Interaction method, system and equipment based on millimeter wave radar figure behaviors

Abstract

The invention relates to an interaction method, system and device based on millimeter wave radar character behaviors, and belongs to the technical field of exhibition halls, the interaction method comprises the following steps: obtaining original point cloud data of each target character from a millimeter wave radar in a target area, and converting the original point cloud data into a target state vector under a unified coordinate system; constructing a dynamic graph structure by taking the target person as a node and taking the relative distance and the motion trend correlation between the target persons calculated based on the target state vector as edges, and extracting a graph index of the dynamic graph structure; according to the graph index and the target state vector corresponding to each target person, identifying a current target scene; and according to the current target scene and the historical interaction vector of each target person in the target area, generating an interaction instruction of the target area, and distributing the interaction instruction to corresponding interaction equipment. The method and the device have the beneficial effects of improving the interaction effect and the user experience.

Description

Technical Field

[0001] This application relates to the technical field of exhibition halls, and in particular to an interactive method, system, and device based on millimeter-wave radar for human behavior. Background Technology

[0002] In modern exhibition halls, providing visitors with personalized and intelligent interactive experiences is key to enhancing the hall's appeal and service quality. With the continuous development of technology, people's demands for exhibition halls are no longer limited to simple exhibit displays; they hope to engage in deeper and more interesting interactions with the exhibits and the environment.

[0003] Currently, common interactive methods in exhibition halls mainly rely on traditional sensors and fixed interactive programs. For example, some exhibition halls use infrared sensors to detect the presence and approximate location of people, triggering preset audio explanations or lighting effects when a person approaches a specific exhibit. Other exhibition halls use touchscreens, buttons, and other devices, allowing visitors to actively select the exhibit information they want to learn about. In addition, there are methods that utilize cameras for people behavior analysis, but this method is easily affected by factors such as lighting and obstructions.

[0004] Whether using cameras or infrared sensors, the interactive methods are relatively simple and fixed, making it difficult to provide personalized interactive experiences based on the behavior and needs of different visitors. Moreover, the preset interactive programs lack flexibility and cannot be dynamically adjusted according to the real-time distribution and behavioral changes of people in the exhibition hall, resulting in limited interactive effects and failing to fully meet the diverse needs of visitors. Summary of the Invention

[0005] To improve the interactive effect and user experience, this application provides an interactive method, system and device based on millimeter-wave radar human behavior.

[0006] Firstly, this application provides an interactive method based on millimeter-wave radar character behavior, employing the following technical solution: An interactive method based on millimeter-wave radar-based human behavior includes: The original point cloud data of each target person is obtained from the millimeter-wave radar of the target area, and the original point cloud data is converted into a target state vector in a unified coordinate system. The target state vector includes at least position and velocity. Using the target person as the node and the relative distance and movement trend correlation between the target persons calculated based on the target state vector as the edge, a dynamic graph structure is constructed, and graph indicators of the dynamic graph structure are extracted. The graph indicators include at least one of node connectivity components, clustering coefficient, and degree centrality. Based on the graph indicators and the target state vectors corresponding to each target person, identify the current target scenario; Based on the current target scene and the historical interaction vectors of each target character in the target area, an interaction command for the target area is generated and dispatched to the corresponding interactive device.

[0007] By employing the aforementioned technical solution, a graph structure is constructed using target individuals as nodes and relationships between them as edges. Graph metrics are then calculated to intuitively represent the relationships and interaction patterns between target individuals. Through the graph structure and metrics, potential connections and behavioral patterns between individuals can be uncovered, providing deeper information for accurately identifying target scenarios. Identifying the current target scenario based on graph metrics and target state vectors, combining information about relationships between individuals and individual state information, allows for a more accurate assessment of the actual situation within the target area. Interactive commands are generated based on the current target scenario and the historical interaction vectors of each target individual in the target area. This fully considers the real-time situation of the target area and the historical behavioral patterns of the individuals, making the generated interactive commands more targeted and personalized, better meeting the needs of different individuals and scenarios, and improving the effectiveness of the interaction and the user experience.

[0008] Optionally, the step of identifying the current target scenario based on the graph indicators and the target state vectors corresponding to each target person includes: The graph indicators are spatiotemporally aligned and feature concatenated with the target state vectors corresponding to each target person to construct a scenario feature matrix. The scenario feature matrix is ​​input into a pre-trained scenario classification model, which outputs the category label of the current target scenario. The scenario confidence is calculated and verified based on the probability distribution or confidence score output by the scenario classification model.

