A csi-based target spatial object and spatial state perception method and system
By constructing a wireless fingerprint of an vacant environment based on the CSI method and using a student time series model, the problems of privacy risks, environmental robustness and recognition accuracy in indoor space perception are solved, and efficient, low-cost and refined spatial status monitoring under contactless perception is achieved.
Patent Information
- Application Number
- CN202611005453.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-07
- Publication Date
- 2026-11-06
AI Technical Summary
Existing technologies for indoor space perception suffer from privacy risks, poor environmental robustness, coarse-grained state recognition, unstable recognition results, and high deployment costs, failing to address multiple industry pain points simultaneously.
A target spatial object and spatial state perception method based on CSI is adopted. By constructing a wireless fingerprint of the vacant environment and combining multi-dimensional CSI time series features and student time series model, a refined judgment of spatial state is achieved. Channel quantization index, model confidence and wireless quality metadata are integrated to output standardized spatial perception events.
It achieves contactless sensing, reduces the risk of privacy leakage, adapts to various environmental conditions, reduces hardware and maintenance costs, can finely distinguish spatial states, improves recognition accuracy and robustness, and is suitable for scenarios such as smart spaces, industrial safety and elderly care.
Smart Images
Figure CN123049232A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of target spatial object and spatial state perception, specifically relating to a target spatial object and spatial state perception method and system based on CSI. Background Technology
[0002] Indoor space object and space status perception is widely used in smart security, smart home, building monitoring and other scenarios. Existing perception methods mainly include two types: visual camera perception and CSI wireless perception. Both of them have technical defects that make it difficult to achieve stable monitoring in multiple scenarios.
[0003] Camera-based spatial perception relies on optical imaging to monitor space. The device continuously collects image and video data, which poses a serious risk of privacy leakage and is not suitable for private indoor areas. The perception effect is highly dependent on ambient light. The recognition performance drops sharply in low light, strong light, and backlight scenes, and furniture, people, and clutter can easily create blind spots. At the same time, the cost of hardware procurement, wiring, and maintenance is high, and large-scale long-term deployment is not economically viable.
[0004] Existing CSI indoor passive sensing solutions, such as CN116168499B, mostly use CSI amplitude stability discrimination or fixed threshold detection logic, which can only roughly determine whether there is signal disturbance in the space. They do not have the ability to distinguish fine state, and cannot effectively distinguish four different spatial states: empty background, low-motion lingering of target object, target activity, and slowly changing environmental interference such as temperature and humidity. The state recognition granularity is coarse and the discrimination is insufficient.
[0005] Existing CSI passive sensing or human activity recognition models generally adopt a single-window independent output mechanism. In complex indoor wireless environments, interference factors such as random fluctuations in wireless signals, relocation of transceiver equipment, adjustment of sampling parameters, and long-term drift of indoor background can cause frequent jitter in the model output results, resulting in a high probability of false detection and false negative, and insufficient robustness for long-term continuous monitoring.
[0006] Therefore, existing technologies have not yet built a complete closed-loop perception system, and only remain at the level of single disturbance detection. They cannot simultaneously solve multiple industry pain points such as privacy risks, poor environmental robustness, weak state subdivision capability, unstable recognition results, and high deployment costs. Summary of the Invention
[0007] To address the technical problems mentioned above, this invention provides a method and system for sensing target spatial objects and spatial states based on CSI.
[0008] This invention adopts the following technical solution: a target spatial object and spatial state perception method based on CSI, comprising the following steps: Establish a wireless link within the target space to continuously collect raw CSI I / Q data; The raw CSI I / Q data is preprocessed to obtain a processed CSI frame sequence, and radio quality metadata is extracted simultaneously; an adapted timing window is generated as required, and multi-dimensional CSI timing features corresponding to each timing window are extracted; The process of building or updating the wireless fingerprint of the vacant environment is based on the wireless quality metadata, the target space vacancy confirmation information, and the multi-dimensional CSI timing characteristics within the timing window. If the stable CSI window determination condition is met, the process of building or updating the wireless fingerprint of the vacant environment is executed. The channel quantization index is obtained by calculating the difference between the multi-dimensional CSI timing characteristics of the current timing window and the wireless fingerprint of the idle environment. Input the multi-dimensional CSI time series features of the current time series window into the pre-trained CSI student time series model, and output the probabilities of various spatial states and the model confidence. The system integrates channel quantization metrics, various spatial state probabilities, model confidence, and radio quality metadata, and outputs standardized spatial awareness events that match the duration of the time window after time-series smoothing.
[0009] In a further embodiment, the process for obtaining the processed CSI frame sequence is as follows: The raw CSI I / Q data is parsed to obtain a CSI frame sequence consisting of several metadata elements, and wireless quality metadata is extracted and generated based on these metadata elements. The parsed CSI frame sequence is sequentially subjected to abnormal frame filtering, time sorting, and sampling quality detection to obtain the intermediate CSI frame sequence. A dimension consistency check is performed based on the intermediate CSI frame sequence, wherein the dimension consistency check is a check of the input tensor of the intermediate CSI frame sequence; If the dimension consistency check fails, channel component adaptation is performed on the intermediate CSI frame sequence to obtain the processed CSI frame sequence.
[0010] In a further embodiment, the timing window is a fixed window or an adaptive window, and the corresponding timing window selection method is as follows: Both the fixed window and the adaptive window can be applied to the process of constructing or updating wireless fingerprints of object states and vacant environments in real time; The adaptive window is applied to scenarios where sampling rate fluctuations, frame loss, or target space object state switching occur during data acquisition or data processing.
