment of the present invention, the activity behaviors include standing, sitting, and lying down. The activity behaviors are identified and classified based on a behavior recognition model based on visual-text joint embedding. The specific method is as follows: based on the image sequence of the person to be analyzed, it is converted into a visual feature vector by a visual encoder. The preset text of different activity behaviors is text-encoded. Based on the visual feature vector and the text encoding of different activity behaviors, the activity behavior classification decision is made to obtain the activity behavior category of the person to be analyzed.

[0023] As a preferred embodiment of the present invention, the specific method for obtaining behavior curves from behavior record data processing is as follows: extracting and preprocessing behavior record data, the preprocessing including moving average processing and smoothing processing, then extracting behavior features from the smoothed data, and drawing behavior curves based on the behavior features.

[0024] In the above technical solution, behavioral record data is transformed into behavioral curves through extraction and preprocessing. The transformed behavioral curves are easy to decompose and can more clearly identify different behavioral characteristics, providing a data foundation for predicting the activity-rest state transition point and the transition point time series curve.

[0025] As a preferred embodiment of the present invention, the behavioral characteristic includes activity intensity, which is the ratio of the frequency to the duration of the activity.

[0026] As a preferred embodiment of the present invention, the neural network model includes a decomposition layer, a CTCL Block module, and a KP Block module. The decomposition layer is used to decompose the behavior curve into a Trend part and a Season part. The CTCL Block module is used to process the Trend part of the behavior curve to predict the transition point between the activity and rest states. The KP Block module is used to fuse the Trend part and the Season part to predict the time series curve of the transition point.

[0027] As a preferred embodiment of the present invention, the CTCL Block module includes at least one KANLinear layer and a linear projection layer. The KANLinear layer can extract high-dimensional features of the input behavior curve to capture complex nonlinear relationships in the data, which helps to accurately identify the transition points of behavior. Its output will be passed through a linear projection layer to map the features to a low-dimensional space for predicting the transition points between active and resting states. The KANLinear layer includes a KAN layer and a linear layer to compensate for the limitations of the linear layer in processing complex time-series data.

[0028] As a more preferred embodiment of the present invention, the KP Block module includes at least two KANLinear layers, a CAT layer and an LSTM module. The KANLinear layer includes a KAN layer and a linear layer, the CAT layer is used for splicing, and the LSTM module serves as a gating mechanism.

[0029] As a preferred embodiment of the present invention, the activity-rest state transition point includes the rest-activity boundary time point CT0 and the activity-rest boundary time point CT12, and the transition point timing curve includes the CT0 timing curve and the CT12 timing curve.

[0030] Another aspect of the present invention provides a method for regulating human rhythm based on behavior recognition. The method includes: using the above-mentioned human rhythm analysis method based on behavior recognition to predict the activity-rest state transition point and the transition point timing curve of the person to be analyzed; setting a standard activity-rest state transition point; comparing the activity-rest state transition point with the standard activity-rest state transition point; and designing a lighting control scheme for regulating the human rhythm of the person to be analyzed based on the comparison result, so that the activity-rest state transition point reaches the standard activity-rest state transition point.

[0031] In the above technical solution, the human body rhythm is predicted in real time based on behavior recognition without relying on any wearable devices. The lighting control scheme can be designed and output according to the changes in the rhythm of the person being analyzed, so as to achieve real-time detection and real-time adjustment. This eliminates the inconvenience of the person being analyzed having to wear the device continuously, provides a more natural and interference-free user experience, and can automatically adjust according to the individual's rhythm changes, thereby improving adjustment efficiency and quality of life.

[0032] The present invention also provides a human rhythm analysis system based on behavior recognition, the system comprising:

[0033] The recognition module is used to collect images of the activities of the person to be analyzed to form an image sequence, and to identify the behavior record data by recognizing the image sequence.

[0034] The processing module is used to process behavior record data to obtain behavior curves;

[0035] The prediction module is used to calculate the activity-rest transition point and its time-series curve based on the neural network model that predicts the activity-rest transition point.

[0036] The present invention also provides an electronic device, including at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described behavior recognition-based human rhythm analysis method.