[0009] By employing the above technical solution, the graph index reflects information such as the relationship structure between target individuals, while the target state vector contains individual state information for each target individual, such as position and speed. Combining these two through multimodal feature fusion fully leverages the complementarity of different modalities, resulting in a more comprehensive and richer information matrix for the constructed scenario feature matrix, thus more accurately depicting the current target scenario. Inputting the scenario feature matrix into a pre-trained scenario classification model allows for the rapid and accurate output of the category label for the current target scenario; then, the confidence score reflects the model's certainty regarding the output category label, helping users assess the reliability of the recognition results.

[0010] Optionally, the step of generating interaction instructions for the target area based on the current target scene and the historical interaction vectors of each target character in the target area includes: The category labels and context confidence of the current target scene are time-series embedded and feature-concatenated with the historical interaction vectors of each target person to construct a dynamic response feature matrix; The dynamic response feature matrix is ​​input into a pre-trained instruction generation model, which outputs a set of candidate interaction instruction sequences. The generated candidate instruction sequence is evaluated and filtered to output the optimized interactive instruction; Behavioral pattern matching is performed on the dynamic response feature matrix to retrieve interactive instruction templates under similar historical scenarios; Based on the search results, optimized interactive instructions, and a pre-defined interactive strategy rule base, personalized interactive instruction parameters are generated through parameter interpolation or rule reasoning.

[0011] By adopting the above technical solution, the category label clarifies the type of the current scene, the context confidence reflects the reliability of scene recognition, and the historical interaction vector records the target person's past behavioral patterns within the target area. By fusing the category label and context confidence of the current target scene with the historical interaction vectors of each target person in a spatiotemporal manner, real-time scene information can be combined with the target person's past interaction history. During training, the instruction generation model learns the potential relationship between different scene features and interaction instructions. It can mine potentially applicable interaction instructions based on the input dynamic response feature matrix, and through evaluation and filtering checks, it can ensure that the output instructions have high reliability. In similar scenes, historically effective interaction instructions may also be applicable in the current scene. By referencing these historical instruction templates, suitable instruction frameworks can be quickly found, reducing the difficulty and workload of instruction generation while improving the effectiveness and success rate of instructions. Combining the retrieved historical instruction templates, optimized interaction instructions, and a preset interaction strategy rule base, personalized interaction instruction parameters can be generated according to the specific situation of the target person and the characteristics of the current scene, providing a tailored interactive experience for each target person and improving user satisfaction and engagement.

[0012] Optionally, the steps following the verification of scenario confidence include: When the scenario confidence level is lower than the scenario threshold, the scenario complexity index is monitored in real time from the millimeter-wave radar of the target area; The adjustment vector of the radar scanning parameters is calculated based on the scene complexity index, the category label of the current target scene, and the scene threshold. Based on the adjustment vector, the scanning frequency, resolution, or beamwidth of the millimeter-wave radar are dynamically adjusted. After adjustment, the quality indicators of the point cloud data were reacquired and verified to ensure that the target state vector error under the unified coordinate system was lower than the preset threshold. After adjustment, record the parameter changes before and after adjustment in the history log.

[0013] By adopting the above technical solution, when the scenario confidence level is lower than the scenario threshold, it means that the reliability of the currently acquired scenario information is low and may not accurately reflect the actual situation of the target area. In this case, calculating the adjustment vector of radar scanning parameters based on the monitored scenario complexity index, the current target scenario category label, and the scenario threshold enables the system to adaptively cope with different complex scenarios and low-confidence situations. This not only enhances the accuracy of the target state vector, providing a more reliable foundation for scenario recognition, but also allows for continuous system evolution through log recording feedback to model training, improving overall energy efficiency and real-time performance.

[0014] Optionally, the step of evaluating and filtering the generated candidate instruction sequence and outputting optimized interactive instructions includes: Spatiotemporal conflict detection is performed on the candidate instruction sequence to identify resource competition or behavioral logic contradictions between instructions; Based on the real-time location distribution and motion trajectory prediction of the target person, the spatial feasibility score for command execution is calculated; By integrating contextual confidence, historical interaction vector similarity, and device status parameters, a dynamic decay factor is generated. Candidate instructions are ranked using a weighted formula, which is: Optimization Score = Scenario Confidence × Spatial Feasibility Score / Dynamic Decay Factor; Select instructions whose optimization scores are higher than the score threshold, inject security constraints, and output optimization instructions.

[0015] By adopting the above technical solutions, the conflict resolution mechanism can avoid device resource contention or behavioral logic conflicts caused by multiple concurrent commands; the spatial feasibility score combined with real-time positioning data from millimeter-wave radar ensures that commands are executable in the physical scenario; the dynamic attenuation factor introduces deviations in device status and historical behavior to reduce the risk of false triggering in abnormal scenarios; and the weighted optimization model achieves quantitative screening, improves the priority of high-confidence and high-security commands, and ensures smooth interaction.