[0011] In a further embodiment, the stable CSI window determination condition is: When constructing a wireless fingerprint in an empty environment for the first time, the effective frame rate, maximum disconnection duration, subcarrier dimension and link quality corresponding to the wireless quality metadata within the time window all meet the corresponding preset standards, and the short-term motion intensity is lower than the corresponding threshold. After an empty environment wireless fingerprint is established, it is further required that the probability of the target object's existence, the short-term motion intensity and the background residual value within the time window are all lower than the corresponding threshold. Correspondingly, the vacant environment wireless fingerprint includes at least one or more of the following: vacant channel shape, gain reference value, noise scale, residual distribution, or motion distribution.
[0012] In a further embodiment, the channel quantization metric further includes a gain offset metric; The channel quantization metrics include at least one or more of the following: background residual value, short-time motion intensity, baseline drift value, radio quality quantization value, and gain offset metric.
[0013] In a further embodiment, the training process of the CSI student time series model is as follows: Acquire spatial annotation information and align the spatial annotation information to the temporal window; the spatial annotation information is teacher modal labeling or manual annotation labeling. Using aligned spatial annotation information, a CSI student temporal model is trained based on a temporal network; the trained CSI student temporal model outputs a temporal multi-class classification.
[0014] In a further embodiment, the spatial state includes at least: a spatial vacancy state, a relatively static existence state of the target object, and a dynamic activity state of the target object; The corresponding spatial state probabilities are calculated for each spatial state, and all spatial state probabilities are integrated to form a spatial state probability distribution feature. Based on the spatial state probability distribution characteristics, the corresponding model confidence calculation method is selected to obtain the model confidence.
[0015] In a further embodiment, the information fields of the standardized spatial perception event include at least basic fields, perception results, environmental evidence, and output purpose; The basic fields include at least the space identifier, device identifier, timestamp, and space status type; The perception results include at least the occupancy status, activity category, model confidence, and activity intensity; The environmental evidence includes at least background residual values, short-term motion intensity, vacant baseline status, and wireless quality quantization values.
[0016] In a further embodiment, the spatial identifier is generated by the target spatial numbering information synchronously recorded when acquiring raw CSI I / Q data; The activity intensity and idle baseline state are calculated by the channel quantization index, and the idle baseline state is used to characterize the effectiveness of wireless fingerprinting in idle environments. The health status of the device is determined by a combination of wireless quality metadata and wireless quality quantification values.
[0017] A target spatial object and spatial state perception system based on CSI, used to implement the target spatial object and spatial state perception method described above, includes: At least one wireless signal transmitter and at least one wireless signal receiver are deployed in the target space according to preset requirements to form a wireless link, and the wireless link covers the activity area in the target space. The CSI acquisition and quality assessment module is used to continuously acquire raw CSI I / Q data and synchronously record and generate wireless quality metadata. The preprocessing and windowing module is configured to preprocess the raw CSI I / Q data to obtain a processed CSI frame sequence, generate an appropriate time series window as required, and extract the multi-dimensional CSI time series features corresponding to each time series window. The idle environment fingerprint module, based on wireless quality metadata, determines the multi-dimensional CSI timing characteristics within the timing window. If the stable CSI window determination conditions are met, the idle environment wireless fingerprint construction or update process is executed. The channel quantization module is used to calculate the difference between the multi-dimensional CSI time-series features of the current time-series window and the wireless fingerprint of the idle environment to obtain the channel quantization index; the CSI student model module is used to input the multi-dimensional CSI time-series features of the current time-series window into the pre-trained CSI student time-series model and output the probabilities of various spatial states and the model confidence. The state fusion module is used to process channel quantization metrics, various spatial state probabilities, model confidence, and radio quality metadata, and combines them with a time-series stabilization mechanism to complete time-series smoothing. The event interface module is used to output standardized spatial awareness events that match the duration of the timing window to downstream systems.
[0018] The beneficial effects of this invention are as follows: This invention relies solely on wireless transceiver nodes to collect CSI wireless signals for contactless sensing. During the inference phase, visual acquisition devices such as cameras can be turned off, eliminating the risk of leakage of privacy data such as images and faces. It is suitable for highly privacy-sensitive scenarios such as hospital wards, nursing home rooms, and residential bedrooms. No additional high-definition camera equipment is required; sensing can be achieved using only conventional wireless communication hardware, reducing hardware procurement, wiring, and long-term maintenance costs. It is suitable for large-scale deployment in various spaces such as homes, offices, warehouses, industrial production lines, and robotic work areas. Furthermore, it is not constrained by lighting or obstruction conditions; spatial status monitoring can still be completed in low-light or obstructed environments, overcoming the shortcomings of visual sensing solutions that rely on ambient light and are prone to creating blind spots.
[0019] This invention differs from traditional CSI schemes, which rely solely on a single image threshold to roughly determine the presence of disturbances and only support binary classification detection. This invention uses a wireless fingerprint baseline of an empty environment, a time-series CSI student multi-classification model, and multi-dimensional channel quantization indicators to jointly determine three differentiated spatial scenarios: vacant space, relatively static state of the target object, and dynamic activity state of the target object. It can distinguish between different conditions such as no target, low-motion target dwelling, dynamic target activity, and environmental baseline drift. This solves the problems of existing technologies being unable to distinguish between relatively static dwelling and empty backgrounds, and having coarse recognition accuracy, thus meeting the needs of intelligent spaces, industrial safety, and elderly care for refined state interpretation.
[0020] This invention establishes a stable CSI window verification mechanism. When the target space is confirmed to be in an vacant or low-disturbance state by manual verification, installation process verification, or joint verification by the model and wireless evidence, the wireless fingerprint baseline of the vacant environment is learned or updated. During real-time operation, the slow changes in the environment are quantified by indicators such as background residual, baseline drift, and gain offset, distinguishing between real target disturbances and slow-disturbance interference from environmental factors such as temperature and humidity, and furniture displacement. At the same time, a complete CSI preprocessing workflow is provided to sequentially complete abnormal frame filtering, dimension unification, and adaptive adjustment of the temporal window, filtering invalid data such as packet loss, disconnection, sampling anomalies, and device amplitude saturation, reducing identification interference caused by random wireless fluctuations, device migration, and unstable sampling rates, and reducing the probability of false detection and missed detection under long-term continuous monitoring.
