Human body perception detection method and system based on machine learning

Abstract

The invention relates to a human body sensing detection method and system based on machine learning, and the method comprises the steps: extracting target micro-motion change, space trajectory and time sequence features through the collection and high-precision marking of original time-space behavior data of millimeter wave, infrared, pressure and sound multi-type sensors in combination with preprocessing, noise removal and normalization standards, and carrying out the detection of a human body. A human body space-time behavior fingerprint database is constructed in an off-line or on-line state through a machine learning algorithm, and regional difference behavior characteristic models are respectively established for different space regions and installation materials. And automatically triggering multiple rounds of redundancy detection and learning optimization. The scheme reduces the false alarm rate of the system and supports continuous evolution of the model.

Description

Technical Field

[0001] This invention relates to the field of indoor environmental intelligent sensing and human presence detection technology, and in particular to a human perception detection method and system based on machine learning. Background Technology

[0002] With the increasing prevalence of smart buildings, smart homes, energy-saving scheduling, and smart offices, indoor human presence detection technology has become a crucial component of intelligent energy consumption control and automated scene management systems. Currently, the industry's widely adopted methods for human presence detection primarily rely on single or multiple indoor environmental sensors (such as millimeter-wave radar, infrared sensors, pressure sensors, and sound sensors) to acquire information on environmental changes within the space. These are then supplemented by traditional signal analysis or machine learning-based classification algorithms to infer the presence of human beings within the space. These solutions typically employ feature extraction and pattern recognition techniques to model and classify data such as dynamic and static changes, thermal signals, and pressure changes output by the sensors, achieving relatively reliable human presence detection results under ideal conditions.

[0003] However, several prominent problems exist in existing technological systems, creating practical bottlenecks for intelligent energy consumption control and scene-adaptive management: High false alarm rates occur when abnormal interference is frequent. Current human detection systems are often hampered by the fuzzy distribution of a single signal mode in abnormal conditions such as the presence of reflective objects, pet activity, or complex mechanical movements, leading to the misidentification of non-human targets as human beings. For example, reflective objects can generate millimeter-wave echoes similar to those of the human body; the motion characteristics of pets and small robots, when frequently active, easily overlap with those of the human body; the operation of mechanical devices in conference rooms and the handling of large items can easily induce misjudgments by the detection system. High false alarm rates directly lead to malfunctions in energy consumption control (such as inappropriate activation of lighting and air conditioning), reducing energy-saving benefits and affecting the user experience. Summary of the Invention

[0004] This application provides a human perception detection method and system based on machine learning, which aims to solve one of the problems or problems of the existing technology mentioned in the background.

[0005] This application provides a human perception detection method based on machine learning, which specifically includes: S1: Collect raw spatiotemporal behavior data from environmental sensors (including millimeter-wave sensors, infrared sensors, pressure sensors, and sound sensors) at different locations and of different types, and label the sampled data with the corresponding spatial region and installation material category.

[0006] S2: Standardize and preprocess the collected raw spatiotemporal behavior data, including removing static background noise and normalizing the amplitude distribution of each sensor channel to maintain the consistency of the output of various sensors and facilitate subsequent feature extraction.

[0007] S3: Extract target micro-motion change features, spatial trajectory sequences, and appearance / disappearance time sequence statistics from preprocessed spatiotemporal behavior data, and form a multi-dimensional feature vector for modeling.

[0008] S4: Based on the multidimensional feature vector, construct a spatiotemporal behavioral fingerprint database of human existence in offline or online states through machine learning algorithms, and establish regionally differentiated behavioral feature models for different spatial regions and installation materials.

[0009] S5: Perform similarity matching between the spatiotemporal behavioral feature vectors collected in real time and the established behavioral fingerprint database, and output the human presence confidence score for each detected instance based on the discriminative features.

[0010] S6: For detection instances with confidence scores below the preset threshold or exhibiting highly anomalous distributions, perform adaptive anomaly weighting discrimination by combining multidimensional contextual data (including door magnetic opening and closing status, pressure mutations, historical energy consumption characteristics, and regional attribute labels).

[0011] S7: When the adaptive anomaly weighted discrimination result is a non-human behavior feature, short-term high-density multi-sensor modal redundancy detection is automatically triggered, and the newly collected multi-modal behavior data is further compared with the regional behavior feature model.

[0012] S8: If the redundant detection result still differs from the features in the human behavior fingerprint database after multiple samplings, the current detection is automatically determined to be an abnormal interference source, false alarm filtering is performed, and the abnormal data and its features are added to the database for model self-learning. If the redundant detection result is the same as the features in the human behavior fingerprint database, it is determined that a human body is present.

[0013] S9: After the system receives the annotation information from the user or the operation and maintenance end, it automatically updates the spatiotemporal behavioral fingerprint database and optimizes the anomaly detection algorithm and regional difference behavioral feature model in real time to dynamically improve the detection robustness and generalization ability.

[0014] This application also provides a human perception detection system based on machine learning, which uses the above-mentioned human perception detection method based on machine learning to detect the presence of a human body.

[0015] This application provides a human perception detection method and system based on machine learning, which has the following beneficial effects: (1) By using multiple types of environmental sensors such as millimeter wave, infrared, pressure, and sound, high-precision clock synchronization, spatial position marking and installation material depth marking are uniformly performed to realize the structured data collection of human and non-human micro-movements, trajectories, and time sequences in the target space, which greatly improves the accuracy of multimodal feature joint analysis and regional environmental adaptability.

[0016] (2) Combining multiple indicators such as micro-motion changes, spatial trajectories, and appearance / disappearance time sequence, and using algorithms such as unsupervised clustering and isolated forest, a behavioral fingerprint database of human existence is constructed in layers for different spatial regions and material environments. Through normalization, grouping, clustering and dynamic time sequence analysis, the ability to distinguish behavioral patterns in different scenarios is significantly enhanced.

[0017] (3) Perform efficient similarity calculation between the detected instance and the historical behavioral fingerprint. Low confidence or abnormal instances are combined with spatial context (door magnet, pressure mutation, historical energy consumption, regional attributes, etc.) and multimodal anomaly weighted discrimination is used to greatly improve the rapid identification and filtering efficiency of non-human abnormal interference.

[0018] (4) For instances initially judged as abnormal behavior, high-density multimodal sampling and re-examination are automatically triggered. Combined with multiple rounds of feature comparison and confidence tracking, atypical anomalies are captured in a timely manner and false alarms are masked. At the same time, abnormal data is archived in real time to provide feedback to optimize the fingerprint model and realize the continuous adaptive evolution of the system. Attached Figure Description

[0019] Appendix Figure 1 This is the main flowchart of a human perception detection method based on machine learning.

[0020] Appendix Figure 2 This is a sub-flowchart of a human perception detection method based on machine learning.

[0021] Appendix Figure 3 This is another sub-flowchart of a human perception detection method based on machine learning. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0024] As attached Figure 1 As shown, this application provides a human perception detection method based on machine learning, specifically including: S1: Collect raw spatiotemporal behavior data from environmental sensors (including millimeter-wave sensors, infrared sensors, pressure sensors, and sound sensors) at different locations and of different types, and label the sampled data with the corresponding spatial region and installation material category.

[0025] S2: Standardize and preprocess the collected raw spatiotemporal behavior data, including removing static background noise and normalizing the amplitude distribution of each sensor channel to maintain the consistency of the output of various sensors and facilitate subsequent feature extraction.

[0026] S3: Extract target micro-motion change features, spatial trajectory sequences, and appearance / disappearance time sequence statistics from preprocessed spatiotemporal behavior data, and form a multi-dimensional feature vector for modeling.

[0027] S4: Based on the multidimensional feature vector, construct a spatiotemporal behavioral fingerprint database of human existence in offline or online states through machine learning algorithms, and establish regionally differentiated behavioral feature models for different spatial regions and installation materials.

[0028] S5: Perform similarity matching between the spatiotemporal behavioral feature vectors collected in real time and the established behavioral fingerprint database, and output the human presence confidence score for each detected instance based on the discriminative features.

[0029] S6: For detection instances with confidence scores below the preset threshold or exhibiting highly anomalous distributions, perform adaptive anomaly weighting discrimination by combining multidimensional contextual data (including door magnetic opening and closing status, pressure mutations, historical energy consumption characteristics, and regional attribute labels).

[0030] S7: When the adaptive anomaly weighted discrimination result is a non-human behavior feature, short-term high-density multi-sensor modal redundancy detection is automatically triggered, and the newly collected multi-modal behavior data is further compared with the regional behavior feature model.