[0016] Optionally, the steps following the dispatch of the interactive command include: The physiological response indicators of the target person are monitored in real time, and an emotional feedback vector is constructed by combining the preset emotional mapping rules. The physiological response indicators are extracted based on the micro-motion features of millimeter-wave radar. Compare the deviation between the emotional feedback vector and the expected interaction effect; If the deviation value exceeds the deviation threshold, an emergency intervention protocol is triggered; Once the emergency protocol is activated, the current interactive command flow is frozen, and the system switches to the preset security scenario mode. Generate intervention logs and associate them with the historical interaction vectors of each target individual.

[0017] By employing the aforementioned technical solution and constructing an emotional feedback vector, these physiological indicators can be integrated into a comprehensive and quantifiable emotional expression, enabling the system to more accurately perceive the target person's emotional response to interactive commands. By comparing the deviation value between the emotional feedback vector and the expected interaction effect, the interaction effect can be quantitatively evaluated. The deviation value intuitively reflects the gap between the actual interaction effect and the expected effect, allowing the system to clearly understand whether the interactive command has achieved its intended goal. When the deviation value exceeds the deviation threshold, triggering an emergency intervention protocol ensures the safety and health of the target person. During the interaction, if the target person's emotional reaction is too strong or abnormal, it may cause physical and psychological harm. Activating the emergency intervention protocol can promptly interrupt potentially risky interactive commands, protecting the target person from potential harm. Generating an intervention log can record the emergency intervention process and related information in detail, providing an important basis for subsequent problem analysis and summarization.

[0018] Optionally, the interaction method further includes: Collect interaction feedback data from each target person, and calculate the interaction effect index based on the interaction feedback data; The interaction effect indicators, the distributed interaction instruction parameters, the current target scene category label and the scenario confidence are feature-encoded and fused to construct an effect evaluation feature vector; The effect evaluation feature vector is input into the pre-trained effect evaluation model, which outputs an interaction effect score. Based on the interaction effect score, the historical interaction vectors of each target character are dynamically updated, the parameters of the scenario classification model and the instruction generation model are optimized, and the interaction strategy rule base is iteratively adjusted.

[0019] By adopting the above technical solutions, and collecting interaction feedback data from various target individuals and calculating interaction effect metrics, abstract interaction effects can be transformed into concrete quantitative indicators. This provides a direct reflection of the execution effect of interaction commands and enables timely detection of problems that arise during the interaction process. Updating historical interaction vectors allows the system to better track changes in the behavior of target individuals, providing a more accurate basis for subsequent interaction command generation. The interaction effect score is an intuitive indicator that allows system administrators and relevant personnel to quickly understand the execution effect of interaction commands. Optimizing the parameters of the scenario classification model and command generation model based on the interaction effect score can improve the accuracy and adaptability of the models, enabling them to better cope with different scenarios and target individuals. Iteratively adjusting the interaction strategy rule base allows the interaction strategies to continuously adapt to new situations and needs.

[0020] Secondly, this application provides an interactive system based on millimeter-wave radar-based human behavior, employing the following technical solution: An interactive system based on millimeter-wave radar-based human behavior includes: The data acquisition module is used to acquire raw point cloud data of each target person from the millimeter-wave radar of the target area, and convert the raw point cloud data into a target state vector in a unified coordinate system. The target state vector includes at least position and velocity. The data processing module is used to construct a dynamic graph structure with the target person as the node and the relative distance and movement trend correlation between the target persons calculated based on the target state vector as the edge, and to extract the graph index of the dynamic graph structure. The graph index includes at least one of node connectivity component, clustering coefficient, and degree centrality. The scenario recognition module is used to identify the current target scenario based on the graph indicators and the target state vectors corresponding to each target person; The instruction generation and dispatch module is used to generate interactive instructions for the target area and dispatch them to the corresponding interactive devices based on the current target scene and the historical interaction vectors of each target character in the target area.

[0021] Thirdly, this application provides a computer device that adopts the following technical solution: A computer device includes a memory, a processor, and a computer program stored in the memory, the processor executing the computer program to implement the interactive method for human behavior based on millimeter-wave radar as described in the first aspect.