[0021] This invention integrates multiple complementary evidences, including channel quantization metrics, spatial state probability, model confidence, and wireless quality, and adds a time-series smoothing and anti-shaking mechanism to eliminate the frequent state jumps caused by independent inference in a single window. The final output carries spatial awareness events with standardized fields, uniformly including complete information such as spatial identifier, spatial state type, activity category, activity intensity, idle baseline state, and device health status. The output format is standardized and can be directly connected to downstream systems such as smart homes, security systems, robot control platforms, and industrial monitoring platforms, making data integration convenient and highly scalable. Attached Figure Description
[0022] Figure 1 This is a flowchart of the target spatial object and spatial state perception method based on CSI in Example 1.
[0023] Figure 2 This is a schematic diagram of the two-stage process of model training and inference in Example 1.
[0024] Figure 3 This is a schematic diagram of the information fields of a standardized spatially perceived event in Example 1.
[0025] Figure 4 This is the architecture diagram of the target space object and space state perception system based on CSI in Example 2.
[0026] Figure 5 This is a schematic diagram of the node deployment of the wireless link in Example 2. Detailed Implementation
[0027] The present invention will now be further described with reference to the accompanying drawings and embodiments.
[0028] Example 1 like Figure 1 As shown, this embodiment discloses a target spatial object and spatial state perception method based on CSI, including the following steps: Establish a wireless link within the target space to continuously collect raw CSI I / Q data; The raw CSI I / Q data is preprocessed to obtain a processed CSI frame sequence, and radio quality metadata is extracted simultaneously; an adapted timing window is generated as required, and multi-dimensional CSI timing features corresponding to each timing window are extracted; The process of constructing or updating a wireless fingerprint for an idle environment involves determining the multi-dimensional CSI timing features within a timing window based on wireless quality metadata. If the conditions for a stable CSI window are met, the process of constructing or updating the wireless fingerprint for an idle environment is executed. In this embodiment, the initial construction of the wireless fingerprint for an idle environment is achieved by determining a stable CSI window based on manual confirmation of vacancy, installation process triggering, or wireless quality and short-term motion conditions, and using this window as the initial wireless fingerprint for an idle environment.
[0029] The channel quantization index is obtained by the difference between the multi-dimensional CSI timing characteristics and the wireless fingerprint of the idle environment. Input the multi-dimensional CSI time series features of the current time series window into the pre-trained CSI student time series model, and output the probabilities of various spatial states and the model confidence. The system integrates channel quantization metrics, various spatial state probabilities, model confidence, and radio quality metadata, and outputs standardized spatial awareness events that match the duration of the time window after time-series smoothing.
[0030] It should be noted that the target space described in this embodiment includes residential spaces, elderly care spaces, hospital wards, office spaces, industrial production lines, warehouse areas, robot work areas, vehicle cabins, and public areas of buildings. Correspondingly, the objects within the target space can be personnel, robots, mobile devices, materials, vehicles, or other objects that can cause changes in the wireless channel. The current preferred embodiment uses personnel as the target object for verification, illustrating the technical approach of using target objects to cause wireless channel disturbances that can be identified by the spatial state model; personnel detection does not constitute a limitation on the type of target object.
[0031] This embodiment does not require long-term use of visual acquisition devices such as cameras to stably output various refined spatial state perception results, such as spatial vacancy, relatively static dwelling of target objects, dynamic activity of target objects, and environmental baseline drift.
[0032] Specifically, the process for obtaining the processed CSI frame sequence described in this embodiment is as follows: Edge computing nodes receive serial or network transmission frames output by wireless sensing nodes and perform parsing operations on these frames to obtain a CSI frame sequence composed of several types of metadata. In this embodiment, the metadata constituting the CSI frame sequence includes acquisition timestamps, data packet sequence numbers, subcarrier dimensions, I / Q sampling values, received signal strength indicators, wireless transmission rates, channel noise floor, wireless sensing device identifiers, current frame sampling status, etc. Based on all the above metadata, information characterizing the wireless link transmission quality is filtered out to extract and generate wireless quality metadata, achieving synchronous generation of CSI frame body data and wireless quality metadata.
[0033] The raw CSI I / Q data is parsed to obtain a CSI frame sequence consisting of several metadata elements, and wireless quality metadata is extracted and generated based on these metadata elements. The parsed CSI frame sequence is sequentially subjected to abnormal frame filtering, time sorting, and sampling quality detection to obtain the intermediate CSI frame sequence. For ease of understanding, the abnormal frame filtering described in this embodiment refers to removing various invalid abnormal frames generated during the acquisition and transmission stages. These invalid abnormal frames include frames with missing or reversed acquisition timestamps, duplicate or abrupt data packet sequence numbers, mismatches between the number of subcarriers and the dimension of the model input tensor, all channel amplitudes being zero or reaching saturation, non-numerical or infinite values in channel data, received signal strength indicators or noise levels exceeding the device's rated operating range, and acquisition intervals exceeding a preset upper limit, etc.
[0034] Furthermore, time sorting refers to reordering the CSI frame sequence after anomaly filtering in ascending order based on the acquisition timestamps of each frame, unifying the temporal order of frame data, eliminating data misalignment caused by out-of-order transmission, and ensuring the accuracy of the temporal logic for subsequent temporal window division and feature extraction.
[0035] In this embodiment, the sampling quality detection is used to determine the stability of the wireless link corresponding to the current frame sequence. The detection indicators include effective coverage duration, effective frame rate, maximum disconnection duration, packet loss rate, duplicate frame ratio, and effective frame ratio.
