Operation determination system and operation determination method

Abstract

To enable stable and versatile detection of the actions of targets within a monitored space. [Solution] In the operation determination system according to one embodiment of the present invention, the processor 41 of the operation determination device 30 performs a CSI acquisition process to acquire radio channel state information (CSI) relating to radio signals propagating within the monitoring space 10, and an operation determination process to map complex feature quantities representing the phase and amplitude included in the CSI to a hyperbolic space having negative curvature characteristics, and determine the operation of a target present in the monitoring space 10 based on geometric indicators representing the temporal change of the mapping points in the hyperbolic space.

Description

[Technical Field]

[0001] This invention relates to an action determination system for determining the behavior of a target present in a monitoring space. [Background technology]

[0002] In recent years, development has progressed on a technology called WLAN sensing, which uses Wi-Fi® wireless signals to detect the movement and position of targets (people and objects), and it has been standardized as "IEEE 802.11bf". WLAN sensing detects the movement and position of objects in the communication environment by analyzing the minute changes in phase and amplitude that occur when wireless signals transmitted and received by Wi-Fi-enabled wireless devices are reflected and attenuated by objects in the communication environment (people, walls, furniture, etc.) based on CSI (Channel State Information). With WLAN sensing, it is possible to detect the movement of objects without using video from cameras, etc., making it possible to understand the state of people while respecting their privacy.

[0003] Prior art related to the present invention includes the following. For example, Patent Document 1 discloses an invention that causes a computer to perform the following actions: to obtain the initial position of an object before it moves; to obtain at least one radio signal from a multipath channel affected by the object's movement; to extract a time-series CSI of the multipath channel from at least one radio signal; to determine the distance of the object's movement based on the time-series CSI; to estimate the direction of the object's movement; and to determine the new position of the object after the movement based on the distance, direction, and initial position. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Special Publication No. 2019-518202 [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] The technology described in Patent Document 1 is an approach based on statistical analysis in Euclidean space, and uses time reversal theory to analyze statistics such as the correlation of multipath signals, RSSI (Received Signal Strength Indicator), and the variance and standard deviation of CSI to track objects. However, the above method has problems with its low versatility, as it requires significant computational costs and prior training on the communication environment. Furthermore, it is susceptible to environmental fluctuations and noise, resulting in a lack of stability.

[0006] This invention has been made in view of the above-mentioned conventional circumstances, and aims to enable stable and versatile determination of the operation of targets present in a monitoring space. [Means for solving the problem]

[0007] An action determination system according to one aspect of the present invention is an action determination system for determining the action of a target present in a monitoring space, comprising one or more processors, wherein the processors perform a CSI acquisition process for acquiring wireless channel state information relating to wireless signals propagating in the monitoring space, and an action determination process for mapping complex feature quantities representing phase and amplitude included in the wireless channel state information to a hyperbolic space having negative curvature characteristics, and determining the action of the target based on geometric indices representing the temporal changes of the mapping points in the hyperbolic space.

[0008] In the above-described motion determination system, the motion determination process may map the complex feature quantities obtained for each frame of a predetermined time length to the hyperbolic space, calculate a collapse index representing the degree of collapse of the target structure for each frame based on the geodesic distance between frames of the mapping points in the hyperbolic space, convert the time-series data of the collapse index into a feature vector, refer to a vector database in which reference feature vectors related to motions that are candidates for detection have been registered in advance, and if a reference feature vector that matches the feature vector exists, detect the motion corresponding to that reference feature vector as a candidate for the target motion.

[0009] Furthermore, in the above-described motion determination system, the motion determination process may estimate and graph the transition of the target's attitude state based on the time-series data of the collapse index and geodesic distance in the frame interval in which the candidate motion was detected, refer to a graph database in which reference graphs representing the transition of the attitude state that occurs when the candidate motion occurs are pre-registered, and determine that the target has performed the candidate motion if a reference graph exists that matches the estimated graph of the transition of the attitude state.

[0010] Furthermore, in the above-described operation determination system, the complex feature quantities may be mapped to the hyperbolic space such that the smaller the fluctuations in phase and amplitude, the closer they are to the origin of the hyperbolic space, and the larger the fluctuations in phase and amplitude, the closer they are to the boundary of the hyperbolic space.