[0022] In summary, this application includes at least one of the following beneficial technical effects: By constructing a graph structure with target individuals as nodes and relationships between them as edges, and calculating graph metrics, the relationships and interaction patterns between target individuals can be intuitively represented. Through the graph structure and graph metrics, potential connections and behavioral patterns between individuals can be uncovered, providing deeper information for accurately identifying target scenarios. Identifying the current target scenario based on graph metrics and target state vectors, combining information about relationships between individuals and individual state information, enables a more accurate assessment of the actual situation within the target area. Interactive commands are generated based on the current target scenario and the historical interaction vectors of each target individual in the target area. This fully considers the real-time situation of the target area and the historical behavioral patterns of individuals, making the generated interactive commands more targeted and personalized, better meeting the needs of different individuals and scenarios, and improving the effectiveness of interaction and user experience. Attached Figure Description

[0023] Figure 1 This is a first flowchart of an embodiment of the method of this application; Figure 2 This is a second flowchart of an embodiment of the method of this application; Figure 3This is a third flowchart of an embodiment of the method of this application; Figure 4 This is the fourth flowchart of an embodiment of the method of this application; Figure 5 This is the fifth flowchart of an embodiment of the method of this application; Figure 6 This is the sixth flowchart of an embodiment of the method of this application; Figure 7 This is the seventh flowchart of an embodiment of the method of this application. Detailed Implementation

[0024] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figure 1-7 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0025] The first embodiment of this application discloses an interactive method based on millimeter-wave radar character behavior. (Refer to...) Figure 1 The interaction method may include S110-S140: S110 acquires raw point cloud data of each target person from millimeter-wave radar in the target area and converts the raw point cloud data into target state vectors in a unified coordinate system. The target state vectors contain at least position and velocity. S120: Using the target person as the node and the relative distance and movement trend correlation between the target persons calculated based on the target state vector as the edge, a dynamic graph structure is constructed, and graph indicators of the dynamic graph structure are extracted. The graph indicators include at least one of node connectivity components, clustering coefficient, and degree centrality. S130, Identify the current target scenario based on the graph indicators and the target state vectors corresponding to each target person; S140: Based on the current target scene and the historical interaction vectors of each target person in the target area, generate interaction instructions for the target area and send them to the corresponding interactive devices.

[0026] Specifically, for S110-S120, a 77GHz millimeter-wave radar array is deployed in the target area of ​​the exhibition hall, and raw point cloud data is acquired through MIMO technology. This data is transformed into a target state vector through coordinate transformation algorithms (e.g., mapping the radar local coordinate system to a unified global world coordinate system using homogeneous coordinate system transformation). This vector contains at least position (XYZ coordinates) and velocity (instantaneous velocity components calculated based on the Doppler effect); and a Kalman filter is used to smooth noise and enhance data accuracy.

[0027] Using each target person as a graph node, the relative distance between people is calculated using the Euclidean distance formula, and the correlation of motion trends (such as the cosine value of the angle between velocity and direction) is analyzed using the cosine similarity algorithm. The two are then weighted and fused as the graph edge weights to construct a dynamic graph structure. Graph metrics are extracted using graph analysis libraries (such as Python's NetworkX), including connected components (to identify group segmentation), clustering coefficients (to measure the density of the group), and degree centrality (to assess individual influence). These metrics are updated in real time to reflect dynamic changes in the scene.

[0028] Reference Figure 2 S130, the steps for identifying the current target scenario based on the graph indicators and the target state vectors corresponding to each target person include S210-S230: S210, Spatiotemporally align and concatenate the graph indicators with the target state vectors corresponding to each target person to construct a scenario feature matrix; S220: Input the scenario feature matrix into the pre-trained scenario classification model and output the category label of the current target scenario; S230, Calculate the scenario confidence score for verification based on the probability distribution or confidence score output by the scenario classification model.

[0029] Specifically, the timestamp alignment method is used to synchronize the graph indicators with the target state vectors at the corresponding times in time and space. The 6-dimensional state vector of each target is concatenated with the 3-dimensional graph indicators to form a 9-dimensional feature through feature concatenation operation. Then, the scenario feature matrix is ​​constructed by sorting by person ID (dimension N×9, where N is the number of target people). Missing data is handled by forward imputation method.

[0030] The scenario feature matrix is ​​input into a pre-trained scenario classification model, which adopts a CNN-LSTM hybrid architecture: convolutional layers (3 convolutional blocks, 3×3 kernel size, ReLU activation) extract spatial features, LSTM layers (2 bidirectional structures, 64 hidden units) capture temporal dependencies, and fully connected layers output 6 scenario labels (such as "gathering and talking", "single person moving", etc.). The model uses the cross-entropy loss function during training, and the optimizer adopts Adam (learning rate 0.001, weight decay 1e-5).