[0031] S8: If the redundant detection result still differs from the features in the human behavior fingerprint database after multiple samplings, the current detection is automatically determined to be an abnormal interference source, false alarm filtering is performed, and the abnormal data and its features are added to the database for model self-learning. If the redundant detection result is the same as the features in the human behavior fingerprint database, it is determined that a human body is present.

[0032] S9: After the system receives the annotation information from the user or the operation and maintenance end, it automatically updates the spatiotemporal behavioral fingerprint database and optimizes the anomaly detection algorithm and regional difference behavioral feature model in real time to dynamically improve the detection robustness and generalization ability.

[0033] Step S1: Collect raw spatiotemporal behavioral data from environmental sensors (including millimeter-wave sensors, infrared sensors, pressure sensors, and sound sensors) at different locations and of different types, and label the sampled data with the corresponding spatial region and installation material category. Specifically, this includes: S1.1: Perform a unified data acquisition initialization process for all types of environmental sensors (including millimeter-wave sensors, infrared sensors, pressure sensors, and sound sensors) deployed in the target space, and configure a sampling clock synchronization mechanism to obtain raw sensor data streams containing timestamps, laying a synchronization foundation for subsequent spatiotemporal behavior analysis.

[0034] For various environmental sensors (including millimeter-wave sensors, infrared sensors, pressure sensors, and sound sensors) deployed in the target space, a physical distributed initialization calibration method (parameters: sensor physical ID, spatial location code, interface specification) is adopted to activate the sensor modules of each type and initialize the acquisition channels, ensuring that the equipment is in a working standby state.

[0035] Furthermore, through a centralized sampling scheduling control algorithm (parameters: unified scheduling clock, global NTP server address, local high-precision clock source), the unified configuration and synchronization of sampling clocks for cross-type sensors are achieved, so that the data acquisition process of all sensors can be carried out in parallel strictly according to the master clock beat.

[0036] Furthermore, an event-driven timestamp tagging mechanism (parameters: unified timing benchmark, data packet sequence number, event trigger flag) is adopted to add a high-precision timestamp to each raw sensor data stream in real time, and to ensure that the multi-source raw data has strict temporal traceability.

[0037] Furthermore, by applying original signal multi-channel buffering technology (parameters: ring buffer depth, edge device buffer resources, network transmission bandwidth), the collected timestamp data streams are temporarily stored in separate channels to support timing alignment and packet loss monitoring, laying the foundation for eliminating acquisition time differences in post-processing.

[0038] By using a joint sensor online status diagnostic program (parameters: heartbeat signal frequency, data packet integrity verification rules), the stability and integrity of each sensor data acquisition link are monitored in real time, and abnormal nodes are automatically reported, ensuring the consistency, integrity, and reliability of multi-source alignment of the acquired data.

[0039] The data output obtained through the above chain derivation is a multi-source raw sensor data stream containing a unified high-precision timestamp, which lays a solid synchronous foundation for subsequent spatial location mapping, data fusion, and spatiotemporal behavior analysis, and achieves temporal consistency in behavior detection data acquisition across types and regions.

[0040] For example, in an open-plan office environment with an area of ​​100m², four millimeter-wave sensors (model AWR6843), eight infrared sensors (model HC-SR501), six pressure sensors (model FSR404), and two high-fidelity microphones (model MP34DT05) are deployed. All sensors are physically connected to the same local area network and synchronized using the NTP protocol (master clock deviation less than 1ms). The sampling period is set to 20ms. All acquisition tasks are initiated by the central controller (embedded Linux system, configured with an 8-core Cortex-A53 CPU and 2GB of memory). After power-on, the sensors automatically establish a heartbeat signal with the central server and complete the data channel initialization. During the acquisition process, all raw signal data are automatically accompanied by timestamps accurate to milliseconds and pushed to the edge buffer nodes via the UDP protocol. The data packet integrity is checked using the CRC32 algorithm on the acquisition link, and the packet loss rate is consistently less than 0.001%. When a large-scale behavior detection task is triggered (such as when people leave the site during peak hours), the system can ensure that every piece of raw data from all the above sensor types is strictly time-aligned, with clock synchronization error controlled within ±1ms, fully meeting the needs of downstream multimodal signal spatiotemporal joint analysis applications.

[0041] S1.2: Based on the above raw sensor data stream, call the data access interface, collect the data output of each sensor according to the preset physical location code, map the spatial location label to the raw data one by one, obtain the sampling data with unique spatial coordinates, and provide a source of location information for subsequent spatial trajectory modeling.

[0042] For the unified high-precision timestamp raw data stream output by various types of environmental sensors that have been deployed and initialized, a location tag mapping method (parameters: physical location coding scheme, sensor physical ID, unified spatial coordinate system) is adopted to achieve unique spatial location identification of all collected data.

[0043] Furthermore, through the data access interface protocol abstraction layer (parameters: interface standard type, data packet structure definition, location label field), data capture and organization are performed on each acquisition channel using preset physical location codes, ensuring that each piece of acquired data is structurally archived according to the spatial distribution of the source device.

[0044] Furthermore, a one-to-one spatial coordinate mapping algorithm (parameters: office area coordinate system, area grid partitioning table, sensor installation point number) is adopted to convert the physical location ID in the original sensor data into spatial coordinate labels, and the spatial labels are concatenated and bound with the original data to form a spatially unique mapping dataset.

[0045] Furthermore, through a multi-dimensional spatial label consistency verification mechanism (parameters: spatial distribution density, label conflict detection rules), the spatial labels and physical identifiers corresponding to all sampled data are automatically scanned, and duplicate, missing, and abnormally inconsistent data are checked and removed, and spatial location allocation conflicts are automatically corrected.

[0046] Furthermore, an efficient data indexing structure is used to construct (parameters: hash table or B+ tree index, spatial label hash key) to generate a unique index table for the sampled dataset that can be retrieved by spatial coordinates, enabling rapid location retrieval and traceable data source for sampled data from any spatial region or sensor node.

[0047] Through the chain-like acquisition and location tag binding process described above, all raw sensor sampling data is transformed into a dataset with unique spatial coordinates, thus realizing the foundation for integrated data spatial tracing and behavioral trajectory analysis.

[0048] For example, in a multi-zone open-plan office space containing 8 physical zones (labeled Z1 to Z8), each zone having an area of ​​12.5 m², 4 millimeter-wave sensors are deployed (installation point COORD). XY =[5,3]、[12,6]、[19,8]、[24,14]), 8 infrared sensors (coordinate points distributed in the center of each area), 6 pressure sensors (accuracy ±0.05m in floor coordinate system), 2 high-fidelity microphones (ceiling corners), all sensors are configured with independent physical IDs and spatial location codes. Using a unified spatial grid coding (such as Z3-X2Y5) and physical ID association rules, each set of collected data is automatically embedded with spatial location tags and organized into a (spatial tag to data stream) mapping structure through hash indexing. For a large-scale sampling task (such as the flow of people in multiple areas during the same period), the system output dataset can identify the spatial block to which each sampled data belongs with millisecond-level accuracy, and the retrieval latency is less than 2ms. After consistency verification, 99.99% of the data in each spatial partition has achieved unique code identification, and the original data can be traced unambiguously to the installation point, partition and physical coordinates. This step significantly improves the performance of subsequent spatial trajectory modeling and accurate location of abnormal behavior, reducing data retrieval and location time by at least 80% compared to traditional manual labeling systems.

[0049] S1.3: The parameters of key areas such as sensor mounting base, wall / ground material, and obstruction structure are digitally collected using an engineering identification entry mechanism and associated with data tags of each sampling point. This enables in-depth labeling of the original spatiotemporal behavioral data and installation material categories, providing a data benchmark for data domain adaptability in multi-regional and multi-material environments.

[0050] S1.4: Establish a unified data structure encapsulation format for various types of raw sensor data (millimeter-wave echo values, infrared thermal data, pressure response curves, acoustic amplitude / spectrum). Through protocol conversion and multi-source fusion interfaces, ensure that different types of data can be normalized and integrated into a unified data acquisition base, and realize the initial structured fusion of multimodal data.

[0051] S1.5: Based on the actual sampling period and preset trigger events (such as significant changes in the appearance / disappearance of targets in space), dynamically adjust the data acquisition frequency and period, and stream and persistently store the original spatiotemporal behavioral data at key sampling moments to provide original data support for subsequent multi-temporal behavioral fingerprint feature extraction and spatiotemporal trajectory modeling.