[0037] The present invention also provides a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the above-described behavior recognition-based human rhythm analysis method.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] 1. This invention provides a method for analyzing human rhythms based on behavior recognition. It acquires images based on a vision system, performs behavior recognition on the images to generate behavior recording data, and transforms the behavior recording data into behavior curves that can be calculated using deep learning. By optimizing the data processing algorithm, the human rhythm of the person to be analyzed is obtained by using a neural network model for prediction. This ensures efficiency in processing large amounts of behavior recording data and calculating the transition points between activity and rest states, and has good accuracy.

[0040] 2. The neural network model in the analysis method of this invention includes a decomposition layer, a CTCL Block module, and a KP Block module. The CTCL Block module is used to process the Trend part of the behavior curve to predict the transition point between the activity and rest states. The KP Block module is used to fuse the Trend part and the Season part to predict the time series curve of the transition point. The CTCL Block module includes at least two KANLinear layers and a linear projection layer. The KANLinear layer can extract high-dimensional features of the input behavior curve and capture complex nonlinear relationships in the data, which helps to accurately identify the transition point of the behavior. The KP Block module includes at least two KANLinear layers, a CAT layer, and an LSTM module. The LSTM module, as a gating mechanism, captures complex behavior patterns to achieve prediction of the time series. The KP Block module generates more accurate time series prediction results through expert weighted selection. The model can capture complex time series patterns and, through multi-scale processing, can handle multi-level features, thereby improving the model's prediction accuracy.

[0041] 3. This invention provides a human rhythm regulation method based on behavior recognition. It predicts human rhythms in real time based on behavior recognition, without relying on any wearable devices, thus avoiding interference with daily activities. It can perform personalized lighting control according to the changes in the rhythm of the person being analyzed, achieving real-time detection and regulation, providing a more natural and interference-free user experience, thereby improving regulation efficiency and quality of life.

[0042] 4. This invention is applicable to the analysis of human circadian rhythms of personnel in enclosed spaces (such as space stations). It analyzes human circadian rhythms based on visual behavior recognition and adjusts lighting according to human circadian rhythms. It uses a pure vision system to accurately adjust the physiological rhythms of the human body in enclosed spaces, thereby improving their health and comfort. Attached Figure Description

[0043] Figure 1 This is a flowchart of the human rhythm analysis method based on behavior recognition according to the present invention;

[0044] Figure 2 Behavioral curves plotted based on activity intensity;

[0045] Figure 3 The behavior curve consists of the Season and Trend components;

[0046] Figure 4 The behavior curve is broken down into Season and Trend curves.

[0047] Figure 5 This is a structural diagram of the neural network model for predicting the activity-rest transition point of this invention.

[0048] Figure 6 This is a structural diagram of the KANLinear layer. Detailed Implementation

[0049] To more clearly describe the inventive objectives, technical solutions, and advantages of the specific embodiments of this invention, the solutions in the specific embodiments will be described in detail below with reference to the accompanying drawings. The specific technical solutions involved in the following embodiments are merely for the purpose of clearly and completely describing the innovative technical solutions of this invention. They are only a part of the specific implementation methods that this invention can adopt, not all embodiments, and should not be construed as limiting the innovative solutions of this invention. Any solution that adopts the same inventive concept as this invention should be included within the protection scope of this invention.

[0050] Secondly, the descriptions in the accompanying drawings of the specific embodiments of this invention are merely for the convenience of those skilled in the art to understand the invention. The details shown in the drawings are for the purpose of clearly presenting the technical solution, and should not be construed as including all technical features in the drawings in the specific implementation examples, nor should the details in the drawings be considered as additional limitations on the innovative technical solution of this invention. The components in the various embodiments described and shown in the drawings can be combined and arranged in different configurations. These variations in combination and arrangement should be considered as part of all embodiments of the innovative solution of this invention and included within the scope of protection of this invention.

[0051] In summary, the solutions or descriptions presented in the specific embodiments and accompanying drawings of this invention are not intended to limit the scope of protection claimed, but merely to illustrate selected embodiments / examples to help those skilled in the art understand the relevant innovative solutions. All other equivalent or parallel embodiments obtained by those skilled in the art based on these embodiments without inventive effort are within the scope of protection claimed by this invention.