[0031] When calculating the scenario confidence, the maximum value of the Softmax probability distribution output by the model is taken as the initial confidence. It is then corrected by combining the Bayesian rule with the prior scenario probability (based on historical data statistics). The formula is P-correction = (P-model × P-prior) / ∑(P-model × P-prior). The scenario threshold is set to 0.75. If it is lower than this value, the radar parameter adjustment process is triggered. P-model represents the initial confidence output by the model, and P-prior represents the prior scenario probability obtained based on historical data statistics.

[0032] Reference Figure 3The steps following the verification of scenario confidence include S310-S350: S310 monitors scene complexity indicators in real time from millimeter-wave radar in the target area when the scene confidence is lower than the scene threshold. S320 calculates the adjustment vector of radar scanning parameters based on scene complexity index, current target scene category label and scene threshold; S330 dynamically adjusts the scanning frequency, resolution, or beamwidth of the millimeter-wave radar based on the adjustment vector. S340, after adjustment, reacquire and verify the quality indicators of point cloud data to ensure that the target state vector error under the unified coordinate system is lower than the preset threshold. After adjusting the S350, record the parameter changes before and after adjustment to the historical log.

[0033] Specifically, the scene complexity index is calculated by weighting three factors: target density (number of people per unit area, weight set to 0.4), motion speed standard deviation (weight set to 0.3), and occlusion frequency (overlap rate of targets in adjacent frames, weight set to 0.3). The comprehensive value of scene complexity is normalized to [0,1].

[0034] Based on the complexity index (C), the current scenario label, and the threshold, a piecewise function is used to calculate the adjustment vector: when C < 0.3, the scanning frequency is reduced by 20%; when 0.3 ≤ C < 0.7, the resolution is increased by 1 level (e.g., the distance resolution is increased from 0.1m to 0.05m); when C ≥ 0.7, the beamwidth is reduced by 30% (e.g., the horizontal beam angle is reduced from 60° to 42°). At the same time, a scenario correction coefficient is introduced, such as an additional 10% increase in scanning frequency for the "fast movement" scenario.

[0035] Parameter adjustment is achieved through the radar control API, and commands are sent using the UDP protocol. The command format is JSON, which includes parameter type, target value, and checksum. The hardware response timeout is set to 500ms, and a retry mechanism is triggered on failure, with a maximum of 3 retries.

[0036] After adjustment, verify the point cloud quality: calculate the root mean square error (RMSE) of the target state vector, set the position error threshold to 0.2m and the velocity error threshold to 0.1m / s, and compare the RMSE average of 30 frames of data before and after adjustment. If the standard is met, record the parameter changes (including timestamp, old parameter value, and adjustment range) to the SQLite database historical log table.

[0037] Reference Figure 4 In S140, the steps for generating interaction commands for the target area based on the current target scene and the historical interaction vectors of each target character in the target area include S410-S450: S410: The category label and context confidence of the current target scene are embedded and feature-stitched with the historical interaction vectors of each target person in a time series to construct a dynamic response feature matrix; S420 inputs the dynamic response feature matrix into the pre-trained instruction generation model and outputs a set of candidate interaction instruction sequences; S430 evaluates and filters the generated candidate instruction sequence and outputs the optimized interactive instruction; S440, perform behavioral pattern matching on the dynamic response feature matrix to retrieve interactive instruction templates under similar historical scenarios; S450 generates personalized interactive instruction parameters based on search results, optimized interactive instructions, and a preset interactive strategy rule base through parameter interpolation or rule reasoning.

[0038] Specifically, when constructing the dynamic response feature matrix, the current scenario label, scenario confidence, and historical interaction vectors (instruction types and execution results of the past 5 interactions) are embedded in a time series: the historical vectors are processed using a Transformer encoder (4 layers, 8 heads), and then concatenated with the current features to form a matrix of dimension N×25 (N is the target number of people), with missing historical data filled with zero vectors.

[0039] The instruction generation model adopts a Seq2Seq architecture. The encoder is a GRU (128 hidden units) that processes the dynamic response feature matrix, and the decoder is an attention mechanism LSTM (256 hidden units). The output contains candidate sequences (5 by default) containing instruction type (such as "voice prompt" or "light guidance"), target person ID, and execution parameters (volume, brightness, etc.). During training, beam search (3-beam width) is used to optimize the output.

[0040] Reference Figure 5 S430, the steps of evaluating and filtering the generated candidate instruction sequence and outputting the optimized interactive instruction include S510-S550: S510 performs spatiotemporal conflict detection on candidate instruction sequences to identify resource competition or behavioral logic contradictions between instructions. S520 calculates the spatial feasibility score for command execution based on the real-time location distribution and motion trajectory prediction of the target person. S530 integrates contextual confidence, historical interaction vector similarity, and device status parameters to generate a dynamic decay factor; S540 ranks candidate instructions using a weighted formula, which is: Optimization Score = Scenario Confidence × Spatial Feasibility Score / Dynamic Decay Factor; S550 selects instructions whose optimization scores are higher than the score threshold, injects safety constraints, and outputs optimization instructions.