[0011] Furthermore, in the above operation determination system, the origin of the hyperbolic space may be adjusted based on a statistical index of environmental noise estimated from the wireless channel state information so that the environmental noise is compensated for.

[0012] Furthermore, in the above-described operation determination system, the operation determination process may be performed using a geometric model in which the hyperbolic space is represented in Euclidean space or pseudo-Euclidean space.

[0013] Also, in the above operation determination system, a request signal for requesting a measurement signal of radio channel state information from a first radio station to a second radio station may be transmitted at a predetermined interval, and in response to the request signal, the measurement signal may be transmitted from the second radio station to the first radio station, and the radio channel state information may be generated based on the measurement signal received by the first radio station.

[0014] Also, in the above operation determination system, the request signal may include a specification of a frequency and a bandwidth to be used for transmitting the measurement signal, and the second radio station may transmit the measurement signal using the frequency and the bandwidth specified by the request signal.

[0015] An operation determination method according to another aspect of the present invention is an operation determination method for determining an operation of a target existing in a monitoring space by one or a plurality of processors, including a CSI acquisition step of acquiring radio channel state information regarding a radio signal propagating in the monitoring space, and mapping a complex feature amount representing a phase and an amplitude included in the radio channel state information to a hyperbolic space having a negative curvature characteristic, and an operation determination step of determining the operation of the target based on a geometric index representing a temporal change of a mapped point in the hyperbolic space.

Effects of the Invention

[0016] According to the present invention, it becomes possible to stably and generally determine the operation of a target existing in a monitoring space.

Brief Description of the Drawings

[0017] [Figure 1] It is a diagram showing a configuration example of an operation determination system according to an embodiment of the present invention. [Figure 2] It is a diagram showing an example of hardware constituting the operation determination device of FIG. 1. [Figure 3] It is a diagram showing an example of functions of the operation determination device of FIG. 1. [Figure 4] It is a diagram showing an example of a processing flow by the operation determination device of FIG. 1.

Best Mode for Carrying Out the Invention

[0018] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 shows a configuration example of an operation determination system according to an embodiment of the present invention. The operation determination system shown in FIG. 1 includes a plurality of wireless devices 20 (here, two wireless devices 20A and 20B) that perform wireless communication in a communication environment including a monitoring space 10, and an operation determination device 30 that determines the operation of a target existing in the space 10 based on measurement data related to wireless communication. The monitoring space 10 is, for example, the main living space of a person 12 who is the target of operation determination. Hereinafter, a case where it is determined whether a target existing in the monitoring space 10 has fallen will be described as an example.

[0019] The wireless devices 20A and 20B are wireless communication devices compliant with the Wi-Fi standard. The wireless devices 20A and 20B may be dedicated devices having the functions described later, or may be general-purpose devices. Examples of general-purpose devices include wireless routers, smartphones, or household appliances having a wireless communication function such as air-conditioning equipment and lighting devices (so-called "smart home appliances"). Hereinafter, for convenience of explanation, the wireless device 20A will be taken as the transmission side (transmission station), and the wireless device 20B will be taken as the reception side (reception station), but a configuration in which each wireless device has both a transmission function and a reception function may also be adopted.

[0020] Upon receiving a radio signal transmitted from radio 20A, radio 20B calculates the CSI, which represents the state of the radio channel, as measurement data related to wireless communication. Wi-Fi uses methods such as OFDM (Orthogonal Frequency Division Multiplexing) and OFDMA (Orthogonal Frequency Division Multiple Access), and radio 20A divides the transmission data into multiple subcarriers and transmits it as a radio signal. The radio signal transmitted from radio 20A reaches radio 20B after passing through multiple propagation paths (multipath), which affects its phase and amplitude. Radio 20B estimates the phase and amplitude for each subcarrier of the received radio signal and outputs the CSI as measurement data.

[0021] Furthermore, CSI measurement is not limited to a configuration based on radio signals transmitted and received between multiple radio units 20. For example, radio unit 20B may send a radio signal towards the monitoring space 10, and radio unit 20B itself may receive the radio signal reflected by an object in the monitoring space 10 to perform CSI measurement. This configuration is effective for monitoring relatively small spaces (for example, the interior of a car).