[0052] Step S2: Standardize and preprocess the collected raw spatiotemporal behavior data, including removing static background noise and normalizing the amplitude distribution of each sensor channel to maintain the consistency of the outputs of various sensors and facilitate subsequent feature extraction. Specifically, this includes: S2.1: Perform sensor type normalization processing on the raw spatiotemporal behavior data. The corresponding input is raw spatiotemporal behavior data from multiple environmental sensors including millimeter wave, infrared, pressure, and sound. Channel mapping and signal resampling techniques are used. The output is channel-normalized spatiotemporal behavior data with a consistent format, providing a unified data foundation for subsequent noise cancellation and amplitude normalization processing.

[0053] The input consists of raw spatiotemporal behavioral data collected by multiple types of environmental sensors (millimeter wave, infrared, pressure, sound) with unified clocks and spatial location annotations. The data is composed of a multi-channel data stream with unified high-precision timestamps, spatial tags, and sensor type attributes.

[0054] By employing a channel mapping method (parameters: sensor physical ID mapping table, channel standard index), the raw outputs of different types of sensors are mapped to pre-defined unified channel identifiers, standardizing the channel attribution of various input data and ensuring that subsequent processing can unambiguously identify the data source.

[0055] Furthermore, by using a signal resampling algorithm (parameters: target uniform sampling rate fs, list of original sampling rates for each channel), the sampling rate of multi-source data streams is made consistent. For millimeter wave channel, infrared channel, pressure channel and acoustic channel, upsampling, downsampling or interpolation resampling are performed according to their original sampling rates, and the standardized sample of each data point at the uniform sampling time is output.

[0056] Furthermore, channel verification and synchronization processing (parameter: channel synchronization threshold δ) are performed. t(Packet loss identifier mask) to perform timestamp synchronization and channel integrity verification on the mapped and resampled data of each channel, remove abnormal samples with excessive timing alignment error or signal loss, and output strictly spatiotemporally consistent multi-channel reference data.

[0057] Furthermore, by using a format normalization encapsulation method (parameters: standard data structure template, signal domain name list), different physical quantity signals (such as millimeter wave echo, voltage, current, heat, sound pressure) are uniformly mapped to the standardized encapsulation structure, thereby achieving format consistency of multimodal data and forming a channel normalized spatiotemporal behavior dataset.

[0058] Through the aforementioned algorithm chain, the original multi-type sensor data is uniformly transformed into normalized spatiotemporal behavior data based on channel coding, strict consistency of sampling time, and standardized structural encapsulation. This provides an orderly and highly comparable unified data foundation for subsequent background noise removal and amplitude distribution normalization analysis, achieving a data consistency benchmark for multi-source signals in the behavior detection chain.

[0059] For example, in an open-plan office area, the system receives data streams simultaneously acquired by 4 millimeter-wave sensors (raw sampling rate 100Hz), 8 infrared sensors (sampling rate 50Hz), 6 pressure sensors (sampling rate 40Hz), and 2 microphones (sampling rate 8kHz). Through channel mapping, millimeter-wave, infrared, pressure, and acoustic signals are mapped to standard channels [CH1-CH20] respectively. A resampling algorithm is used to unify all signals to a 100Hz sampling frequency (fs = 100Hz). Microphone data undergoes bandpass filtering and downsampling, while infrared and pressure signals are extended through linear interpolation or sample-and-hold, ensuring that each sampling moment corresponds to a complete data vector for all channels. Data samples with timestamp alignment deviations greater than 2ms are discarded, and packet loss points are marked as missing. The final generated dataset is a set of channel-normalized spatiotemporal behavior matrices of dimension [N × 20] (N being the number of sampling points). The system automatically verifies that all data domain encapsulation formats are consistent, and different physical quantity signals are encapsulated using standard fields (such as Echo, Thermal, Force, dB) to ensure consistent extraction of multimodal features subsequently. In various office scenarios (such as conference rooms, open areas, and tea areas), the normalized data met the technical requirements for subsequent unified analysis of behavioral patterns and anomaly detection in terms of amplitude distribution, sampling nodes, and signal synchronization accuracy. Performance testing showed that this step reduced the data entry time window error of different sensor channels to within ±1ms and controlled the amplitude truncation rate to below 0.01%, effectively supporting the smooth implementation of multimodal spatiotemporal behavior joint discrimination and high-precision false alarm filtering algorithms.

[0060] S2.2: Perform static background noise modeling and removal on the channel normalized spatiotemporal behavior data. The input object is the format-normalized spatiotemporal behavior data. Apply background modeling and dynamic baseline filtering algorithms to obtain dynamic response spatiotemporal behavior data that has removed persistent environmental interference, creating a stable signal baseline for amplitude normalization.

[0061] S2.3: Perform multi-channel amplitude normalization processing on dynamic response spatiotemporal behavior data. The input object is dynamic response spatiotemporal behavior data after static background noise removal processing. The algorithm adopts Z-score normalization, min-max normalization and other algorithms to output multi-channel normalized spatiotemporal behavior data with consistent amplitude, thereby improving the comparability of outputs from different sensors in the feature extraction stage.

[0062] S2.4: Perform anomalous signal truncation and extreme value correction processing on multi-channel normalized spatiotemporal behavior data. The input is multi-channel normalized spatiotemporal behavior data with consistent amplitude. Based on the 3σ principle or IQR outlier algorithm, perform anomalous interval pruning to obtain high-quality signal data with stable extreme values, effectively suppressing the influence of transient anomalous peaks on subsequent spatiotemporal micro-motion feature extraction.

[0063] S2.5: Perform specific parameter optimization processing on the high-quality signal data after extreme value correction according to spatial region and installation material category. The input data is the high-quality signal that has been extreme value corrected. The algorithm of regional attribute label optimization and material related parameter correction is adopted to output a standardized spatiotemporal behavior dataset with regional adaptability and material consistency. This lays an integrated data foundation for subsequent multi-dimensional feature modeling of target micro-motion changes and fingerprint database construction.

[0064] Step S3: Extract target micro-motion change features, spatial trajectory sequences, and appearance / disappearance time sequence statistics from the preprocessed spatiotemporal behavior data, and form a multi-dimensional feature vector for modeling. For example... Figure 2 As shown, it specifically includes: S3.1: Perform target micro-motion change feature detection processing on the standardized preprocessed multi-source spatiotemporal behavior data to obtain micro-motion change parameters that reflect the fine-grained motion state of the target, and output them as target micro-motion change feature values.

[0065] S3.2: Based on the target's micro-motion change feature value, execute the spatial trajectory sequence construction algorithm to model the target's continuous motion trajectory in space, realize the generation of spatial trajectory sequence data, and provide spatial distribution basis for further distinguishing human body from abnormal interference behavior.

[0066] The input is standardized preprocessed multi-source spatiotemporal behavior data, including calibrated channels, high-quality amplitude signals, spatial location, installation material labels, and other attributes.

[0067] A spatial trajectory reconstruction algorithm (parameters: spatiotemporal synchronous data stream, target micro-motion change feature value set, spatial region coordinate system reference) is adopted to realize real-time position estimation of each detected target at multiple time points and multiple sensor sampling points in space, and generate a spatial trajectory point sequence.

[0068] Furthermore, by using multi-point trajectory interpolation and fitting methods (parameters: interpolation type such as cubic spline, polynomial fitting order, outlier removal threshold), the continuous expression of spatiotemporal discrete sampling trajectories is realized, and a complete and smooth motion trajectory curve of the target in the monitoring space is obtained, clarifying the motion path distribution.

[0069] Furthermore, by using trajectory clustering analysis algorithms (parameters: trajectory Euclidean distance, DBSCAN or K-means cluster radius, minimum number of sampling points), the spatial trajectories of a single target or multiple active targets are classified into categories and behavioral patterns, separating the main trajectory and abnormal trajectory of independent targets to form trajectory cluster identifiers.

[0070] Furthermore, a trajectory attribute statistical analysis method (parameters: trajectory length, average speed, stopping node judgment threshold, turning angle threshold, etc.) is used to extract key feature indicators of trajectory segments, including multi-dimensional behavioral parameters such as movement distance, speed distribution, dynamic change rate, dwell heat distribution, and activity range, to enhance the ability to distinguish the trajectory characteristics of different types of behaviors (such as human, pet, and mechanical).

[0071] Through the above chain processing, the single or multi-body activities detected based on the target's micro-motion change characteristics are transformed into high-resolution, structured target motion trajectory data through spatial trajectory sequence construction, continuous fitting, clustering labeling, and behavioral parameter analysis. This provides a solid spatial distribution basis for subsequent time-series statistics and multi-dimensional feature fusion modeling, achieving the technical effect of accurately distinguishing human bodies from abnormal interference behavior patterns.