[0052] Example 1

[0053] This embodiment provides a method for human rhythm analysis based on behavior recognition, such as... Figure 1 As shown, it includes the following steps:

[0054] S1. Collect activity images of the personnel to be analyzed to form an image sequence, and identify the image sequence to obtain behavioral record data;

[0055] The specific method for obtaining behavioral record data by recognizing image sequences is as follows:

[0056] S11. The image acquisition mechanism collects images of the personnel to be analyzed in real time to obtain an image sequence. ;

[0057] S12. Assign identification codes (IDs) to the personnel to be analyzed based on the image sequence;

[0058] S13. For the personnel to be analyzed who have been assigned IDs, use a logistic regression model to determine whether the personnel to be analyzed are in a resting or active state at different time points.

[0059] S14. When it is determined that the person to be analyzed is in an active state, the activity behavior is identified and classified to obtain behavior record data.

[0060] In step S12, the specific method for assigning identification codes (IDs) to the personnel to be analyzed based on the image sequence is as follows: Feature extraction is performed on the image sequence using a convolutional network (CNN). The extracted feature maps are then processed through a region proposal network (RPN) to generate candidate regions. RoI (Region of Interest) alignment is performed on the candidate regions, and bounding boxes and class predictions are generated through fully connected layers. Finally, a segmentation mask is generated using a fully convolutional network (FCN). Human features are extracted from the segmentation mask and compared with the identity database to assign an identity identification code (ID) to the person to be analyzed.

[0061] The formula for generating bounding boxes and class predictions using fully connected layers is as follows:

[0062]

[0063] in It is a boundary prediction for time t. This refers to the category prediction at time t, where RoIAlign represents the region alignment operation. This indicates that candidate regions are generated at time t. This is the extracted feature map.

[0064] The formula for generating the segmentation mask is:

[0065]

[0066] in For segmentation mask, FCN represents the fully convolutional operation.

[0067] Extracting human features from a segmentation mask refers to extracting identity information from the segmentation mask. The formula used is:

[0068]

[0069] in This represents the activation function, typically ReLU. For linear layer weights, This is the linear layer bias term.

[0070] The extracted identity information is compared with the identity information in a pre-established identity database. A unique ID is assigned to each matching individual to be analyzed, using the following formula:

[0071]

[0072] in This represents the unique identity ID assigned at time t. Cosine similarity (F) refers to the value of the independent variable that makes a function reach its maximum value within a specific set or interval. j This represents the feature vector of the j-th identity in the identity database.

[0073] In step S13, a logistic regression model is used to determine whether the person being analyzed is in a resting or active state. The decision rule is as follows:

[0074]

[0075] In the formula Indicates the activity status. This represents the resting state, and τ is the decision threshold. The calculation formula is:

[0076]

[0077] Where θ and θ0 are the weights and biases of the logistic regression, and T represents the transpose. This represents the Euclidean distance variation between the bounding boxes of two consecutive frames. The calculation formula is:

[0078]

[0079] To extract features from image sequences using a convolutional network (CNN), the extracted feature maps are passed through a region proposal network (RPN) to generate candidate regions. The candidate regions are then aligned with RoIs and bounding boxes are generated through a fully connected layer.

[0080] In this embodiment, τ is set to 0.5. The logistic regression model outputs a probability, which represents the probability that the person being analyzed is in a resting or active state given the input features. The results are analyzed to determine whether the person being analyzed is in a resting or active state at different time points.

[0081] In step S14, the activity behaviors include standing, sitting, and lying down. When it is determined that the person being analyzed is in an active state, it is necessary to determine which specific activity behavior it is.

[0082] This embodiment uses a behavior recognition model based on visual-text joint embedding to classify the activity behavior of the person to be analyzed. The specific method is as follows: based on the image sequence of the person to be analyzed, it is converted into a visual feature vector by a visual encoder. The text of different pre-set activity behaviors is encoded. Based on the visual feature vector and the text encoding of different activity behaviors, the activity behavior classification decision is made to obtain the activity behavior category of the person to be analyzed.

[0083] Specifically, using the image sequences of the person to be analyzed. It is transformed into a visual feature vector by a visual encoder:

[0084]

[0085] The pre-defined text categories for different activity behaviors—"standing", "sitting", and "laying"—are then encoded.

[0086]

[0087] Activity behavior classification decision is made after calculating cosine similarity:

[0088]

[0089] Select the text feature vector T with the highest similarity, i.e., the highest confidence. k The corresponding activity behavior category is used as the identification result.