[0022] The motion determination device 30 determines the motion of a target present in the monitoring space 10 based on time-series data of the CSI obtained with respect to the wireless signal propagating in the monitoring space 10. When a target present in the monitoring space 10 performs any motion, fluctuations in phase and amplitude occur in the wireless signal propagating near the target, and a fluctuation pattern specific to the motion appears in the time-series data of the CSI. Therefore, the motion of the target can be determined by analyzing the temporal fluctuations of the CSI.

[0023] The motion detection device 30 may be a device deployed inside or near the monitoring space 10, or it may be an external server (including a cloud server) connected via the Internet. Also, although the motion detection device 30 is represented as a separate device in Figure 1, it may be integrated into the wireless device 20B.

[0024] As shown in Figure 2, the operation determination device 30 is composed of a computer equipped with hardware resources such as a processor 41, memory 42, and storage 43. For example, the processor 41 loads a predetermined program stored in the storage 43 onto the memory 42 and executes it, thereby realizing each of the functions described later (see Figure 3). The processor 41 can be freely selected from existing processors such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), and two or more can be used in any combination. The storage 43 can be freely selected from existing devices such as an HDD (Hard Disk Drive), SSD (Solid State Drive), or flash memory, and two or more can be used in any combination.

[0025] Figure 3 shows an example of the functions of the motion determination device 30. The motion determination device 30 shown has a CSI acquisition function 51, a motion determination function 52, an origin adjustment function 53, and a notification function 54. Each of these functions will be described in detail below.

[0026] The CSI acquisition function 51 performs a CSI acquisition process to acquire CSI data relating to radio signals propagating within the monitoring space 10 from radio devices 20 (e.g., radio device 20B) deployed inside or near the monitoring space 10. If CSI is measured by multiple radio devices 20, CSI data may be acquired from each radio device 20.

[0027] The motion determination function 52 receives CSI data that has undergone external noise suppression processing, maps complex features representing the phase and amplitude contained in the CSI to a hyperbolic space with negative curvature characteristics, and performs motion determination processing to determine the target motion based on geometric indicators in the hyperbolic space. The motion determination processing can also be implemented using AI processing with deep learning models such as CNN (Convolutional Neural Network) or RNN (Recurrent Neural Network).

[0028] The operation determination process includes (1) hyperbolic space mapping, (2) collapse index calculation, (3) vector database search, and (4) graph database search. These processes will be explained in detail below. Here, the details of the operation determination process will be explained using the Poincaré sphere model. The Poincaré sphere model is an example of a geometric model that represents a hyperbolic space with negative curvature properties in Euclidean space.

[0029] (1) Hyperbolic space mapping The complex features representing the phase and amplitude of the CSI are mapped to a Poincaré sphere model such that smaller phase and amplitude fluctuations are closer to the origin of the Poincaré sphere, and larger phase and amplitude fluctuations are closer to the boundary (sphere surface) of the Poincaré sphere. The mapping of complex features to the Poincaré sphere model is performed, for example, by the following procedure.

[0030] The complex channel response for each subcarrier k (where k=1,2,…,n) in a time frame t of a predetermined length is defined as H. t (k) Let H t (k) is a complex number including phase and amplitude, and can be defined by the following equation (1). Here, |H t (k)| represents the amplitude of the subcarrier k, θ t (k) represents the phase of subcarrier k.

number

[0031] Complex channel response H for each sub - carrier k t (k) are concatenated to form an n - dimensional complex feature vector z as shown in the following equation (2). t to form.

Number

[0032] Complex feature vector z t is converted into a 2n - dimensional real vector z' by concatenating the real part and the imaginary part as shown in the following equation (3). t and is normalized so as to be within the Poincaré sphere. The Poincaré sphere is an n - dimensional unit open ball B defined by the following equation (4). n is.

Number

Number

[0033] Normalized feature vector z ~ t is mapped to a point u on the sphere by an exponential mapping from the origin 0 of the Poincaré sphere as shown in the following equation (5). t By the tanh function, the norm of the mapped point is always restricted to be less than 1 (i.e., staying inside the Poincaré sphere).