[0072] For example, in a 60-square-meter open-plan office area, the system deploys 8 millimeter-wave, 12 infrared, 6 pressure, and 4 sound sensor channels. After S3.1 micro-motion feature detection, it is determined that two groups of targets are moving simultaneously within a certain time period. Spatial trajectory reconstruction is performed on the multi-channel, synchronously sampled spatiotemporal data of each target, with the spatial area divided into 0.5-meter grids. Trajectory points are smoothed using cubic spline interpolation, and abnormal stationary or jumping points are identified and removed using the 3σ criterion. The trajectory clustering radius is set to 1 meter, and the minimum number of sampling points is 10, ultimately distinguishing two non-overlapping target movement trajectories. The total length (8.1 meters, 4.7 meters), average speed (0.62 m / s, 0.38 m / s), and main distribution areas of dwelling hotspots (conference table area, doorway) of each trajectory are statistically analyzed. Comparison with typical human and pet trajectory patterns in the fingerprint database shows that the movement range width, dwelling pattern, and speed distribution indicators highly match those of typical human activities. The output consists of two time-series spatial trajectory data vectors, including the spatial coordinates (x, y, z), motion velocity, and dwell time for each sampling moment. This output data is used in the subsequent S3.3 step to extract the timing statistics of target appearance / disappearance, enabling accurate subsequent discrimination between human presence patterns and abnormal interference patterns.

[0073] S3.3: For spatial trajectory sequence data, a dynamic time series analysis algorithm is used to extract the appearance / disappearance time series statistics of the target, including dwell time, intermittent activity cycle, and the time of entry and exit events, in order to characterize the target's time series activity pattern and obtain appearance / disappearance time series statistics parameters.

[0074] A dynamic temporal analysis algorithm (parameters: spatial coordinate sequence of trajectory points, sampling timestamp sequence, spatial region encoding) is used to model the temporal behavior changes of each target's motion trajectory within the sampling period.

[0075] Furthermore, a temporal segmentation method based on state transition detection (parameter: occurrence / disappearance amplitude threshold) is used. act Minimum dwell time T min This allows for the division of a set of trajectory points into consecutive "existing" and "missing" intervals in chronological order, and the assignment of an activity status marker to each interval.

[0076] Furthermore, a residence time calculation algorithm is adopted (parameters: start and end time points of the existing interval (t)). in , t out The target dwell time is calculated for each "existence" interval according to the following formula: Among them, T stay For the length of stay, t in t represents the start time of regional activities.out This is the time when the activity ends.

[0077] Furthermore, an intermittent periodicity discrimination algorithm is used (parameter: trajectory presence / missing period threshold T). gap Shortest interval sampling number N gap Periodic statistics are performed on consecutive "missing" state intervals to extract the target's departure / re-entry pattern. For entry and departure events, the corresponding time points are recorded as time-series event features.

[0078] Furthermore, through time series statistical analysis (parameters: sliding window (△t), number of active intervals), the number of times the target appears, the average dwell time, and the event density index are statistically analyzed within the sampling period to form a complete set of time series statistical parameters for the appearance / disappearance of the target.

[0079] Through the above-mentioned multi-level chain-like time-series algorithm processing method, the spatial trajectory sequence data system is transformed into multi-dimensional time-series statistical indicators covering dwell time, activity cycle, and entry / exit time, so as to achieve high-resolution characterization of the target time-series activity pattern and effectively improve the accuracy of subsequent multi-dimensional feature fusion and human body-abnormal interference separation and discrimination.

[0080] For example, in a conference room scenario, the system deploys 6 millimeter-wave and 8 pressure sensors, dividing the space into 3 areas (podium, table area, and doorway). The system continuously acquires the target space trajectory over 1200 seconds, outputting trajectory points in 100Hz time-series format. A state transition detection threshold θ is used. act =0.3, minimum dwell time T min =2s (i.e., continuous existence state determination for no less than 200 frames). After trajectory data segmentation, a target was analyzed to have 4 consecutive "existence" intervals, corresponding to (t... in , t out The intervals are (135s, 211s), (218s, 276s), (782s, 913s), and (989s, 1164s), and the dwell times are 76s, 58s, 131s, and 175s, respectively, according to the formula. The average value of the intermittent gap interval is 32s. The entry / exit events are recorded as the start and end points of each of the above intervals. Based on this, the system counts that the target appeared 4 times during the monitoring period, with an average dwell time of 110s and an intermittent period of 32s. The entry and exit events are aligned with the door magnet and area pressure changes. The above output parameters are combined into a multi-dimensional feature vector and compared with the typical human activity time sequence template in the fingerprint database. It highly conforms to the normal human activity pattern in an office setting, effectively avoiding false alarms caused by significant time sequence differences between the activity intervals under mechanical movement and pet interference and human patterns, and significantly improving the robustness and false alarm separation capability of the detection system.

[0081] S3.4: The behavioral attributes of the target, including the micro-motion change feature values, spatial trajectory sequence data, and appearance / disappearance time sequence statistical parameters, are multi-dimensional structured feature vectors generated through a feature vector encoding strategy. This ensures that the parameters of the causal chain in each processing dimension can collaboratively express the overall behavioral state of the target.

[0082] The input consists of target micro-motion change feature values, spatial trajectory sequence data, and appearance / disappearance time-series statistical parameters generated by the preceding sub-steps S3.1-S3.3. All of these data are derived from standardized synchronous spatiotemporal behavior datasets from multi-source multimodal environmental sensors (millimeter wave, infrared, pressure, and sound).

[0083] A feature-based hierarchical coding method (parameters: micro-motion change feature group, trajectory sequence group, and time series statistics group) is adopted to encode the behavioral parameters of various physical modes in a specific format, and to unify discrete and continuous variables into vector form.

[0084] Furthermore, through feature normalization and scale alignment algorithms (parameters: standardized mean and variance of various feature types, component weight settings), the normalization of multidimensional feature components in the numerical domain and distribution domain is achieved, and different physical quantities and different types of behavioral parameters are compressed into a unified range.

[0085] Furthermore, a multi-dimensional feature splicing and synthesis strategy (parameters: feature splicing sequence, feature missing processing strategy) is adopted to process the feature vectors with subtle changes. Spatial trajectory embedding vector Time series statistical parameter vector Vectors are concatenated in causal chain order to form structured high-dimensional feature vectors. in, For the micro-motion parameter set, For the set of spatial trajectory parameters, For the time series of occurrence / disappearance.

[0086] Furthermore, through a feature correlation analysis algorithm (parameters: mutual information threshold, correlation matrix, redundant attribute filtering rules), the linear and nonlinear correlations within the concatenated high-dimensional feature vector are determined, highly redundant or irrelevant components are removed, and the feature combination with maximum information is retained.

[0087] Furthermore, for the final multidimensional structured feature vector encoding result, a label embedding mapping method (parameters: spatial region label, sensor type label, material category label) is used to attach causal chain meta-information to each vector entry, thereby realizing the collaborative expression of each behavioral attribute.

[0088] Through the integrated chain derivation of encoding, normalization, splicing, correlation analysis and label embedding, multidimensional structured feature vectors of multi-source target behavior information are generated, which fully and synergistically reflect the spatial, temporal and micro-motion triple behavioral attributes of the detected target, providing a high-quality input foundation for machine learning modeling and the construction of spatiotemporal behavioral fingerprint database.

[0089] For example, in an open-plan office environment, the input consists of two sets of targets, each corresponding to a set of micro-motion feature groups. (e.g., amplitude variation of millimeter waves, mean movement of infrared hot spots, root mean square of pressure fluctuations), spatial trajectory parameters (e.g., total trajectory length 7.5 meters, average speed 0.47 m / s, number of dwell points 2, maximum turning angle 0.82 radians), time series statistical parameters (e.g., total dwell time 112 seconds, 4 events, average interval 41 seconds, first entry time 135 seconds). After standardization, each component is normalized and falls into... Interval. For example, after amplitude normalization and synthesis: Subsequently, based on correlation analysis, turning angle components that were highly linearly correlated with spatial attributes but contributed nothing to classification were removed, reducing the number of synthesized features to 10. Finally, through label embedding, the spatial region number 4, the sensor combination label (millimeter wave + infrared + pressure), and the adjustable material weight encoding were superimposed to generate a complete multi-dimensional collaborative feature vector. Applying this vector to a historical fingerprint database for comparison, human activity and non-human abnormal trajectories can be efficiently distinguished, achieving the high-precision modeling input requirements for anomaly filtering in intelligent human perception and detection systems.