[0036] Furthermore, a dimensionality consistency check is performed on the intermediate CSI frame sequence, which is a check on the input tensor of the intermediate CSI frame sequence; if the dimensionality consistency check fails, channel component adaptation is performed on the intermediate CSI frame sequence to obtain the processed CSI frame sequence.
[0037] In this embodiment, the input tensor includes the following dimensional parameters: number of antenna links, number of subcarriers, number of timing window frames, number of I / Q feature channels or amplitude-phase feature channels corresponding to the model feature dimensions, and selected subcarrier mask.
[0038] Based on the dimension parameters contained in the above input tensor, the specific method of channel component adaptation is as follows: select the effective subcarrier set shared by all antenna links, perform truncation or padding of the subcarrier dimensions according to a preset mask, resample the time frame sequence, fill abnormal antenna links with default values, and remove time windows with dimension anomalies; after the above adaptation process, if the time window still does not meet the model input requirements, then the window will not be sent into the model for inference or training.
[0039] Considering the differences in data processing scenarios, link sampling stability, and business identification requirements, the time-series window provided in this embodiment is divided into a fixed window or an adaptive window, and the corresponding selection method for the time-series window is as follows: Both the fixed window and the adaptive window can be applied to the process of constructing or updating the wireless fingerprint of an vacant environment and identifying the state of objects within the target space in real time. For example, when identifying the state in real time, the window duration in this embodiment ranges from 0.5 seconds to 10 seconds, preferably from 1 second to 3 seconds; when used for constructing or updating the wireless fingerprint of an vacant environment, the window duration ranges from 10 seconds to 120 seconds, preferably from 20 seconds to 60 seconds.
[0040] The adaptive window is applied to scenarios where sampling rate fluctuations, frame loss, or target space object state switching occur during data acquisition or data processing. Specifically, the window duration is dynamically extended or shortened based on the number of effective frames within the window, the maximum interruption duration, the short-term motion intensity within the window, and the confidence stability index, until the window meets the minimum number of effective frames and data coverage constraints.
[0041] In a preferred embodiment, the current verification prototype uses a fixed window for real-time inference and empty background learning; in other embodiments, an adaptive window can be enabled based on sampling rate fluctuations, frame loss, or target object state switching.
[0042] The multi-dimensional CSI timing features described in this embodiment include one or more of the following: amplitude variation sequence of each subcarrier signal, phase information of each subcarrier signal, normalized channel shape, inter-frame difference, window variation range, short-time motion intensity, and link quality features. These features can be flexibly selected according to the deployment scenario and model type.
[0043] Further examples illustrate the acquisition methods of feature quantization in multi-dimensional CSI time-series features.
[0044] The amplitude variation sequence and phase information of each subcarrier signal are directly extracted from the original CSI complex channel data within the time window, and will not be elaborated here.
[0045] The normalized channel shape is obtained as follows: Define the first... The frame-normalized channel shape is , represents the standardized channel vector of a single frame CSI after removing device gain and distance attenuation, eliminating amplitude deviations caused by hardware and distance.
[0046] Specifically, the data collected based on the wireless network card... Frame number The original in-phase sampled value of the subcarrier and the original orthogonal sampled values The CSI complex channel value corresponding to this subcarrier is obtained by combining the following formulas. : , It is the imaginary unit.
[0047] In the calculations below, the CSI complex channel values are directly reused. Built-in raw in-phase sampled values and the original orthogonal sampled values When solving for amplitude and normalized characteristics, no new I / Q variables are generated.
[0048] The corresponding subcarrier amplitude is calculated using the following formula. : .
[0049] Based on the subcarrier amplitude Logarithmic compression of the amplitude yields the logarithmic amplitude characteristic. : .
[0050] Calculate the frame-level gain reference value by iterating through the logarithmic magnitude characteristics of all subcarriers in the current frame. This embodiment uses either the median or the mean for calculation: or ,in, To take the median of the logarithmic magnitude feature across all subcarrier dimensions within the current frame, This represents the average of the logarithmic magnitude features across all subcarrier dimensions within the current frame.
[0051] Based on the frame-level gain reference value After completing the normalization operation, we obtain the first... Frame number Normalized channel components corresponding to subcarriers This embodiment includes two normalization implementations: or ,in, It is a very small constant used to avoid calculation errors caused by a denominator of zero.
[0052] All subcarriers within the same frame By combining these elements, the complete normalized channel shape can be obtained. .
[0053] Inter-frame difference is determined using the following formula: In the formula, For inter-frame difference, and Normalized channel shape and normalized channel shape The Middle The standardized channel values corresponding to the subcarriers. This is a calculation to obtain the mean value across all subcarrier dimensions.
[0054] The window variation range is used to characterize the channel fluctuation amplitude of a single subcarrier within a complete timing window; for the first... For each subcarrier path, the normalized channel values corresponding to all single-frame CSI data within the time window are extracted. The range of variation for that subcarrier path is obtained by subtracting the P95 quantile and P5 quantile within the window, or by subtracting the maximum and minimum channel values within the window. .
[0055] Example Explanation: A timing window contains 100 frames of valid single-frame CSI data. The normalized channel values of the 12th subcarrier within all frames of the window are extracted. After removing the extreme values at both ends (5%), the difference between the P95 quantile and the P5 quantile is calculated to obtain the window variation range of that subcarrier. .
[0056] Short-time motion intensity is used to quantify the overall channel fluctuation severity within a single time window. This embodiment provides two optional calculation methods, which can be selected according to the time window type and service scenario requirements. The two methods have differentiated adaptation characteristics: For example, when the time window is a fixed window, the sampling rate is stable, and the business objective is empty baseline learning or conventional real-time smooth recognition, the short-time motion intensity is calculated using the arithmetic mean method. The calculation formula is: ; The arithmetic mean method is not sensitive to short-term sudden spikes, can smooth out small-amplitude random noise, and is suitable for low-noise continuous monitoring scenarios.