Number

[0034] When the target in the space 10 is stationary, the fluctuations in the phase and amplitude of the CSI are small, so the mapped point is close to the origin of the Poincaré sphere. On the other hand, when the target in the space 10 is in operation, the more rapid the operation is, the greater the fluctuations in the phase and amplitude of the CSI become, and the mapped point is close to the boundary of the Poincaré sphere.

[0035] (2) Avalanche exponent calculation process If u is the mapping point in the current time frame t, and v is the mapping point in the next time frame t+1, then the geodesic distance d(u,v) between the two mapping points u and v can be defined by the following equation (6).

number

[0036] Due to the geometric properties of the Poincaré sphere, the space increases exponentially near the boundary, so even a small increase in the norm causes a rapid increase in the geodesic distance d. Therefore, when the skeletal structure changes rapidly, such as when the target overturns, the time-series change in the measured CSI will appear as a rapid increase in the geodesic distance d. This property is used to calculate the collapse index C, which represents the degree of collapse of the target's skeletal structure, for each frame.

[0037] The collapse index C can be calculated using the following equation (7) based on the geodesic distance d(u,v) between consecutive time frames. In equation (7), φ(G) is a correction term based on the origin adjustment process described later.

number

[0038] (3) Vector database search process The vector database search process uses time-series data of the collapse index C to detect candidates related to the target's behavior. First, the time-series data of the collapse index C for the most recent W frames [C1, C2, ..., C W Convert the data into a feature vector. Methods for this conversion include extracting statistical measures from the time series data (e.g., maximum value, average rate of change, peak time, etc.) and using them as components, or utilizing the embedding representation of the time series data itself.

[0039] Next, the vector database is searched using the resulting feature vector (hereinafter referred to as the "measured feature vector"). The vector database contains a behavior vector library, which is a set of reference feature vectors that have been collected and registered in advance for the actions of the detection candidates. For each action of the detection candidate (for example, falling), one or more reference feature vectors can be registered as part of the behavior vector library.

[0040] Searching a vector database can be achieved, for example, by nearest neighbor search based on the cosine similarity or Euclidean distance between the measured feature vector and a reference feature vector. As a result, if a reference feature vector exists that matches the measured feature vector (has a predetermined similarity), the action corresponding to that reference feature vector is detected as a candidate for the target action.

[0041] (4) Graph database search processing The graph database search process verifies whether a candidate action actually occurred when a candidate action is detected by the vector database search process. For frame intervals in which a candidate action is detected, the target's attitude state at each point in time is estimated based on time-series data of the collapse index C and geodesic distance d, and a state transition graph (hereinafter referred to as the "measured state transition graph") representing the transitions in the target's attitude state is generated. The state transition graph is a graph composed of nodes representing attitude states and edges representing their transitions.

[0042] Next, the graph database is searched using the measured state transition graph. The graph database contains reference state transition graphs that have been collected and registered in advance regarding the transitions in posture states that occur when a candidate action occurs. One or more reference state transition graphs can be registered for each candidate action (e.g., falling over).

[0043] For example, if the target "falls over," it will go through the following state transitions: "Standing" → "Unstable" → "Rapid Collapse" → "Stationary." Therefore, for the detection candidate action "falling over," a state transition graph representing the above state transitions is registered as the reference state transition graph. Furthermore, constraints such as the time range required for the transition (for example, rapid collapse is 0.3 seconds to 1.5 seconds) can be attached to each edge of the reference state transition graph.

[0044] Graph database searches are performed by graph matching between the measured state transition graph and the reference state transition graph. Graph matching can be performed, for example, using subgraph isomorphism testing. That is, it is determined whether the reference state transition graph is contained within the measured state transition graph. The measured state transition graph represents the state transitions that actually occurred in the target, and includes nodes and edges due to preceding and succeeding daily operations and noise (i.e., it contains more nodes and edges than the reference state transition graph which represents ideal state transitions). Therefore, by checking whether the reference state transition graph is contained within the measured state transition graph using subgraph isomorphism testing, it is possible to determine whether the detected candidate operation actually occurred. Furthermore, if constraints are attached to the edges of the reference state transition graph, it is also determined whether the constraints are satisfied. As a result, if a reference state transition graph that matches the measured state transition graph exists (and the constraints are satisfied), it is determined that the detected candidate operation was actually performed by the target.