[0041] Specifically, spatiotemporal conflict detection is achieved through a rule engine: temporal conflict detection checks the overlap rate of instruction execution periods (e.g., >50% is considered a conflict), spatial conflict calculates the IoU (Intersection over Union) of the instruction's effective area (e.g., >0.3 is considered resource competition), logical contradictions are judged through a decision tree (e.g., "move left" and "move right" both point to the same target), and conflicting instructions are marked before entering the feasibility assessment stage.

[0042] The spatial feasibility score is based on the A* algorithm to predict the target's trajectory in the next 3 seconds, and calculates the spatial intersection ratio between the instruction execution area and the predicted trajectory (if the ratio is >60%, 1 point is awarded; otherwise, it is linearly decayed to 0 points). The distance to obstacles is also considered (if >1.5m, 0.2 points are added). The comprehensive score range is [0, 1.2].

[0043] The dynamic decay factor integrates three factors: scenario confidence (with a weight of 0.5), historical interaction vector cosine similarity (with a weight of 0.3), and remaining device power (normalized to [0,1], with a weight of 0.2). The formula is dynamic decay factor = 1 / (0.5P + 0.3S + 0.2E), which ensures increased decay when the similarity is high or the power is low. The weights can be set and modified according to the actual situation. P is scenario confidence, S is historical interaction vector cosine similarity, and E is remaining device power.

[0044] After the optimization score is calculated according to the formula, the candidate instructions are sorted in descending order. The top two instructions with scores greater than the set value (e.g., 0.85) are selected and safety constraints are injected (e.g., volume ≤ 60dB, light intensity ≤ 500lux). Finally, the optimized instructions are output.

[0045] Historical interaction templates are retrieved from the FAISS vector library. The dynamic response feature matrix is ​​used as the query vector. Euclidean distance < threshold (e.g., 0.3) is considered a similar scenario. The parameter range in the template is extracted for personalized adjustment.

[0046] Personalized instructions are generated using parameter interpolation: if the historical template parameters are [a, b] and the current scenario features correspond to the weight w, then the new parameter = a + w × (ba). The new parameter is then modified by combining the rule base (e.g., "increase volume by 20% for elderly users"), and finally the parameters are dispatched to interactive devices (e.g., smart speakers, LED light strips) via the MQTT protocol. The device response status is fed back to the system in real time.

[0047] Reference Figure 6 In S140, the steps following the dispatch of the interactive instruction include S610-S650: The S610 monitors the physiological response indicators of the target person in real time and constructs an emotional feedback vector by combining the preset emotional mapping rules. The physiological response indicators are extracted based on the micro-motion features of millimeter-wave radar. S620, the deviation value between the emotional feedback vector and the expected interaction effect; S630, if the deviation value exceeds the deviation threshold, the emergency intervention protocol is triggered; S640: After the emergency protocol is activated, the current interactive command stream is frozen and switched to the preset security scenario mode; S650 generates intervention logs and links them to the historical interaction vectors of each target individual.

[0048] Specifically, physiological response indicators are extracted using millimeter-wave radar micro-motion features: short-time Fourier transform (STFT) is used to analyze the frequency (respiratory rate, 0.1-0.5Hz) and amplitude (0.1-1cm) of chest micro-motions, and combined with heart rate (extracted through Doppler frequency shift, 1-3Hz) to construct an 8-dimensional emotional feedback vector. Based on a pre-trained emotional mapping model (SVM classifier, with features as micro-motion spectrum features), emotional labels such as "pleasure", "neutrality", and "irritability" are output.

[0049] When calculating the deviation value, the cosine distance between the emotional feedback vector and the expected effect vector (based on the context label preset) is used as the deviation value. The threshold is set to 0.4. If it is exceeded, an emergency intervention is triggered: the instruction queue is frozen, the safe context mode is switched (such as turning off the bright light and playing soothing music), a log containing timestamps, abnormal indicators and intervention measures is generated, and the log is associated with the user ID and stored in the MongoDB historical interaction collection.