[0057] In another embodiment, when the timing window is an adaptive window, there are sampling rate fluctuations, state transitions, sudden disturbances, or when it is necessary to improve the sensing sensitivity to large movements, the root mean square calculation method is used: .
[0058] In the formula, This indicates that the root mean square value is calculated for all inter-frame differences within the window along the temporal dimension.
[0059] The root mean square method amplifies the inter-frame difference signal with a large amplitude, enhances the feature response corresponding to violent motion, and can avoid the weakening of sudden motion features by mean smoothing. It is suitable for scenarios such as target entry, rapid movement, local disturbance, and sampling fluctuation.
[0060] For example, an adaptive timing window collects 80 frames of valid single-frame CSI data, and calculates the inter-frame difference frame by frame. to The root mean square of all inter-frame difference values is calculated to obtain the short-time motion intensity M corresponding to the time window.
[0061] Link quality features are generated based on wireless quality metadata extracted synchronously during the preprocessing stage. Specifically, they include one or more of the following: effective frame rate, packet loss rate, maximum disconnection duration, RSSI interval, and noise baseline index. These features are used to help determine the acquisition stability and data reliability of the CSI data in the current time window.
[0062] All the feature quantities corresponding to the above-mentioned multi-dimensional CSI time series features can be normalized to the 0~1 value range according to the baseline distribution of the vacant environment, or the quantile threshold can be used to complete the quantization characterization, eliminating the defect that the absolute amplitude of the channel between different acquisition devices and different target monitoring spaces is not horizontally comparable.
[0063] In a further embodiment, this embodiment introduces a stable CSI window, which is defined as: a time-series CSI window in which the effective frame rate, flow interruption, subcarrier dimension, and link quality all meet the standards within a preset duration, and the probability of the target object's existence, short-term motion intensity, and background residual value are all low. Preferably, this type of window generally corresponds to the target space being in an empty or low-disturbance state.
[0064] Therefore, the criteria for determining a stable CSI window are as follows: the effective frame rate, maximum disconnection duration, subcarrier dimension, and link quality corresponding to the radio quality metadata within the time window all meet the corresponding preset standards, and the probability of the target object's existence, short-term motion intensity, and background residual value within the time window are all lower than the corresponding thresholds; only when all conditions are met simultaneously is the time window determined to be a stable CSI window.
[0065] The processing method for multiple stable CSI windows is as follows: First, maintain a candidate stable window queue. Only when a specified number of windows in a series of consecutive stable windows or a series of windows in a recent period meet the stable CSI window determination criteria will the idle environment wireless fingerprint construction or update process be entered. The number of consecutive windows, the total number of windows, and the number of qualified windows can be pre-configured according to the scenario.
[0066] In this embodiment, if, after screening and judgment, the window that currently triggers the wireless fingerprint construction or update process in the vacant environment is... Then, based on the multi-dimensional CSI time series features within the corresponding window, let the first... Normalized channel shape of a stable CSI window The window-level empty channel shape is obtained by aggregating the normalized channel components of all frames within the window. Window gain reference and subcarrier noise scale The idle environment wireless fingerprint is then updated according to the following formula, thereby determining the window corresponding to the idle environment wireless fingerprint: in, The updated shape of the vacant channel. For the historical idle channel baseline stored before the update, The forgetting coefficient has a value between 0 and 1.
[0067] If the window level has an empty channel shape relative to historical idle channel baseline If the overall offset exceeds the baseline drift threshold, or if the short-term motion intensity and background residual value corresponding to the window do not meet the stability conditions, then the window will not participate in the wireless fingerprint update in the idle environment.
[0068] Correspondingly, the vacant environment wireless fingerprint includes at least one or more of the following: vacant channel shape, gain reference, noise scale, residual distribution, and motion distribution, which can be flexibly selected according to the target space deployment scenario and the type of recognition model.
[0069] Correspondingly, the vacant channel shape described in this embodiment is a vacant environment reference vector formed by aggregating the normalized channel shapes within multiple stable windows by means, medians, or exponential moving averages after the existence of candidate windows that meet the stable CSI window determination conditions in the target space and the queues meet the admission rules.
[0070] Based on real-time confirmation of wireless fingerprints in an vacant environment, the channel quantization metrics described in this embodiment include at least: background residual value, short-time motion intensity, baseline drift value, and wireless quality quantization value.
[0071] The following section provides a further explanation of the quantization metrics for each channel.
[0072] Among them, the background residual value is the background residual after standardization according to the vacant fingerprint noise scale. It belongs to the robust standardized distance and is used to quantify the overall deviation of the current time-series channel from the vacant baseline. Two optional calculation forms are provided.
[0073] If a time-series two-dimensional mean calculation method is used, the background residual value The calculation formula is: ; in, This represents the total number of valid frames within the current time window. The total number of subcarriers involved in the calculation. and The wireless fingerprint sensor in an idle environment is the first one. The mean value of the unused channel shape and the noise scale corresponding to the subcarrier.
[0074] If a single-window statistical mean calculation method is used, the background residual... The calculation formula is: ,in For the current window Normalized channel shape statistics for subcarriers.
[0075] Furthermore, the baseline drift value described in this embodiment The formula used to quantify the deviation between the recent stable window channel mean and the baseline mean of the vacant fingerprint is as follows: ,in, For the first time in the recent stable CSI window The channel mean of the subcarrier.
[0076] Wireless quality quantization value The result is obtained by weighted fusion calculation of effective frame rate, maximum disconnection duration, packet loss rate, subcarrier dimensional consistency, and RSSI stability. The quantization result is normalized to the range of 0 to 1. The higher the value, the better the wireless link acquisition quality in the current time window.
[0077] In a further embodiment, to compensate for the deficiency that the core channel quantization index cannot identify the overall gain synchronization drift of the device, a gain offset index can be added. As an independent auxiliary verification item, the verification object is the offset of the overall channel gain of the current time window relative to the wireless fingerprint benchmark gain in the idle environment. The gain offset index... The calculation formula is: ,in, For the channel gain of the recent stable CSI window; , These represent the baseline gain and gain fluctuation scale for wireless fingerprinting in an idle environment, respectively.