[0045] For example, if a measured state transition graph is obtained such as "standing" → "rapid collapse" → "standing," it represents a case where the target stumbled and regained their balance, and since there is no transition to "stationary," it is not judged as a "fall." In this way, by verifying the consistency of state transitions, it is possible to suppress misjudgments that cannot be eliminated by vector database search processing alone.

[0046] The origin adjustment function 53 performs an origin adjustment process to adjust the origin of the Poincaré sphere model so that environmental noise is compensated for. Electrical equipment such as air conditioners, refrigerators, and inverter circuits for lighting equipment are present in the monitoring space 10, generating quasi-periodic EMI (Electromagnetic Interference). The EMI generated by these electrical devices imparts a steady bias and drift to the CSI, systematically displacing the position of the mapping point on the Poincaré sphere. As a result, the collapse index C increases even when the target is stationary, causing false detections.

[0047] In the origin adjustment process, the origin of the Poincaré sphere model is adjusted to compensate for environmental noise based on a dynamic gain factor G, which represents the statistical characteristics of environmental noise estimated from the CSI. The dynamic gain factor G is calculated from the statistical characteristics of the CSI during a quiescent period, during which no target activity is estimated. A quiescent period is the time during which the collapse index C remains at a low value below a baseline for a certain period of time.

[0048] The dynamic gain factor G can be calculated, for example, by the following equation (8). Here, Q is the set of time frames of the quiescent period, T is the number of frames in the quiescent period (=|Q|), and z' τ Var[z'] is a 2n-dimensional real feature vector in the time frame τ∈Q during the quiescent period, where Var[z'] τ ] is z' τ This is the sample variance for each component (i.e., the variance of the phase and amplitude variations of each subcarrier).

number

[0049] In other words, the dynamic gain factor G is obtained by calculating the variance of the complex feature vector of the CSI during the quiet period (the variance of the phase and amplitude fluctuations of each subcarrier) and summing it up as a scalar value representing the magnitude of the environmental noise. The operation of summing it up as a scalar value involves summing, mean, maximum value, and L. 2Any statistical aggregation operation, such as a norm, can be used. The dynamic gain factor G is a dynamic value that is updated at regular intervals (e.g., every few minutes).

[0050] Based on the dynamic gain factor G, the origin (reference point) in the Poincaré sphere model is adaptively shifted in a direction that cancels out the bias caused by environmental noise. Specifically, the centroid of the mapping points during the quiet period (Fréchet mean μ) is set as the new reference point. The Fréchet mean μ can be calculated, for example, by the following equation (9).

number

[0051] Using the Fréchet mean μ, all mapping points are rearranged based on μ using a Möbius transform, as shown in equation (10) below. By performing this origin adjustment process, steady-state shifts due to environmental noise are removed, and only the true fluctuations originating from the target's operation can be reflected in the collapse index C.

number

[0052] Here, in the collapse index calculation formula (7) mentioned above, φ(G) is a correction term based on the dynamic gain factor G. The role of φ(G) is to adaptively adjust the scale of the collapse index C according to the environmental noise level. φ(G) can be calculated, for example, by the following formula (11). Here, α is a tuning parameter set according to the environment, and α>0.

number

[0053] The larger the dynamic gain factor G (higher noise), the smaller φ(G) becomes, reducing sensitivity and suppressing false detections. Conversely, the smaller the dynamic gain factor G (lower noise), the closer φ(G) approaches 1, enabling full-scale detection. In other words, the origin adjustment process is a spatial correction that adapts the reference point of the mapping to the environment, while φ(G) is a scalar correction for the calculation of the collapse index. Both function in conjunction as compensation mechanisms for environmental noise, but they operate at different layers.

[0054] The notification function 54 performs notification processing to inform a user, such as a monitor, of the target's fall if, as a result of the motion detection processing, it is determined that a target located within the monitoring space 10 has fallen over. The notification processing can be performed through the device itself or another device using any method such as screen display, audio output, lamp illumination, or a combination thereof.