[0050] Reference Figure 7 Furthermore, the interaction methods also include S710-S740: S710 collects interaction feedback data from each target person and calculates interaction effect indicators based on the interaction feedback data. S720 encodes and fuses the interactive effect indicators, the parameters of the distributed interactive instructions, the current target scene category label and the situation confidence to construct an effect evaluation feature vector. S730: Input the effect evaluation feature vector into the pre-trained effect evaluation model and output the interaction effect score; S740 dynamically updates the historical interaction vectors of each target character, optimizes the parameters of the scenario classification model and the instruction generation model, and iteratively adjusts the interaction strategy rule base based on the interaction effect score.

[0051] Specifically, the collected interactive feedback data includes user action response (posture changes detected by radar) and device execution status (success / failure). Interactive effect indicators are calculated as follows: completion rate (number of successfully executed commands / total number of commands), response time (time difference between command issuance and action detection), and positive emotion rate (percentage of "pleasure" tags). These are weighted and integrated into a comprehensive indicator with a weighting of 4:3:3.

[0052] A 32-dimensional effect evaluation feature vector is constructed by feature hashing (interaction effect index), one-hot encoding (context label), and normalization (confidence). This vector is then input into the evaluation model, which is a 3-layer fully connected network with 128-64-32 hidden units. The output score is [0,10].

[0053] Dynamic updates are based on ratings: For example, when the rating is >8, the current interaction vector is appended to the user's history; for ratings between 6 and 8, only some effective features are updated; for ratings <6, model optimization is triggered—the parameters of the fully connected layer of the scenario classification model are fine-tuned using the current data (e.g., learning rate 0.0001, 10 iterations), and the attention weights of the instruction generation model are regularized (e.g., L2 penalty coefficient 0.01); for high-frequency, low-rating scenarios (e.g., "conflict" scenario ratings consistently <50), the rule engine is invoked to iteratively adjust the interaction strategy rule base (e.g., increasing the instruction priority weight for conflict scenarios). The optimization process uses incremental learning to avoid forgetting historical data and ensure that model performance dynamically improves with the number of interactions. The entire update process is executed asynchronously during system idle periods (e.g., CPU utilization <30%).

[0054] Based on the above method embodiments, the second embodiment of this application discloses an interactive system based on millimeter-wave radar character behavior. The interactive system based on millimeter-wave radar character behavior of this application embodiment can implement any of the above-described interactive methods based on millimeter-wave radar character behavior, and the specific working process of each module in the interactive system based on millimeter-wave radar character behavior can be referred to the corresponding process in the above method embodiments.

[0055] For ease of understanding, an example is given below: An interactive system based on millimeter-wave radar-based human behavior includes: The data acquisition module is used to acquire the raw point cloud data of each target person from the millimeter-wave radar in the target area, and convert the raw point cloud data into a target state vector in a unified coordinate system. The target state vector contains at least position and velocity. The data processing module is used to construct a dynamic graph structure with the target person as the node and the relative distance and movement trend correlation between the target persons calculated based on the target state vector as the edge, and to extract the graph index of the dynamic graph structure. The graph index includes at least one of node connectivity component, clustering coefficient, degree centrality. The scenario recognition module is used to identify the current target scenario based on graph indicators and the target state vectors corresponding to each target person; The instruction generation and dispatch module is used to generate interactive instructions for the target area and dispatch them to the corresponding interactive devices based on the current target scene and the historical interaction vectors of each target character in the target area.

[0056] The third embodiment of this application provides a computer device that may include a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement an interactive method based on millimeter-wave radar human behavior.

[0057] The memory can communicate with the processor via a communication bus, which can be an address bus, a data bus, a control bus, etc.

[0058] Additionally, the memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device.

[0059] Furthermore, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0060] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce a good effect.

[0061] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for human behavior-based interaction based on millimeter wave radar, characterized by, include: The original point cloud data of each target person is obtained from the millimeter-wave radar of the target area, and the original point cloud data is converted into a target state vector in a unified coordinate system. The target state vector includes at least position and velocity. Using the target person as the node and the relative distance and movement trend correlation between the target persons calculated based on the target state vector as the edge, a dynamic graph structure is constructed, and graph indicators of the dynamic graph structure are extracted. The graph indicators include at least one of node connectivity components, clustering coefficient, and degree centrality. Based on the graph indicators and the target state vectors corresponding to each target person, identify the current target scenario; Based on the current target scene and the historical interaction vectors of each target character in the target area, an interaction command for the target area is generated and dispatched to the corresponding interactive device.

2. The method of claim 1, wherein, The steps for identifying the current target scenario based on the graph indicators and the target state vectors corresponding to each target person include: The graph indicators are spatiotemporally aligned and feature concatenated with the target state vectors corresponding to each target person to construct a scenario feature matrix. The scenario feature matrix is ​​input into a pre-trained scenario classification model, which outputs the category label of the current target scenario. The scenario confidence is calculated and verified based on the probability distribution or confidence score output by the scenario classification model.