[0090] S3.5: Perform consistency checks and outlier removal on the multidimensional structured feature vectors. Use outlier detection algorithms to remove noise or extreme attributes to ensure the stability and robustness of the input features in the subsequent spatiotemporal behavioral fingerprint modeling stage, and output a high-quality multidimensional structured feature vector for final modeling.

[0091] Step S4: Based on the multi-dimensional feature vector, a spatiotemporal behavioral fingerprint database of the human body is constructed offline or online using machine learning algorithms, and regionally differentiated behavioral feature models are established for different spatial regions and installation materials. For example... Figure 3 As shown, it specifically includes: S4.1: Spatial region grouping is performed on the preprocessed multidimensional feature vectors. Using spatial region identifiers and installation material categories as indices, a subset of discriminative behavioral features is obtained, providing raw input for the subsequent construction of a regionally differentiated behavioral feature model.

[0092] The input is a high-quality, multi-dimensional structured feature vector processed by the S3 step and passing the consistency check, carrying key label information, including spatial region identification and sensor installation material category.

[0093] A spatial label decoding method (parameters: spatial block number, physical location coordinate system, spatial grid division granularity) is adopted to classify each feature vector according to its corresponding spatial region identifier, thereby realizing primary feature grouping based on location.

[0094] Furthermore, by using an installation material attribute extraction algorithm (parameters: installation substrate material type, wall reflectivity, ground damping coefficient), feature vectors belonging to the same area but with different material environments are grouped a second time, so that the feature vectors under spatial area grouping can be further refined into subsets of the same material.

[0095] Furthermore, based on the spatial region-material two-level grouping results, a hierarchical index construction algorithm (parameter: region-material multi-level index tree structure) is adopted to generate a feature index structure with "region identifier-material category" as the primary key, thereby improving the efficiency of subsequent model clustering and retrieval.

[0096] Furthermore, feature screening and discriminative measurement algorithms (parameters: Mahalanobis distance, mutual information score, principal component contribution rate threshold) are applied to automatically screen feature components with high discriminative power within each region-material group, and remove highly redundant or weakly discriminative samples to form a subset of highly informative and discriminative behavioral features.

[0097] Through the above chain derivation, the original multidimensional structured feature vectors are grouped and filtered layer by layer according to spatial regions and material categories, outputting a region-material behavior feature subset with clear indexes and high discriminativeness. This provides an accurate input benchmark for the subsequent construction of unsupervised regional differential behavior models, realizing adaptive regional recognition and robust human behavior modeling in multi-material environments.

[0098] S4.2: The feature normalization algorithm is used to normalize the multidimensional feature vectors under the region grouping to eliminate the amplitude scale differences caused by different regions and installation materials, and to ensure that the feature vectors processed by the machine learning algorithm are homogeneous in the spatial domain and material domain.

[0099] The input is a high-dimensional, multi-source structured behavioral feature subset after spatial region grouping and material category indexing in the S4.1 sub-step. Each feature vector carries spatial region identifier, installation material category, and corresponding behavioral attribute parameters.

[0100] A feature normalization algorithm (parameters: mean μ, standard deviation σ, minimum min, maximum max, normalization interval [0,1], region group index, material group index) is employed. Based on all feature components within each region-material group, amplitude scale normalization is performed on each feature dimension to achieve distribution consistency correction. Furthermore, Z-score normalization is used (parameters: original value of single feature component within the group). mean Standard deviation For each feature component, the following normalization calculation is performed: in, This represents the original value of the i-th sample on the j-th feature component. This is the mean of the feature component within the group. Standard deviation, This is the result of Z-score normalization.

[0101] Furthermore, if a specific behavioral feature component exhibits extreme outliers or heterogeneous physical quantity ranges, a Min-Max normalization algorithm is employed (parameter: minimum value of a single feature component within a group). Maximum value Normalized interval [a, b]), amplitude scaling is performed according to the following formula: in, These are the feature component results after Min-Max normalization. This represents the original value of the i-th sample on the j-th feature component.

[0102] Furthermore, for multimodal feature linkage under different region and material groups, a grouping and segmentation normalization mechanism is adopted (parameters: normalization strategy selection, feature weight factor). Local distribution normalization correction is performed on each spatial region-material subset, and then the embedding is unified through a feature reweighting method. For the same physical quantity type under different region-material groupings, a unified feature domain dimension is used to ensure that the unsupervised clustering algorithm can perform feature similarity discrimination on a unified scale.

[0103] Furthermore, for some highly discriminative feature components, quantile normalization is performed based on historical statistical distributions (parameter: group distribution quantiles). By normalizing the P-th quantile scale, extreme diffusion scenarios are corrected, thereby improving the discriminative comparability of each feature component in scenarios with abnormal spatial distribution or specific material reflection gain.

[0104] Through the above chain-like normalization process, the multidimensional structured feature vectors under the original region-material grouping are all transformed into a standard feature dataset that is evenly distributed, scaled, and comparable across regions. This provides a data benchmark for the unified modeling of unsupervised regional differential behavior clustering models and achieves homogeneity calibration across space and materials.

[0105] For example, in an open-plan office setting with an area of ​​120 square meters, three types of sensors—millimeter wave, infrared, and pressure—are deployed in Area 1 (glass partition) and Area 2 (plasterboard wall), respectively. After processing in the preceding steps, Area 1-glass group contains 170 feature vectors, each containing 12 feature components (such as micro-motion amplitude, trajectory length, and dwell time); the micro-motion amplitude component of its original data ranges from [0.12, 0.87], and the dwell time component ranges from [10, 214] seconds; Area 2-plasterboard group has 112 feature vectors, with micro-motion amplitude components ranging from [0.19, 0.98].

[0106] In this embodiment, following the grouping normalization process, for region 1-glass, the mean value is calculated for the micro-motion amplitude component. Standard deviation After Z-score normalization, all components converge to the interval [-1.6, 1.7]. For the residence time component, the minimum value is 10 and the maximum value is 214. The Min-Max normalization is to [0,1], where 20 seconds corresponds to 0.05 after normalization and 180 seconds corresponds to 0.85. The pressure change component and other components are corrected for distribution tails according to quantile normalization.

[0107] For Region 2 - the gypsum board group, Z-score and Min-Max normalization of the micro-motion amplitude components are performed simultaneously and independently, and the extreme value distribution is corrected by grouping and segmenting scale. Finally, after chain-like normalization, the entire region-material feature subset achieves a mean close to 0, a standard deviation of 1, and a scale in the [0,1] interval across all 12 feature components, significantly improving data consistency across groups. In this embodiment, after normalization processing, the input to the unsupervised clustering algorithm (K-means, Isolation Forest) improves the behavior pattern discrimination accuracy from 78.4% under the original data to 91.2%. The cross-region clustering boundaries are clear, and there is no cross-drift under different material environments, significantly improving the accuracy and robustness of regional behavior fingerprint modeling, achieving the key technical effect of constructing a high-precision human behavior feature database.

[0108] S4.3: Based on the normalized multidimensional feature vector, call the unsupervised clustering algorithm (such as isolated forest or homogeneous clustering algorithm) to perform behavioral pattern clustering on the data under each spatial region and material category, generate regional difference behavioral feature centers, and form a highly discriminative prototype cluster.

[0109] The input is a set of high-dimensional structured feature vectors of region-material groups after normalization processing in sub-step S4.2. It covers the target micro-motion parameters, spatial trajectory attributes and time series statistical features collected by multi-source sensors (millimeter wave, infrared, pressure, sound). Each data entry is labeled with spatial region and installation material.

[0110] An unsupervised clustering algorithm (parameters: clustering method type K-means, DBSCAN, isolated forest, discriminative feature dimension, initial number of clusters K=3~10, anomaly detection confidence threshold) is used to independently perform clustering operations on the normalized feature subsets under each spatial region and installation material category, thereby achieving automatic identification and similarity grouping of target behavior patterns.

[0111] Furthermore, an automatic cluster number selection mechanism (parameters: silhouette coefficient, Davies-Bouldin index, Calinski-Harabasz score) is used to evaluate the clustering effect of different K values ​​under the same region-material subset. The maximum silhouette coefficient method is used to select the optimal number of clusters so that the clustering results have the best cohesion and external separation.

[0112] Furthermore, for the isolated forest algorithm branch (parameter: number of trees N) estimators =100, Maximum sample proportion Max samples =1.0, anomaly scoring threshold Contamination=0.05), perform behavior density distribution modeling in high-dimensional feature space, automatically classify unique behavior samples into isolated classes based on the difference in abnormal path length, and assign other highly similar samples to the same cluster, thereby achieving automatic stripping of atypical behaviors.