[0055] Figure 4 shows an example of the processing flow using the operation determination device shown in Figure 1. The operation determination device 30 acquires CSI data obtained with respect to the wireless signal propagating in the monitoring space 10 at each time frame of a predetermined length (step S11). Next, it maps the complex features representing the phase and amplitude included in the CSI data to a hyperbolic space with negative curvature characteristics (for example, a Poincaré sphere model) (step S12), and calculates a collapse index C representing the degree of collapse of the target skeletal structure based on the geodesic distance between consecutive time frames (step S13). Subsequently, it determines whether or not it is a quiescent period in which the collapse index C remains at a low value below a reference value for a certain period of time (step S14). If it is determined to be a quiescent period (step S14: Yes), it performs an origin adjustment process (step S22) to adaptively shift the origin of the hyperbolic space in a direction that cancels out the bias due to environmental noise.

[0056] If it is determined that it is not a period of calm (Step S14: No), the time-series data of the collapse index C is transformed to generate an actual feature vector (Step S15), and the vector database is searched using the actual feature vector (Step S16). If the vector database search reveals that a reference feature vector matches the actual feature vector (Step S17: Yes), the transition of the target's attitude state is estimated to generate an actual state transition graph (Step S18), and the graph database is searched using the actual state transition graph (Step S19). If the graph database search reveals that a reference state transition graph matches the actual state transition graph (Step S20: Yes), notification processing is performed (Step S21). If there is no reference feature vector that matches the actual feature vector (Step S17: No), if there is no reference state transition graph that matches the actual state transition graph (Step S20: No), or after the notification processing (Step S21) is performed, the process proceeds to the next frame.

[0057] In the explanations so far, we have used the Poincaré sphere model as an example of a geometric model that represents hyperbolic space within Euclidean space. However, it is also possible to use other models such as the Poincaré disk model, the Poincaré half-plane model, and the Klein model. Furthermore, it is also possible to use geometric models that represent hyperbolic space within pseudo-Euclidean space, such as the Lorentz (hyperboloid) model.

[0058] In target operation detection using Wi-Fi CSI, the quality of the CSI data directly affects the accuracy of the detection. In conventional systems, the parameters of the beacon frames and data frames transmitted by the access point (AP) are fixed and not optimized for sensing purposes. Therefore, by controlling the CSI data collection parameters (e.g., beacon interval, bandwidth mode, frame transmission rate, frequency band), the SNR (Signal-to-Noise Ratio) and temporal resolution in operation detection are improved. The control of CSI data collection parameters will be explained in detail below, with radio 20A as the transmitting station (AP side) and radio 20B as the receiving station (STA side).

[0059] Radio 20B, which performs CSI measurements, sends a request signal (e.g., a UDP packet) to radio 20A requesting a CSI measurement signal. The request signal is sent at regular intervals (e.g., every 10 milliseconds). The request signal includes control information specifying the frequency and bandwidth to be used for transmitting the measurement signal. When radio 20A receives a request signal from radio 20B, it sends a measurement signal (e.g., a unicast response frame) to radio 20B using the frequency and bandwidth specified in the control information contained in the request signal. Radio 20B generates CSI data based on the received measurement signal.

[0060] In this way, by having radio 20B actively transmit a request signal to radio 20A, and having radio 20A transmit a measurement signal in response, CSI acquisition that is independent of the beacon interval becomes possible. Furthermore, by specifying the frequency and bandwidth used for transmitting the measurement signal, high-rate CSI acquisition can be achieved.

[0061] For example, by using the 5GHz band to transmit measurement signals, interference with 2.4GHz band Bluetooth, microwave ovens, and adjacent Wi-Fi networks can be avoided. Furthermore, avoiding interference improves the signal-to-noise ratio (SNR) of CSI data, improving the accuracy of detecting weak signals (e.g., respiratory signals).

[0062] For example, when using a 20MHz bandwidth, the number of subcarriers is 53, but by transmitting the measurement signal using a 40MHz bandwidth, the number of subcarriers can be increased to 57. As a result, the spatial resolution can be improved by approximately 7.5% compared to using a 20MHz bandwidth. Furthermore, when transmitting the measurement signal using an 80MHz bandwidth, the number of subcarriers can be increased to more than 100, resulting in an improvement of approximately 90% in spatial resolution compared to using a 20MHz bandwidth.