[0078] The gain offset metric is used to determine whether there is device gain, distance attenuation, or overall link amplitude synchronization drift in the current window. The system compares the gain offset metric with a preset gain threshold, combines the background residual value and short-term motion intensity to make a comprehensive judgment, and outputs subsequent processing instructions. The specific verification logic is as follows: If the gain offset index is below the threshold, the background residual value is below the corresponding threshold, and the short-term motion intensity is below the corresponding threshold, then maintaining or updating the wireless fingerprint in the idle environment is allowed. If the gain offset index is higher than the threshold, but the background residual value and short-term motion intensity are lower than the corresponding threshold, it is determined that the overall link gain has changed, the equipment location has changed, or the transmit power has changed. The target state switching is suspended, and the gain recalibration or baseline update status is triggered. If the gain offset index is higher than the threshold, and either the background residual value or the short-term motion intensity is higher than the corresponding threshold, it is determined to be a target disturbance or a sudden change in the environment. The update of the wireless fingerprint of the vacant environment is prohibited, and the abnormal environmental evidence or target activity candidate event is output. If the gain offset index remains above the threshold for a period exceeding the preset duration, the output device health status is abnormal or a baseline expiration event occurs.
[0079] In another embodiment, the training process for the CSI student time series model is as follows: Spatial annotation information is acquired and aligned to a temporal window. The spatial annotation information can be manually labeled, teacher modal labels, or a combination of both. Manually labeled labels are used for training samples where the target spatial state is known to the collector and the annotations are reliable. Teacher modal labels are used for training samples with synchronous video, millimeter-wave, UWB, or other sensor outputs from the teacher's end. Teacher modal labels can be hard labels, probability distribution soft labels, or intermediate features from the teacher's end, and can serve as primary labels, auxiliary labels, soft labels, or data quality diagnostic criteria. When the confidence level of a teacher modal label is insufficient, the collection viewpoint is obstructed, or there is inconsistency between the teacher's end and manual annotations, the teacher modal label is not used as the primary label, or it is marked as a low-confidence sample and removed.
[0080] The usage scenarios for manually labeled and teacher modal labels are differentiated as follows: In controlled acquisition scenarios where spatial states or action categories are explicitly selected manually, manually labeled labels are used as the primary labels. During the acquisition phase, when teacher modalities such as video pose models, millimeter-wave radar, UWB, and robot sensors are used simultaneously, and the confidence level of the teacher modal output reaches a threshold, teacher modal labels can be used as auxiliary or soft labels in training. When manually labeled and teacher modal labels conflict, manually labeled labels are prioritized, or the window is marked as a low-confidence / required-verification sample and not directly used to update the CSI student model. Teacher modal labels are not input during the online inference phase.
[0081] Using aligned spatial annotation information, a CSI student temporal model is trained based on a temporal network; the trained CSI student temporal model outputs a temporal multi-class classification. The temporal network described in this embodiment includes any one or more combinations of lightweight temporal networks such as TCN, 1D-CNN, recurrent neural networks, and Transformer. TCN is only a preferred implementation and is not a unique limitation on the type of temporal network.
[0082] The trained CSI student time series model described in this embodiment meets the following conditions: without teacher modality labels or manually labeled input, only real-time multi-dimensional CSI time series features are input to output the probabilities of each spatial state.
[0083] The spatial state includes at least: a spatial vacancy state, a relatively static existence state of the target object, and a dynamic activity state of the target object; in other embodiments, further subdivided static categories may be added.
[0084] The corresponding spatial state probabilities are calculated for each spatial state, and all spatial state probabilities are integrated to form a spatial state probability distribution feature.
[0085] To further illustrate, consider the input of a continuous, multi-dimensional CSI time-series feature sequence. : , This represents the total number of CSI sub-windows contained in a single input time series. For the first Multidimensional CSI temporal features of frames .
[0086] The model extracts multi-dimensional CSI time-series feature sequences. The model extracts channel shape temporal variation, frame-level gain variation, and short-term disturbance mode, and then outputs the probabilities of each category. The model output expression is as follows: , is the CSI student time series model, is the set of all trainable weight parameters of the model, and is the output vector, which is used to represent the probability of various spatial states such as spatial vacancy, the relatively static existence state of the target object, and the dynamic activity state of the target object.
[0087] Based on the spatial state probability distribution characteristics, the corresponding model confidence calculation method is selected to obtain the model confidence: The model confidence score is calculated using the highest class probability. : ,in, For spatial state category identification, This represents the predicted probability for the corresponding spatial state category. This indicates that the maximum value is taken among all class probabilities.
[0088] The highest class probability confidence score is suitable for scenarios where the model output probability distribution is sharp and the highest class probability has exceeded a preset threshold. It can be directly used to generate the model confidence score field within standardized spatially perceived events.
[0089] In another embodiment, the model confidence is calculated as a confidence score. : ,in, It represents the highest probability value among all spatial state probabilities. The second highest probability value is the difference between the two values, which represents the degree of classification and discrimination of the model.
[0090] Confidence margin is applicable to scenarios where the probability values of the highest and second-highest categories are close, the state boundaries are ambiguous, or the stability of state transitions needs to be constrained. For example, it is used to determine the transition between two states: relatively static existence and dynamic activity, or vacancy and relatively static existence. A state transition event can only be output when the confidence margin reaches a preset threshold and the condition is met for several consecutive time windows.
[0091] Finally, the information fields of the standardized spatial perception event described in this embodiment include at least the basic fields, perception results, environmental evidence, and output purpose; The basic fields include at least the space identifier, device identifier, timestamp, and space status type; The perception results include at least the occupancy status, activity category, model confidence, and activity intensity; The environmental evidence includes at least background residual values, short-term motion intensity, vacant baseline status, and wireless quality quantization values.