[0113] Furthermore, a cluster center generation algorithm (parameters: feature mean, weighted centroid, Euclidean distance minimization rule) is applied to the clustered sample subsets to calculate the center vector of each group of high-information behavioral features, thereby obtaining the set of behavioral feature centroids under the region-material grouping.

[0114] Through the above chain-like clustering analysis steps, the distribution of behavioral features in each spatial region-material environment is effectively summarized into several high-resolution, highly distinctive behavioral feature prototype clusters, providing cluster centroid standards for the subsequent generation of regional behavioral fingerprint templates.

[0115] S4.4: Apply time-series modeling methods to perform statistical analysis on the appearance / disappearance order and micro-changes of behavioral feature centers obtained from each cluster, generate information records with spatiotemporal labels, and further enrich the dynamic dimension of the regional difference behavioral feature model.

[0116] S4.5: The clustering behavior feature center is integrated with the temporal information. Based on the behavior label and spatial attributes, it is automatically archived into spatiotemporal behavior fingerprint entries corresponding to the spatial region and installation material, and stored in the behavior fingerprint database to provide a standard template for subsequent real-time similarity matching.

[0117] S4.6: Perform consistency checks and model optimization on the entire spatiotemporal behavioral fingerprint database, including detecting feature redundancy, evaluating model cohesion, and removing noisy segments, to ensure the generalization ability and detection accuracy of the behavioral fingerprint database in multiple scenarios.

[0118] Step S5: The real-time acquired spatiotemporal behavioral feature vectors are matched with the established behavioral fingerprint database for similarity analysis. Based on the discriminative features, a confidence score for the presence of a human being is output for each detected instance. Specifically, this includes: S5.1: Based on the spatiotemporal behavior feature vectors acquired in real time, feature reduction algorithms (such as PCA and linear discriminant analysis) are called to reduce the dimensionality of the multidimensional feature vectors to obtain a highly discriminative spatiotemporal behavior feature subspace, which outputs a low-redundancy and highly discriminative feature expression for the similarity calculation process.

[0119] S5.2: Using the dimensionality-reduced, highly discriminative spatiotemporal behavioral feature subspace as input, cluster-based retrieval technology is employed to pre-screen the corresponding behavioral fingerprint sub-database for the target region to which the feature vector belongs, thereby achieving region-related retrieval in the feature space and effectively improving the response efficiency and accuracy of the subsequent similarity matching system.

[0120] The input is a real-time spatiotemporal behavior feature subspace vector after S5.1 feature reduction processing. Each vector has been reduced to a highly discriminative feature subspace and retains complete spatial region identification label attributes.

[0121] A cluster-based clustering center retrieval method (parameters: target spatial region identifier, behavioral fingerprint database partition index, K-means / isolated forest clustering center pointer table) is employed to perform targeted pre-screening of the behavioral fingerprint sub-database of the region to which the current feature vector belongs. This processing action parses the spatial region marker of the real-time detection instance, along with the additional installation material category parameter, and retrieves the cluster center set under the corresponding spatial region and material group from the global behavioral fingerprint database. The screening result is a specific fingerprint sub-database of the physical domain to which the detection instance belongs.

[0122] Furthermore, by using a spatial region attribute consistency matching algorithm (parameters: region ID function, material matching threshold, hierarchical grouping table), the spatial attributes of the feature vector and the primary key of the corresponding fingerprint sub-database are compared, filtering out all irrelevant fingerprint entries that are inconsistent with the spatial region of the detection instance, thus achieving dual constraints on the granularity of data retrieval in both the spatial and material domains.

[0123] Furthermore, based on the multidimensional distribution of highly discriminative spatiotemporal feature subspaces, a feature index mapping table (parameters: dimensionality reduction feature mapping matrix, fingerprint feature space encoder) is used to automatically map real-time spatiotemporal feature vectors to the target region behavioral fingerprint subspace, significantly reducing the set of samples to be matched and effectively improving the similarity calculation speed and retrieval resource utilization of the online system.

[0124] Furthermore, the regional cluster center feature retrieval algorithm (parameters: centroid vector set, spatial-material label, feature retrieval depth) is invoked to perform multi-level cluster category aggregation on the selected fingerprint sub-database. The fingerprint centers required for similarity discrimination are further aggregated according to cluster category, and the pool of fingerprint centers to be matched is output, providing a targeted and low-redundancy input basis for the subsequent similarity scoring process.

[0125] Through the above-mentioned multi-layered nested cluster category retrieval and spatial region index filtering, the dimensionality-reduced feature subspace is mapped in real time to a target-related efficient behavioral fingerprint sub-library, thereby greatly improving the efficiency of fingerprint query and the accuracy of online similarity scoring in real-time detection scenarios, and realizing a highly robust human behavior instance fingerprint distribution mechanism in cross-regional and cross-material environments.

[0126] For example, in an intelligent lighting system for an office building, three types of environmental sensors—millimeter-wave, infrared, and pressure—are deployed in four sub-regions (A, B, C, and D) and two material environments (glass and gypsum board), collecting over 8,000 high-dimensional multimodal feature data points daily. After PCA dimensionality reduction, the dimension of a single feature vector is compressed from 12 dimensions to 5 dimensions. In a patrol-type detection example, assume the current feature vector v corresponds to spatial region B-glass. The system's behavioral fingerprint database is organized according to a region-material index tree, and region B-glass contains N=4 cluster centers obtained through isolated forest clustering. The cluster center retrieval process first parses the spatial region and material label of v, compressing the search scope from 346 fingerprint centers in the entire database to only 4 centers in this region, focusing the retrieval granularity on highly relevant target classes. For the projection of v into the 5-dimensional feature subspace of region B-glass, the feature index mapping table is called to map it to the 5-dimensional subspace of the required fingerprint center. In actual comparison, only 4 samples need to be accessed, improving retrieval efficiency by over 86%. The output result will be directly used as the fingerprint similarity score input in step S5.3, enabling low-latency, high-precision real-time detection of human behavior, and optimizing the subsequent false alarm discrimination rate and model response performance.

[0127] S5.3: For the behavioral fingerprint sub-database retrieved from the target region, the weighted cosine similarity algorithm is used to perform a one-to-one similarity matching operation between the current real-time spatiotemporal behavioral feature subspace vector and the fingerprint feature set to obtain the similarity distribution between the detected instance and each fingerprint.

[0128] The real-time spatiotemporal behavioral feature subspace vectors, which are dimensionality-reduced by step S5.2 and filtered by spatial region indexing, will be used as input for the similarity scoring operation, and will be located in the behavioral fingerprint sub-library of the physical space and installation material environment to which the target detection instance belongs.

[0129] The weighted cosine similarity algorithm is adopted (parameter: real-time spatiotemporal behavioral feature subspace vector). The set of center vectors of the fingerprint sub-database is C={c1,c2,...c n Feature component weight vector This enables similarity matching between the feature subspace vector and the center vectors in the fingerprint sub-database.

[0130] S5.4: Based on the similarity distribution results, the human presence confidence score of the detected instance is calculated using a discriminative confidence weighted fusion algorithm (Gaussian distribution evaluation or KNN confidence regression can be selected). The output value reflects the degree of similarity between its real-time behavior and historical human fingerprint feature clusters.

[0131] S5.5: Combine the obtained human presence confidence score with the feature attribution label of the detected instance (including spatial region label, sensor type, and environmental context) to generate the final human presence confidence judgment result, which serves as the input basis for the downstream adaptive anomaly weighted discrimination module, realizing a closed-loop link of human detection, anomaly filtering, and system linkage.

[0132] Step S6: For detection instances with confidence scores below a preset threshold or exhibiting highly anomalous distributions, adaptive anomaly weighting is performed in conjunction with multidimensional contextual data (including door magnetic opening / closing status, pressure surges, historical energy consumption characteristics, and regional attribute labels). Specifically, this includes: S6.1: Use the confidence score output by the real-time detection instance as an anomaly detection trigger signal to automatically filter detection instances with confidence scores below a preset threshold or significantly deviating from the distribution of human behavioral characteristics, so as to obtain a set of high-risk detection instances that need further judgment.

[0133] S6.2: For a set of high-risk detection instances, an event-driven model is applied based on the door magnetic monitoring status data at the time of the instance occurrence to analyze the synchronous door magnetic opening and closing signals, use the door magnetic status change as an anomaly discrimination weighting factor, generate a door magnetic context feature vector, and enhance the discrimination weight of spatial access dynamic information.