[0063] Furthermore, the CSI sampling rate can be optimized according to the application by adjusting the transmission interval of the request signal. For example, in the case of breath detection, the transmission interval of the request signal should be set to 10 milliseconds and the CSI sampling rate to 100 Hz. Respiration is in the 0.1-0.5 Hz band, and increasing the CSI sampling rate can improve the accuracy of power spectral density (PSD) estimation using the Welch method. In the case of motion detection such as walking or falling, the transmission interval of the request signal should also be set to 10 milliseconds and the CSI sampling rate to 100 Hz. Human body movements are in the 0.5-5 Hz band, and 100 Hz is sufficient oversampling. In the case of presence detection, the transmission interval of the request signal should be set to 100 milliseconds and the CSI sampling rate to 10 Hz, allowing operation in power-saving mode.

[0064] As described above, in the operation determination system of this example, the processor 41 of the operation determination device 30 performs a CSI acquisition process to acquire radio channel state information (CSI) relating to radio signals propagating within the monitoring space 10, and an operation determination process to map complex feature quantities representing the phase and amplitude included in the CSI to a hyperbolic space with negative curvature characteristics, and determine the operation of a target present in the monitoring space 10 based on geometric indices representing the temporal change of the mapping points in the hyperbolic space.

[0065] In the motion determination process, complex features obtained for each frame of a predetermined time length are mapped to hyperbolic space. Based on the geodesic distance between frames of the mapped points in hyperbolic space, a collapse index representing the degree of collapse of the target structure is calculated for each frame. The time-series data of the collapse index is converted into a feature vector (generating an actual feature vector). A vector database containing pre-registered reference feature vectors related to candidate motions is then referenced. If a reference feature vector that matches the actual feature vector exists, the motion corresponding to that reference feature vector is detected as a candidate for the target motion.

[0066] Furthermore, the motion determination process estimates and graphs the target's attitude state transitions based on time-series data of the collapse index and geodesic distance in the frame interval where the candidate motion was detected (generating an actual state transition graph). It then refers to a pre-registered graph database containing reference state transition graphs that represent the attitude state transitions that occur when the candidate motion occurs. If a reference state transition graph that matches the actual state transition graph exists, it is determined that the target performed the candidate motion.

[0067] In this example of the operation determination system, the processor 41 of the operation determination device 30 further performs an origin adjustment process to adjust the origin of the hyperbolic space so that the environmental noise is compensated for, based on a statistical index of environmental noise estimated from the wireless channel state information.

[0068] Performing this process will yield the following results. (1) High-sensitivity detection using boundary effects in hyperbolic space By mapping the complex features of the CSI to a hyperbolic space with negative curvature, the CSI fluctuations when the target structure changes rapidly are amplified as geodesic distance. This makes it possible to detect sudden changes in motion, such as toppling, with higher sensitivity compared to statistical analysis in Euclidean space. (2) Determination of environmental independence based on geometric indicators Because it uses geodesic displacement in hyperbolic space as a geometric indicator, it does not require prior training for a specific communication environment. Even if the shape of the monitored space or the furniture arrangement changes, adaptive correction through origin adjustment processing functions, ensuring robustness to environmental changes. (3) Misjudgment suppression through two-stage database processing By performing a two-stage process involving pattern matching using a vector database and state transition verification using a graph database, it is possible to effectively suppress misclassifications (for example, recovery from tripping, dropping of objects, etc.) that cannot be eliminated by threshold judgment of the collapse index alone. (4) Scalability for general-purpose action determination By switching the calculation formula for the collapse index, the threshold for judgment, and the registered data in the database, it can be applied to movement patterns other than falls (e.g., getting out of bed, prolonged immobility, wandering, etc.), giving it versatility as a movement judgment system.

[0069] The above explanation describes how to determine a target's falling motion, but it is also possible to determine other motions that exhibit unique fluctuation patterns in CSI time-series data. Depending on the motion to be determined, the calculation formula for the geometric index, the threshold for determination, and the data used in the vector database or graph database should be switched. In addition, to suppress misjudgments caused by pets with skeletal structures similar to humans, filtering of measured data and adjustment of the data registered in each database may be added.