[0092] The spatial identifier is generated by the target spatial number information synchronously recorded when collecting raw CSI I / Q data; The activity intensity and idle baseline state are calculated by the channel quantization index, and the idle baseline state is used to characterize the effectiveness of wireless fingerprinting in idle environments. The health status of the device is determined by a combination of wireless quality metadata and wireless quality quantification values.
[0093] This embodiment provides exemplary value definitions for each standardized spatial event field: Spatial identifiers include global spatial numbers With local area numbering Among them, global space numbering It can be configured as a global spatial identifier such as home.living_room, factory.line_01, or a local area number. It can be configured as a local area identifier within a space such as door_area or table_area; The selectable values for spatial status type include: vacant, relatively static, dynamic, environmental change, baseline being updated, baseline failure, and unknown status.
[0094] The selectable values for the target object's activity category include: dynamic activity, relatively static existence, lying down, and unknown. The confidence score ranges from 0 to 1, corresponding to the classification confidence score value output by the model. Activity intensity supports two forms of expression: hierarchical labeling. low / medium / high Or normalize the continuous values between 0 and 1.
[0095] The selectable values for the background baseline status include: baseline valid, baseline adaptively updating, baseline expired and invalid, and baseline not initialized. The selectable values for device health status include: device normal, device performance degraded, and device offline.
[0096] Each of the above fields has an independent and clear physical meaning, and supports the addition of new fields and detailed value selection as needed to adapt to different deployment scenarios.
[0097] Example 2 This embodiment discloses a target spatial object and spatial state perception system based on CSI, used to implement the target spatial object and spatial state perception method described in Embodiment 1, including: At least one wireless signal transmitter and at least one wireless signal receiver are deployed in the target space according to preset requirements to form a wireless link, and the wireless link covers the activity area in the target space; the wireless signal transmitter and wireless signal receiver of the wireless link can be located at the space boundary, on both sides of the door, on the desktop and wall, on the top and side, or in a multi-node array position.
[0098] The CSI acquisition and quality assessment module is used to continuously acquire raw CSI I / Q data and synchronously record and generate wireless quality metadata. The preprocessing and windowing module is configured to preprocess the raw CSI I / Q data to obtain a processed CSI frame sequence, generate an appropriate time series window as required, and extract the multi-dimensional CSI time series features corresponding to each time series window. The idle environment fingerprint module, based on wireless quality metadata, determines the multi-dimensional CSI timing characteristics within the timing window. If the stable CSI window determination conditions are met, the idle environment wireless fingerprint construction or update process is executed. The channel quantization module is used to calculate the difference between the multi-dimensional CSI time-series features of the current time-series window and the wireless fingerprint of the idle environment to obtain the channel quantization index; the CSI student model module is used to input the multi-dimensional CSI time-series features of the current time-series window into the pre-trained CSI student time-series model and output the probabilities of various spatial states and the model confidence. The state fusion module is used to process channel quantization metrics, various spatial state probabilities, model confidence, and radio quality metadata, and combines them with a time-series stabilization mechanism to complete time-series smoothing. The event interface module is used to output standardized spatial awareness events that match the duration of the timing window to downstream systems.
[0099] To better illustrate the deployment and practical application effects of this system, two typical application implementation cases are presented below.
[0100] Case 1: Deploying a one-transmitter-one-receiver wireless node in a living room. In an vacant space, the system collects multiple segments of valid CSI data to establish a wireless baseline for the vacant living room environment. During model training, CSI sample data is collected for three scenarios: vacant space, relatively static target object presence, and dynamic target object activity. Training labels are generated using manual annotation, video pose model teacher labels, or a combination of both. During online inference, the video acquisition equipment is turned off, and the system relies solely on CSI temporal features and the vacant background baseline to determine and output the vacant state of the living room, the relatively static state of the target object, and the dynamic state of the target object. In this case, the target object is a person, and this is merely a preferred implementation method for verifying CSI-based spatial state perception routes.
[0101] Case 2: Deploying multiple wireless transceiver links in the industrial robot's operating area. The system first learns and constructs a unique vacancy wireless fingerprint for the operating area during idle periods when there are no personnel and the robot is stationary. During operation, it monitors in real time changes in wireless channel characteristics caused by robot movement and personnel intrusion. Finally, through the event interface module, it directly outputs standardized perception events—such as the activity of target objects in the area, area vacancy, and baseline state changes—to the downstream control system of the robot, supporting the human-machine collaborative safety management logic in the industrial field.
[0102] This solution can be implemented in smart homes for seamless occupancy detection and integrated lighting / air conditioning / security systems; in elderly care and medical care for privacy-friendly on-site / off-site / activity anomaly detection; in bionic or mobile robots for spatial perception, proximity sensing, and human-robot collaboration; in industrial sites for personnel / equipment / material area status monitoring; in warehouse areas for status monitoring; in building public spaces for perception; and in vehicle cabins for target space status recognition. The actual system can be mapped to various spatial states based on industry scenarios, including occupancy, relatively static presence, dynamic activity, proximity, dwelling, anomalies, risk areas, environmental changes, equipment surroundings, robot interaction objects, or industrial safety incidents.
Claims
1. A CSI-based target spatial object and spatial state perception method, characterized in that, Includes the following steps: Establish a wireless link within the target space to continuously collect raw CSI I / Q data; The raw CSI I / Q data is preprocessed to obtain a processed CSI frame sequence, and radio quality metadata is extracted synchronously. Generate suitable time series windows according to requirements, and extract multi-dimensional CSI time series features corresponding to each time series window; The process of building or updating the wireless fingerprint of the vacant environment is based on the wireless quality metadata, the target space vacancy confirmation information, and the multi-dimensional CSI timing characteristics within the timing window. If the stable CSI window determination condition is met, the process of building or updating the wireless fingerprint of the vacant environment is executed. The channel quantization index is obtained by calculating the difference between the multi-dimensional CSI timing characteristics of the current timing window and the wireless fingerprint of the idle environment. Input the multi-dimensional CSI time series features of the current time series window into the pre-trained CSI student time series model, and output the probabilities of various spatial states and the model confidence. The system integrates channel quantization metrics, various spatial state probabilities, model confidence, and radio quality metadata, and outputs standardized spatial awareness events that match the duration of the time window after time-series smoothing.