[0090] S2. Process the behavior record data to obtain the behavior curve; specifically: extract and preprocess the behavior record data, including moving average processing and smoothing processing, and then extract the behavior features from the smoothed data, including activity intensity, and draw the behavior curve based on the behavior features.

[0091] Let the dataset of behavioral record data be... ,in t i Represents a timestamp. b i Indicates time t i The behavior category is determined as either rest or activity, with further subcategories for activity behaviors. Sampling is performed at appropriate time intervals.

[0092]

[0093] The sampled data is processed using a moving average to generate a smooth curve for observation.

[0094]

[0095] Where k is the size of the moving window.

[0096] Behavioral features, such as activity intensity, are extracted from the smoothed data. Activity intensity is the ratio of activity frequency to duration, or the number of behavioral changes per unit time. The calculation formula is as follows:

[0097]

[0098] Based on the activity intensity, the following is obtained: Figure 2 The behavior curve shown can be analyzed to reveal that it consists of a Season component, a Trend component, and random noise. Figure 3 As shown, the Season section reflects the periodic fluctuations in behavior, which are usually related to the body's natural rhythms (such as the circadian rhythm). The periodic peaks and troughs reflect the body's activity and rest states at different times of the day. This part is usually regular and the fluctuation amplitude is relatively stable. The Trend section shows the long-term trend of the behavior curve. This part can reflect the overall trend of the behavior pattern, such as the phenomenon that activity gradually decreases or increases over time in a day. The trend part is usually relatively smooth.

[0099] To better process this data, it is broken down as follows: Figure 4 As shown, since the Season component reflects cyclical fluctuations, it possesses stable and repeatable characteristics. Cyclical data is relatively predictable within a certain range because these fluctuations recur. Therefore, using the Season component as input for time series forecasting can effectively capture future cyclical fluctuations. The Trend component, on the other hand, represents the smoothed long-term changes in the behavioral curve, which more clearly demonstrates the overall direction of behavioral change and important turning points. Therefore, using the rate of change of the Trend component (i.e., the derivative of the curve) to locate the CT0 point can accurately identify the boundary between resting and active states of behavior.

[0100] The fusion of periodicity and trend is illustrated in the figure. The Season component reflects periodic fluctuations, exhibiting stable and repeatable characteristics, while the Trend component displays long-term behavioral trends. Combining the two allows for a more comprehensive prediction of future behavior. Therefore, modeling the Season and Trend components separately and then fusing their outputs captures the combined influence of periodicity and long-term trends, thereby improving the accuracy of time series forecasting. Processing the Season and Trend components separately allows for a clearer identification of different behavioral characteristics (periodicity and trend), and using them separately to predict turning points. This approach enables the model to better handle complex behavioral data and achieve more accurate results in practical applications.

[0101] S3. Based on the neural network model that predicts the transition point between activity and rest states, the behavior curve is input into the neural network model for calculation to obtain the transition point between activity and rest states and the transition point time series curve.

[0102] The neural network model includes a decomposition layer, a CTCL Block module, and a KP Block module. The decomposition layer is used to decompose the input behavior curve into a Trend part and a Season part. The CTCL Block module is used to process the Trend part of the behavior curve to predict the transition point between the active and resting states. The KP Block module is used to fuse the Trend part and the Season part to predict the time series curve of the transition point.

[0103] The CTCL Block module includes at least one KANLinear (KAL) layer and a linear projection layer. The KANLinear layer can extract high-dimensional features of the behavior curve and capture complex nonlinear relationships in the data, which helps to accurately identify the transition points of the behavior. Its output will be passed through a linear projection layer to map the features to a low-dimensional space for predicting CT0 or CT12.

[0104] In this embodiment, the KANLinear layer is a novel structure. The KANLinear layer includes a KAN layer (Kolmogorov-Arnold Networks) and a linear layer, designed to overcome the limitations of linear layers when processing complex time-series data. For example... Figure 6 As shown, the final output of KANLinear is a weighted sum of the linear and B-spline interpolation parts:

[0105]

[0106] In the formula, α represents the weight of the linear part, and β represents the weight of the spline part. Output the results for the linear part. Output the results for the B-spline interpolation part.