[0070] Although embodiments of the present invention have been described above, these embodiments are merely illustrative and do not limit the technical scope of the present invention. The present invention can take many other forms, and various modifications such as omissions and substitutions can be made without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention as described herein, and are included in the scope of the invention as described in the claims or its equivalents.

[0071] Furthermore, the present invention can be provided not only as the devices described above or as systems composed of such devices, but also as methods executed by these devices, programs for a processor to realize the functions of these devices, and storage media for storing such programs in a computer-readable manner. [Industrial applicability]

[0072] This invention can be used in a motion determination system that determines the behavior of a target present in a monitored space. [Explanation of Symbols]

[0073] 10: Surveillance space, 12: Person, 20A, 20B: Wireless device, 30: Motion detection device, 41: Processor, 42: Memory, 43: Storage, 51: CSI acquisition function, 52: Motion detection function, 53: Origin adjustment function, 54: Notification function

Claims

1. An action determination system that determines the action of a target present in a monitoring space, Equipped with one or more processors, The aforementioned processor, A CSI acquisition process that acquires wireless channel status information relating to wireless signals propagating within the aforementioned monitoring space, An operation determination system that performs an operation determination process which maps complex feature quantities representing the phase and amplitude included in the wireless channel state information to a hyperbolic space having negative curvature characteristics, and determines the operation of the target based on geometric indicators representing the temporal change of the mapping points in the hyperbolic space.

2. In the operation determination system according to claim 1, The aforementioned motion determination process maps the complex feature quantities obtained for each frame of a predetermined time length to the hyperbolic space, calculates a collapse index representing the degree of collapse of the target structure for each frame based on the geodesic distance between frames of the mapped points in the hyperbolic space, converts the time-series data of the collapse index into a feature vector, refers to a vector database in which reference feature vectors related to motion candidates are registered in advance, and detects the motion corresponding to the reference feature vector as a candidate for the target motion if a reference feature vector that matches the feature vector exists.

3. In the operation determination system according to claim 2, The aforementioned motion determination process estimates and graphs the state transitions of the target's attitude based on time-series data of the collapse index and geodesic distance in the frame interval in which the candidate motion is detected, refers to a graph database in which reference graphs representing the state transitions that occur when the candidate motion occurs are pre-registered, and determines that the target has performed the candidate motion if a reference graph exists that matches the estimated state transition graph.

4. In the operation determination system according to claim 1, An operation determination system that maps the complex feature quantities to the hyperbolic space such that the smaller the phase and amplitude fluctuations, the closer the feature quantity is to the origin of the hyperbolic space, and the larger the phase and amplitude fluctuations, the closer the feature quantity is to the boundary of the hyperbolic space.

5. In the operation determination system according to claim 1, An operation determination system that adjusts the origin of the hyperbolic space so that the environmental noise is compensated for, based on a statistical index of environmental noise estimated from the wireless channel state information.

6. In the operation determination system according to claim 1, An action determination system that performs the action determination process using a geometric model that represents the hyperbolic space in Euclidean space or pseudo-Euclidean space.

7. In the operation determination system according to claim 1, A request signal is transmitted from the first radio station to the second radio station at predetermined intervals, requesting a signal for measuring radio channel status information. In response to the request signal, the measurement signal is transmitted from the second radio station to the first radio station. An operation determination system that generates the radio channel status information based on the measurement signal received by the first radio station.

8. In the operation determination system according to claim 7, The request signal includes specifying the frequency and bandwidth to be used for transmitting the measurement signal, The second radio station is an operation determination system that transmits the measurement signal using the frequency and bandwidth specified by the request signal.

9. An operation determination method for determining the operation of a target present in a monitoring space using one or more processors, A CSI acquisition step for acquiring wireless channel status information relating to wireless signals propagating within the aforementioned monitoring space, An operation determination method comprising: an operation determination step of mapping complex feature quantities representing phase and amplitude included in the wireless channel state information to a hyperbolic space having negative curvature characteristics, and determining the operation of the target based on geometric indicators representing the temporal change of the mapping points in the hyperbolic space.

Images