2. The method of claim 1, wherein, The process for obtaining the processed CSI frame sequence is as follows: The raw CSI I / Q data is parsed to obtain a CSI frame sequence consisting of several metadata elements, and wireless quality metadata is extracted and generated based on these metadata elements. The parsed CSI frame sequence is sequentially subjected to abnormal frame filtering, time sorting, and sampling quality detection to obtain the intermediate CSI frame sequence. A dimension consistency check is performed based on the intermediate CSI frame sequence, wherein the dimension consistency check is a check of the input tensor of the intermediate CSI frame sequence; If the dimension consistency check fails, channel component adaptation is performed on the intermediate CSI frame sequence to obtain the processed CSI frame sequence.
3. The method of claim 1, wherein, The timing window can be a fixed window or an adaptive window, and the corresponding timing window selection method is as follows: Both the fixed window and the adaptive window can be applied to the process of constructing or updating wireless fingerprints of object states and vacant environments in real time; The adaptive window is applied to scenarios where sampling rate fluctuations, frame loss, or target space object state switching occur during data acquisition or data processing.
4. The target spatial object and spatial state perception method based on CSI according to claim 1, characterized in that, The criteria for determining a stable CSI window are as follows: When constructing a wireless fingerprint in an empty environment for the first time, the effective frame rate, maximum disconnection duration, subcarrier dimension and link quality corresponding to the wireless quality metadata within the time window all meet the corresponding preset standards, and the short-term motion intensity is lower than the corresponding threshold. After an empty environment wireless fingerprint is established, it is further required that the probability of the target object's existence, the short-term motion intensity and the background residual value within the time window are all lower than the corresponding threshold. Correspondingly, the vacant environment wireless fingerprint includes at least one or more of the following: vacant channel shape, gain reference value, noise scale, residual distribution, or motion distribution.
5. The target spatial object and spatial state perception method based on CSI according to claim 1, characterized in that, The channel quantization metric also includes a gain offset metric; The channel quantization metrics include at least one or more of the following: background residual value, short-time motion intensity, baseline drift value, radio quality quantization value, and gain offset metric.
6. The target spatial object and spatial state perception method based on CSI according to claim 1, characterized in that, The training process for the CSI student time series model is as follows: Acquire spatial annotation information and align the spatial annotation information to the temporal window; the spatial annotation information is teacher modal labeling or manual annotation labeling. Using aligned spatial annotation information, a CSI student temporal model is trained based on a temporal network; the trained CSI student temporal model outputs a temporal multi-class classification.
7. The target spatial object and spatial state perception method based on CSI according to claim 1, characterized in that, The spatial state includes at least: a spatial vacancy state, a relatively static existence state of the target object, and a dynamic activity state of the target object; The corresponding spatial state probabilities are calculated for each spatial state, and all spatial state probabilities are integrated to form a spatial state probability distribution feature. Based on the spatial state probability distribution characteristics, the corresponding model confidence calculation method is selected to obtain the model confidence.
8. The target spatial object and spatial state perception method based on CSI according to claim 1, characterized in that, The information fields of the standardized spatial perception event include at least the basic fields, perception results, environmental evidence, and output purpose; The basic fields include at least the space identifier, device identifier, timestamp, and space status type; The perception results include at least the occupancy status, activity category, model confidence, and activity intensity; The environmental evidence includes at least background residual values, short-term motion intensity, vacant baseline status, and wireless quality quantization values.
9. A method for perceiving target spatial objects and spatial states based on CSI according to claim 8, characterized in that, The spatial identifier is generated from the target space number information synchronously recorded when acquiring raw CSI I / Q data; The activity intensity and idle baseline state are calculated by the channel quantization index, and the idle baseline state is used to characterize the effectiveness of wireless fingerprinting in idle environments. The health status of the device is determined by a combination of wireless quality metadata and wireless quality quantification values.
10. A target spatial object and spatial state perception system based on CSI, used to implement the target spatial object and spatial state perception method as described in any one of claims 1 to 9, characterized in that, include: At least one wireless signal transmitter and at least one wireless signal receiver are deployed in the target space according to preset requirements to form a wireless link, and the wireless link covers the activity area in the target space. The CSI acquisition and quality assessment module is used to continuously acquire raw CSI I / Q data and synchronously record and generate wireless quality metadata. The preprocessing and windowing module is configured to preprocess the raw CSI I / Q data to obtain a processed CSI frame sequence, generate an appropriate time series window as required, and extract the multi-dimensional CSI time series features corresponding to each time series window. The idle environment fingerprint module, based on wireless quality metadata, determines the multi-dimensional CSI timing characteristics within the timing window. If the stable CSI window determination conditions are met, the idle environment wireless fingerprint construction or update process is executed. The channel quantization module is used to calculate the difference between the multi-dimensional CSI timing features of the current timing window and the wireless fingerprint of the idle environment to obtain the channel quantization index. The CSI student model module is used to input the multi-dimensional CSI time series features of the current time series window into the pre-trained CSI student time series model and output the probabilities of various spatial states and the model confidence. The state fusion module is used to process channel quantization metrics, various spatial state probabilities, model confidence, and radio quality metadata, and combines them with a time-series stabilization mechanism to complete time-series smoothing. The event interface module is used to output standardized spatial awareness events that match the duration of the timing window to downstream systems.