[0107] The calculation of the base linear mapping in the KANLinear layer is as follows:

[0108]

[0109] In the formula For activation function, The weights are linear, and X is the input. This is a linear bias term.

[0110] The calculation of the B-spline interpolation part in the KANLinear layer is as follows:

[0111]

[0112] In the formula For spline functions, As the weights corresponding to the spline function, the B-spline basis function calculates the interpolation of the input in different intervals using the control points grid and the input x:

[0113]

[0114] Where k is the spline order. It is a B-spline basis function, and its calculation results are based on the input and the grid.

[0115] Let the input data in the CTCL Block module be represented as ,in This represents the data at each input time point. Feature extraction is implemented using at least one KANLinear layer. Each layer can be represented as a non-linear mapping that extracts features from the input data. Let the input features of the i-th layer be... The output features are ,as follows:

[0116]

[0117] in It is the linear transformation part. This is the B-spline interpolation part, which needs to be explained. , yes and The output is used to capture non-linear relationships. A KAN layer can be represented as performing a non-linear transformation on each layer until the desired features are finally extracted:

[0118]

[0119] After extracting features through at least one KANLinear layer, the features are mapped to a low-dimensional space using the output of a linear projection layer. The calculation formula is as follows:

[0120]

[0121] In the formula, W and b are the weights and biases of the linear projection layer. The detection of the CT0 point is essentially a regression task. Therefore, the results of the linear projection layer can be used to directly predict the index or time value of the activity-rest transition point in the behavior curve, thereby obtaining the CT0 point or CT12 point.

[0122] The KP Block module includes at least two KANLinear layers, a CAT layer, and an LSTM module. As mentioned above, the KANLinear layer includes a KAN layer and a linear layer. After at least two KANLinear layers, a CAT layer is used to stitch the results together. The LSTM module serves as a gating mechanism. In some embodiments, a linear projection layer is provided after the CAT layer, such as... Figure 5As shown in the KP Block section, the KP Block module merges the Trend and Season parts, and combines multiple network components by using LSTM gates, KANLinearexpert and linear projection layers to capture complex behavioral patterns to achieve time series prediction.

[0123] This embodiment innovatively employs LSTM as a gating mechanism. LSTM is highly effective in processing time series data because it can capture long-term dependencies within the data. Here, the LSTM takes trend and seasonal characteristics as input and generates a gating output based on dependencies at time steps. gate_out To adjust the final prediction results:

[0124]

[0125] in These represent the encoding vector and the hidden state of the LSTM, respectively, and the gated output is converted into selection weights using softmax:

[0126]

[0127] The score determines which network has the greatest impact on the final prediction. The weighted combination of KANLinear experts is:

[0128]

[0129]

[0130] Final prediction:

[0131]

[0132] The linear layer here maps the predicted multidimensional features to the time series prediction output.

[0133] The prediction mechanism of the KP Block module combines LSTM gating, KANLinear and linear projection layers. It generates more accurate time series prediction results through expert weighted selection. The model can capture complex time series patterns and process multi-level features through multi-scale processing, thereby improving prediction accuracy.

[0134] The activity-rest transition point includes the rest-activity boundary time point CT0 and the activity-rest boundary time point CT12. The transition point time series curves include the CT0 time series curve and the CT12 time series curve. In this embodiment, the behavior curve neural network model is used for prediction, and the CT0 and CT12 time series curves of the person to be analyzed based on the real-time image sequence are output.

[0135] Example 2

[0136] This embodiment provides a human rhythm regulation method based on behavior recognition. The regulation method includes: using the human rhythm analysis method based on behavior recognition in Embodiment 1, predicting the activity-rest state transition point and the transition point time series curve of the person to be analyzed, setting a standard activity-rest state transition point, comparing the activity-rest state transition point with the standard activity-rest state transition point, and designing a lighting control scheme for regulating the human rhythm of the person to be analyzed based on the comparison result, so that the activity-rest state transition point reaches the standard activity-rest state transition point.

[0137] By detecting and analyzing the behavioral data of the individuals being analyzed, the CT0 point of the real-time image sequence is predicted and compared with a standard CT0 point. Based on different results (early, delayed, or identical), corresponding lighting control schemes are implemented to optimize the working and sleeping environment of the individuals being analyzed and improve their quality of life. Specifically, a standard CT0 point needs to be preset first, i.e., the desired CT0 point for the individuals being analyzed. Using the CT0 obtained from the analysis and Comparison:

[0138]

[0139] according to As a result, a corresponding lighting control scheme was adopted.