3. The method of claim 2, wherein, The steps for generating interaction instructions for the target area based on the current target scene and the historical interaction vectors of each target character in the target area include: The category labels and context confidence of the current target scene are time-series embedded and feature-concatenated with the historical interaction vectors of each target person to construct a dynamic response feature matrix; The dynamic response feature matrix is ​​input into a pre-trained instruction generation model, which outputs a set of candidate interaction instruction sequences. The generated candidate instruction sequence is evaluated and filtered to output the optimized interactive instruction; Behavioral pattern matching is performed on the dynamic response feature matrix to retrieve interactive instruction templates under similar historical scenarios; Based on the search results, optimized interactive instructions, and a pre-defined interactive strategy rule base, personalized interactive instruction parameters are generated through parameter interpolation or rule reasoning.

4. The method of claim 2, wherein, The steps following the verification after calculating scenario confidence include: When the scenario confidence level is lower than the scenario threshold, the scenario complexity index is monitored in real time from the millimeter-wave radar of the target area; The adjustment vector of the radar scanning parameters is calculated based on the scene complexity index, the category label of the current target scene, and the scene threshold. Based on the adjustment vector, the scanning frequency, resolution, or beamwidth of the millimeter-wave radar are dynamically adjusted. After adjustment, the quality indicators of the point cloud data were reacquired and verified to ensure that the target state vector error under the unified coordinate system was lower than the preset threshold. After adjustment, record the parameter changes before and after adjustment in the history log.

5. The method of claim 3, wherein, The steps of evaluating and filtering the generated candidate instruction sequences and outputting optimized interactive instructions include: Spatiotemporal conflict detection is performed on the candidate instruction sequence to identify resource competition or behavioral logic contradictions between instructions; Based on the real-time location distribution and motion trajectory prediction of the target person, the spatial feasibility score for command execution is calculated; By integrating contextual confidence, historical interaction vector similarity, and device status parameters, a dynamic decay factor is generated. Candidate instructions are ranked using a weighted formula, which is: Optimization Score = Scenario Confidence × Spatial Feasibility Score / Dynamic Decay Factor; Select instructions whose optimization scores are higher than the score threshold, inject security constraints, and output optimization instructions.

6. The method of claim 1, wherein, The steps following the dispatch of interactive instructions include: The physiological response indicators of the target person are monitored in real time, and an emotional feedback vector is constructed by combining the preset emotional mapping rules. The physiological response indicators are extracted based on the micro-motion features of millimeter-wave radar. Compare the deviation between the emotional feedback vector and the expected interaction effect; If the deviation value exceeds the deviation threshold, an emergency intervention protocol is triggered; Once the emergency protocol is activated, the current interactive command flow is frozen, and the system switches to the preset security scenario mode. Generate intervention logs and associate them with the historical interaction vectors of each target individual.

7. The method of claim 3, wherein the method further comprises: The interaction method also includes: Collect interaction feedback data from each target person, and calculate the interaction effect index based on the interaction feedback data; The interaction effect indicators, the distributed interaction instruction parameters, the current target scene category label and the scenario confidence are feature-encoded and fused to construct an effect evaluation feature vector; The effect evaluation feature vector is input into the pre-trained effect evaluation model, which outputs an interaction effect score. Based on the interaction effect score, the historical interaction vectors of each target character are dynamically updated, the parameters of the scenario classification model and the instruction generation model are optimized, and the interaction strategy rule base is iteratively adjusted.

8. A millimeter wave radar based human behavior interactive system, characterized by, Performing the interactive method based on millimeter-wave radar character behavior as described in any one of claims 1 to 7, comprising: The data acquisition module is used to acquire raw point cloud data of each target person from the millimeter-wave radar of the target area, and convert the raw point cloud data into a target state vector in a unified coordinate system. The target state vector includes at least position and velocity. The data processing module is used to construct a dynamic graph structure with the target person as the node and the relative distance and movement trend correlation between the target persons calculated based on the target state vector as the edge, and to extract the graph index of the dynamic graph structure. The graph index includes at least one of node connectivity component, clustering coefficient, and degree centrality. The scenario recognition module is used to identify the current target scenario based on the graph indicators and the target state vectors corresponding to each target person; The instruction generation and dispatch module is used to generate interactive instructions for the target area and dispatch them to the corresponding interactive devices based on the current target scene and the historical interaction vectors of each target character in the target area.

9. A computer device, comprising: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the interactive method for human behavior based on millimeter-wave radar as described in any one of claims 1 to 7.

Images