[0134] The input data consists of a set of high-risk detection instances and their original door magnetic monitoring status data at the time of occurrence. The instance set has been filtered by the previous step (S6.1) and has corresponding spatiotemporal labels and confidence information. Each instance needs to establish an event association with the door magnetic signal at the time of occurrence.

[0135] An event-driven model (parameters: door magnetic sensor status history sequence, signal timestamp, spatial region mapping table) is adopted to synchronously analyze the door magnetic sensor monitoring status signal at the time of occurrence of high-risk detection instances, and obtain the baseline information of the spatial access status at that moment.

[0136] Furthermore, by using a door magnetic signal analysis algorithm (parameters: open / closed state encoding, time window span, and anomaly marking rules), event recognition and period matching are performed on the door magnetic signal sequence, the door magnetic opening and closing change event sequence under the corresponding detection instance index is extracted, and the event synchronization interval is generated.

[0137] Furthermore, a door magnet context feature encoding method (parameters: event type vector, opening and closing duration, time difference from the nearest door magnet flip, door magnet event frequency statistics) is adopted to establish vectorized feature encoding for the extracted door magnet opening and closing events, thereby mapping the door magnet state changes into door magnet context feature vectors with multi-dimensional semantics.

[0138] Furthermore, by using a door magnetic sensor weighting factor calculation algorithm (parameters: door magnetic sensor event type weight table, spatial area gate control priority, time decay coefficient), the door magnetic sensor weighting components in the anomaly discrimination process are dynamically adjusted based on the door magnetic sensor context feature vector, thereby increasing the weight of the door magnetic sensor context in high-risk event discrimination and realizing differentiated weighted processing of dynamic spatial access behavior.

[0139] By using a door and magnetic sensor context feature fusion method (parameters: confidence score, event co-operation coefficient, multimodal information integration matrix), the door and magnetic sensor context feature vector is integrated with the original detection instance features to form an enhanced feature input, providing a refined context input with spatial access identifiers for the adaptive anomaly weighted discrimination process.

[0140] Through the aforementioned algorithm chain, the door magnetic monitoring data of the high-risk detection instance set is transformed into door magnetic context feature vectors, and the information expression of spatial access dynamic events is improved in the weighted discrimination process, thereby upgrading the discrimination sensitivity of non-human behavior anomaly detection.

[0141] S6.3: Based on the time series data of pressure sensors associated with high-risk detection instances, the abrupt change detection algorithm is used to perform anomaly weighted modeling on the magnitude and frequency of pressure changes, forming a pressure abrupt change context feature quantity that is highly coupled with the detection instance, and incorporating it into the weighted discrimination process to improve the sensitivity of false alarm discrimination.

[0142] S6.4: Combining the historical energy consumption characteristics of the spatial area corresponding to the detection instance, the energy consumption behavior pattern modeling method is adopted to extract instantaneous abnormal energy consumption fluctuations, periodic abnormal energy consumption, and features related to human activities. These features are then transformed into historical energy consumption context vectors and connected to the anomaly weighted discrimination process to achieve multi-dimensional discrimination at the coupling level of behavior and energy consumption.

[0143] S6.5: For detected instances belonging to different spatial regions and building attributes, based on pre-set regional attribute labels and scene metadata, a regional difference analysis algorithm is used to determine the weighting coefficients, transforming regional importance, susceptibility to disturbance, and historical abnormal distribution into regional attribute feature weights, thus supplementing the contextual discrimination information.

[0144] S6.6: Summarizes the door magnetic context feature vector, pressure mutation context feature, historical energy consumption context vector and regional attribute feature weights, and applies an adaptive weighted fusion algorithm based on the multimodal feature set to perform multidimensional anomaly discrimination modeling on high-risk detection instances, outputs anomaly type discrimination results and weighted anomaly scores, and achieves accurate non-human behavior feature recognition and dynamic false alarm filtering triggering.

[0145] Step S7: When the adaptive anomaly weighting discrimination result is a non-human behavior feature, short-term high-density multi-sensor modal redundancy detection is automatically triggered, and the newly acquired multi-modal behavior data is further compared with the regional behavior feature model. Specifically, this includes: S7.1: Based on the non-human behavior feature labels output by adaptive anomaly weighted discrimination, a short-time high-density sampling mechanism is automatically activated for the current spatiotemporal behavior feature vector to shorten the acquisition cycle and significantly improve the data acquisition frequency, thereby obtaining the new cycle of multi-source sensor spatiotemporal behavior raw data.

[0146] S7.2: For the new cycle of multi-source sensor spatiotemporal behavior raw data, a multi-channel synchronous calibration algorithm is used to align data timestamps and map spatial coordinates to ensure the consistency of various sensor data in the spatiotemporal dimension and form a standardized multimodal synchronous dataset.

[0147] S7.3: Based on the standardized multimodal synchronous dataset, a multimodal feature fusion algorithm is applied to extract micro-motion change features, spatial trajectory patterns, and appearance / disappearance time-series statistics to form a multi-dimensional fusion feature vector, in order to fully reflect the multi-source performance of abnormal interference features.

[0148] The input data is a standardized multimodal synchronous dataset obtained through multi-channel synchronous calibration, which includes high-frequency sampling signals and their spatial-temporal labels from various types of sensors such as millimeter wave, infrared, pressure, and sound.

[0149] A multimodal signal decoupling method (parameters: channel separation coefficient, noise reduction weight) is adopted to decouple the data streams of each sensor type, so as to achieve low-interference separation of their respective signal channels in the prototype feature space.

[0150] Furthermore, by using a micro-motion feature extraction algorithm (parameters: short-time sliding window length, time-domain micro-motion sensitivity threshold, frequency-domain energy extraction bandwidth), multi-scale detection of minute motion changes in millimeter-wave echoes, infrared thermal data, and pressure response sequences is performed to obtain multi-channel intensity change features reflecting the micro-motions on the surface and inside of abnormal targets.

[0151] Furthermore, a spatial trajectory modeling method (parameters: spatial coordinate sampling interval, trajectory smoothing window, curvature clustering threshold) is adopted to reconstruct the trajectory of the spatial location data sequence after multi-channel separation, extract the spatial motion path, dwell area and activity trajectory pattern, and form a spatial trajectory index for abnormal interference scenarios.

[0152] Furthermore, by using the occurrence / disappearance event detection and time series statistics algorithm (parameters: dynamic threshold, adaptive event counting window), the start and end points, duration period, and intermittent occurrence events of the target signal in each channel time series are labeled and statistically analyzed to generate target occurrence / disappearance time series, specific state dwell time, and periodic frequency statistics.

[0153] Furthermore, a multimodal feature fusion algorithm (parameters: weighted fusion coefficient, feature normalization model, channel importance adjustment matrix) is applied to construct a unified multidimensional fusion feature vector by employing a feature-level fusion mechanism for the aforementioned micro-motion change features, spatial trajectory patterns, and appearance / disappearance statistics.

[0154] By using multimodal feature fusion algorithms or processing methods, the spatiotemporal heterogeneous features obtained in the previous step are integrated and transformed into a fusion feature vector that includes multimodal and multi-temporal information, thereby achieving a full characterization and robustness improvement of the target's multi-source performance under various abnormal interferences.

[0155] For example, a multimodal sensor array of millimeter wave, infrared, pressure, and sound sensors deployed in conference room A is configured with a sampling period of 200ms, a sliding window of 1s for micro-motion feature extraction, a trajectory reconstruction sampling interval of 30cm, and a temporal event detection threshold of the dynamic mean plus twice the standard deviation. The millimeter wave channel detected short-term micro-motion peaks with a standard deviation greater than 3; the infrared thermal channel showed stable output without periodic enhancement within 10 seconds; the pressure channel exhibited a step event that persisted for 5 seconds; and the sound channel showed no significant frequency band activity. Spatial trajectory modeling was applied, and the trajectory pattern distribution showed localized small-range oscillations. The thermal channel showed dwell times concentrated in the 3-7 second range, and the disappearance event interval exceeded the normal range for human activity. In the multimodal feature fusion stage, weighting coefficients [millimeter wave 0.4, infrared 0.2, pressure 0.2, sound 0.2] were used. After normalization of each channel, the vectors were concatenated to form a total of 15-dimensional fused feature vectors. The final output multi-dimensional fused feature vector significantly deviated from historical human fingerprint features, providing a comprehensive and structured data foundation for subsequent anomaly feature similarity discrimination and refined false alarm filtering.