[0140] when That is, the CT0 point of the person to be analyzed and When the brightness is the same or very similar, it indicates that the user's circadian rhythm is normal and the current sleep-wake cycle needs to be maintained. From CT0 to CT12, white light with a cosine variation ranging from 600 lumens to 2000 lumens is applied, gradually increasing the brightness to help the user stay awake. From CT12 to CT24, white light with a cosine variation ranging from 600 lumens to 0 lumens is applied, gradually decreasing the brightness to help the user fall asleep and maintain a normal circadian rhythm. In this invention, the time difference between CT0 and CT12 is defined as 12 hours.

[0141] when In other words, if the CT0 point of the data is advanced, it indicates that the user's circadian rhythm is earlier than usual, and it needs to be gradually delayed through light regulation to achieve the desired effect. One hour before CT0, blue light with a cosine variation ranging from 800 to 1200 lumens is used to gradually wake the user and delay their biological clock. From CT0 to CT12, white light with a cosine variation ranging from 600 to 2000 lumens is used to simulate the changes in natural sunlight and help the user stay awake. From CT12 to CT23 (CT12 plus 11 hours), red light with a cosine variation ranging from 600 to 0 lumens is used to reduce the suppression of melatonin and help the user gradually fall asleep.

[0142] when If the CT0 point of the data analysis is delayed, it indicates that the user's circadian rhythm is late, and it needs to be gradually advanced through light regulation to achieve the desired effect. From CT0 to CT12, white light with a cosine variation ranging from 600 lumens to 2000 lumens is used to gradually wake the user and advance their biological clock. From CT12 to CT13, yellow light with a cosine variation ranging from 600 lumens to 1500 lumens is used to simulate the gradual warming of sunlight and reduce its impact on wakefulness. From CT13 to CT15, white light with a cosine variation ranging from 400 lumens to 0 lumens is used, and from CT15 to CT24, white light with a cosine variation ranging from 60 lumens to 0 lumens is used to simulate the gradual weakening of natural sunlight and help the user gradually fall asleep.

[0143] By employing behavior recognition technology based on a computer vision system, this method adaptively records and analyzes the behavior curves of the person being analyzed, and calculates the activity-rest transition points (CT0, CT12). Simultaneously, it establishes the recursive feasibility and stability of rhythm regulation, thereby optimizing the lighting control strategy and achieving precise regulation of the biorhythms of the person being analyzed. This method significantly reduces human intervention, improves regulation efficiency, saves energy resources, ensures control performance, and reduces the risk of rhythm disturbance. The recursive feasibility and system stability under the constraints of the person's activity state and lighting control input are demonstrated, and numerical simulations and experiments prove the effectiveness and advancement of the proposed method.

[0144] Example 3

[0145] This embodiment provides a human rhythm analysis system based on behavior recognition. The system includes:

[0146] The recognition module is used to collect images of the activities of the person to be analyzed to form an image sequence, and to identify the behavior record data by recognizing the image sequence.

[0147] The processing module is used to process behavior record data to obtain behavior curves;

[0148] The prediction module is used to calculate the activity-rest transition point and its time-series curve based on the neural network model that predicts the activity-rest transition point.

[0149] The systems or modules described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware, or a module that performs the same function can be implemented by a combination of multiple sub-modules, etc.

[0150] This embodiment also provides an electronic device, including at least one processor, a memory communicatively connected to the at least one processor, and at least one input / output interface communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the behavior recognition-based human rhythm analysis method of the foregoing embodiments. The input / output interface may include a display, keyboard, mouse, and USB interface for inputting and outputting data.

[0151] The electronic device can be an electronic device for the client, such as a mobile phone, laptop, tablet, desktop computer, etc., to execute the behavior recognition-based human rhythm analysis method of Embodiment 1.

[0152] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0153] When the integrated units of this invention are implemented as software functional units and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0154] This embodiment also provides a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the above-described behavior recognition-based human rhythm analysis method.

[0155] A computer-readable storage medium can be a tangible device that holds and stores instructions for use by an instruction execution device. A computer-readable storage medium can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any combination thereof.

[0156] For those skilled in the art, when understanding the solutions described in the specific embodiments of the present invention, conventional technical manuals in the field can be consulted. At the same time, appropriate understandings or adjustments can be made to the above-mentioned terms to deduce the same or similar technical solutions without creative effort.

[0157] The above embodiments describe only the basic principles, main features and / or advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and the description of the invention content in the specification are only the principles or specific cases of the present invention. Without departing from the essence of the innovative idea of ​​the present invention, there are various changes and improvements to the innovative solution of the present invention, and all such changes and improvements fall within the scope of protection claimed by the present invention.

Claims

1. A method for analyzing human rhythms based on behavior recognition, characterized in that, Includes the following steps: Images of the activities of the individuals to be analyzed are collected to form an image sequence, and behavioral record data is obtained by identifying the image sequence. Behavioral curves are obtained by processing behavioral record data; Based on the neural network model that predicts the transition point between activity and rest states, the behavior curve is input into the neural network model for calculation, and the transition point between activity and rest states and the transition point time series curve are obtained. The neural network model includes a decomposition layer, a CTCL Block module, and a KP Block module. The decomposition layer is used to decompose the behavior curve into a Trend part and a Season part. The CTCL Block module is used to process the Trend part of the behavior curve to predict the transition point between the active and resting states. The CTCL Block module includes at least one KANLinear layer and a linear projection layer. The KP Block module is used to fuse the Trend part and the Season part to predict the time series curve of the transition point. The KP Block module includes at least two KANLinear layers, a CAT layer, and an LSTM module.

2. The method for human rhythm analysis based on behavior recognition according to claim 1, characterized in that, The specific method for obtaining behavioral record data by recognizing image sequences is as follows: S11. Real-time acquisition of activity images of the personnel to be analyzed to obtain image sequences; S12. Assign identification codes (IDs) to the personnel to be analyzed based on the image sequence; S13. Use logistic regression model to determine whether the person being analyzed is in a resting or active state at different time points. S14. When it is determined that the person to be analyzed is in an active state, the activity behavior is identified and classified to obtain behavior record data.

3. The method for human rhythm analysis based on behavior recognition according to claim 1, characterized in that, The specific method for obtaining behavior curves from behavior record data is as follows: extract and preprocess the behavior record data, including moving average processing and smoothing processing, then extract behavioral features from the smoothed data, and draw behavior curves based on the behavioral features.

4. The human rhythm analysis method based on behavior recognition according to claim 3, characterized in that, Behavioral characteristics include activity intensity, which is the ratio of the frequency to the duration of the activity.

5. The method for human rhythm analysis based on behavior recognition according to claim 1, characterized in that, KANLinear layers can extract high-dimensional features of behavioral curves; The KANLinear layer consists of a KAN layer and a linear layer.

6. The human rhythm analysis method based on behavior recognition according to claim 1, characterized in that, The KANLinear layer consists of a KAN layer and a linear layer, with a CAT layer used for splicing and an LSTM module used as a gating mechanism.

7. A method for regulating human rhythms based on behavior recognition, characterized in that, The adjustment method includes: Using the human rhythm analysis method based on behavior recognition as described in any one of claims 1-6, the activity-rest state transition point and the transition point time series curve of the person to be analyzed are predicted; Set a standard activity-rest transition point; The activity-rest transition point is compared with the standard activity-rest transition point. Based on the comparison results, a lighting control scheme for the human rhythm regulation of the person to be analyzed is designed so that the activity-rest transition point reaches the standard activity-rest transition point.

8. A human rhythm analysis system based on behavior recognition, characterized in that, The system is used to execute the human rhythm analysis method based on behavior recognition as described in any one of claims 1-6, and the system comprises: The recognition module is used to collect images of the activities of the person to be analyzed to form an image sequence, and to identify the behavior record data by recognizing the image sequence. The processing module is used to process behavior record data to obtain behavior curves; The prediction module is used to calculate the activity-rest transition point and its time-series curve based on the neural network model that predicts the activity-rest transition point.

9. An electronic device comprising at least one processor and a memory communicatively connected to said at least one processor; said memory storing instructions executable by said at least one processor, characterized in that, The instructions are executed by the at least one processor to enable the at least one processor to perform the behavior recognition-based human rhythm analysis method according to any one of claims 1-6.

Images