[0156] S7.4: Based on the extracted multi-dimensional fusion feature vector, a multi-modal feature similarity discrimination algorithm of regional behavior feature model is adopted to calculate the feature distance or similarity score between the current behavior instance and each regional behavior feature model, and form a behavior feature similarity matrix.

[0157] S7.5: Based on the behavioral feature similarity matrix results, execute the anomaly differentiation decision process, statistically judge the similarity distribution characteristics between multiple collection results and the regional behavioral feature model, determine whether it is a typical non-human behavioral anomaly instance, and output a high-confidence anomaly event judgment label for use in the subsequent false alarm filtering process.

[0158] Step S8: If the redundant detection result still differs from the human behavior fingerprint database features after multiple samplings, the current detection is automatically determined to be an abnormal interference source, false alarm filtering is performed, and the abnormal data and its features are stored in the database for model self-learning. If the redundant detection result is the same as the human behavior fingerprint database features, it is determined that a human body is present. Specifically, this includes: S8.1: For the multimodal behavioral feature data output by the multi-sensor modal redundancy detection module, perform multiple rounds of similarity comparison based on the regional behavioral feature model to determine the degree of fit between the behavioral features and the human behavioral fingerprint database, and obtain the similarity distribution parameters between the behavioral features and the human behavioral fingerprint database.

[0159] S8.2: Utilizing the similarity distribution parameters of spatiotemporal behavioral features from multiple rounds, and employing anomaly detection algorithms (such as confidence threshold discrimination and local outlier factor), calculate the confidence sequence of human behavior discrimination for detected instances under repeated sampling, in order to quantify the continuous loss of human features in detected instances.

[0160] S8.3: For detection instances where the confidence sequence is consistently below the anomaly detection threshold, a false alarm judgment mechanism is automatically triggered, defining the current detection instance as an abnormal interference source and generating a false alarm judgment label.

[0161] The input data consists of spatiotemporal behavioral feature fusion vectors obtained through multi-round redundant multimodal detection and confidence sequences generated after comparison with the human behavioral fingerprint database.

[0162] S8.4: Employ a false alarm filtering module to perform false alarm signal shielding operations on abnormal interference source instances with false alarm judgment tags, thereby eliminating abnormal detection signals and closing the loop of the energy consumption control link.

[0163] S8.5: The abnormal interference source behavior feature data marked by false alarm filtering is archived into the abnormal data feature library according to the formatting rules, and used as the input of the model adaptive learning module to drive the subsequent optimization and dynamic evolution of the spatiotemporal behavior fingerprint library and the regional behavior feature model.

[0164] Step S9: After the system receives the annotation information from the user or the operation and maintenance end, it automatically updates the spatiotemporal behavioral fingerprint database and optimizes the anomaly detection algorithm and the regional difference behavioral feature model in real time to dynamically improve detection robustness and generalization ability. Specifically, this includes: S9.1: Receive behavioral data annotation information from users or operations and maintenance personnel, and perform data calibration processing on the original spatiotemporal behavioral feature vector set based on the feedback information to obtain a high-quality spatiotemporal behavioral feature sample library with labels.

[0165] S9.2: Apply homogeneous clustering and anomaly detection algorithms to the calibrated spatiotemporal behavioral feature sample library to recalculate the core spatiotemporal behavioral fingerprint of the human body and its unique cluster centers, thereby achieving adaptive parameter updates of the spatiotemporal behavioral fingerprint model.

[0166] S9.3: Based on the latest spatiotemporal behavioral fingerprint model, transfer learning is applied to the regional difference behavioral feature model of each spatial region and installation material environment to obtain behavioral feature parameters that are highly adapted to regional attributes.

[0167] S9.4: Perform incremental machine learning on the anomaly detection algorithm, incorporating the features of the latest anomaly data into the detection boundary conditions to generate an anomaly detection parameter set that reflects the latest distribution of interference scenarios.

[0168] S9.5: Based on the incrementally updated anomaly discrimination parameter set, the discrimination process of the machine learning inference module is fine-tuned online, dynamically adjusting the confidence score of human presence and the false alarm trigger threshold, and outputting detection model parameters that meet the new working conditions.

[0169] S9.6: The updated spatiotemporal behavioral fingerprint model and anomaly discrimination parameters are synchronized to the global fingerprint database and the regional difference model database, serving as the latest execution benchmark for subsequent human presence detection and anomaly filtering discrimination algorithms, thus realizing the system's self-learning closed loop.

[0170] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0171] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0172] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A human perception detection method based on machine learning, specifically including: S1: Collect raw spatiotemporal behavioral data from environmental sensors of different locations and types, and label the sampled data with the corresponding spatial regions and installation material categories; S2: Standardize and preprocess the collected raw spatiotemporal behavioral data; S3: Extract target micro-motion change features, spatial trajectory sequence, and appearance / disappearance time sequence statistics from preprocessed spatiotemporal behavior data, and form a multi-dimensional feature vector; S4: Based on the multidimensional feature vector, construct a spatiotemporal behavioral fingerprint database of the human body through machine learning algorithms, and establish regionally differentiated behavioral feature models for different spatial regions and installation materials; S5: Perform similarity matching between the spatiotemporal behavioral feature vectors collected in real time and the established behavioral fingerprint database, and output the human presence confidence score for each detected instance based on the discriminative features; S6: For detection instances with confidence scores below a preset threshold or exhibiting highly anomalous distributions, perform adaptive anomaly weighting discrimination by combining multidimensional contextual data; S7: When the adaptive anomaly weighted discrimination result is a non-human behavior feature, short-term high-density multi-sensor modal redundancy detection is automatically triggered, and the newly collected multi-modal behavior data is further compared with the regional behavior feature model; S8: If the redundant detection results still differ from the characteristics of the human behavior fingerprint database after multiple samplings, the current detection will be automatically determined as an abnormal interference source, and false alarm filtering will be performed.

2. The human perception detection method based on machine learning according to claim 1, characterized in that, Following step S8, the following is also included: Step S9: After the system receives the annotation information from the user or the operation and maintenance end, it automatically updates the spatiotemporal behavior fingerprint database and optimizes the anomaly detection algorithm and the regional difference behavior feature model in real time.

3. The human perception detection method based on machine learning according to claim 1, characterized in that, The environmental sensors include millimeter-wave sensors, infrared sensors, pressure sensors, and sound sensors.

4. The human perception detection method based on machine learning according to claim 1, characterized in that, The standardization preprocessing in step S2 includes removing static background noise and normalizing the amplitude distribution of each sensor channel.

5. The human perception detection method based on machine learning according to claim 1, characterized in that, The multidimensional context data in step S6 includes: door magnetic opening and closing status, pressure surge, historical energy consumption characteristics, and regional attribute labels.

6. The human perception detection method based on machine learning according to claim 2, characterized in that, Step S1 specifically includes: A unified data acquisition initialization process is executed for various environmental sensors deployed in the target space, and a sampling clock synchronization mechanism is configured to obtain raw sensor data streams containing timestamps; Based on the above raw sensor data stream, the data access interface is called to collect the data output of each sensor according to the preset physical location code, and the spatial location label is mapped one by one with the raw data. The parameters of key areas such as the sensor mounting base, wall / ground material, and obstruction structure are digitally collected using an engineering identification entry mechanism and linked to the data tags of each sampling point; A unified data structure and encapsulation format is established for various types of raw sensor data. Through protocol conversion and multi-source fusion interfaces, it is ensured that different types of data can be normalized and integrated into a unified data acquisition base. Based on the actual sampling period and preset trigger events, the data acquisition frequency and period are dynamically adjusted, and the original spatiotemporal behavioral data at key sampling moments are streamed, cached, and persistently stored.

7. The human perception detection method based on machine learning according to claim 6, characterized in that, The various types of raw sensor data include: millimeter-wave echo values, infrared thermal data, pressure response curves, and acoustic amplitude / spectrum.

8. The human perception detection method based on machine learning according to claim 2, characterized in that: The system has the ability to archive abnormal events and features, and will label and format multimodal features that are determined to be non-human abnormalities for storage, so as to provide adaptive optimization of the model and subsequent generalization ability improvement.

9. The human perception detection method based on machine learning according to claim 1, characterized in that: The redundant high-frequency multi-channel re-detection and closed-loop false alarm filtering module can automatically increase the sampling frequency when an anomaly is detected, collect multiple multi-source synchronous signals and perform multiple rounds of feature fusion and fingerprint feature comparison. If the confidence level is continuously lower than the threshold, a false alarm judgment is output.

10. A human perception detection system based on machine learning, characterized in that, The presence of a human body is detected using the human perception detection method based on machine learning as described in any one of claims 1-9.

Images