JP2021157275A - Action recognition server, action recognition system, and method for recognizing action - Google Patents

Abstract

To suppress reduction of the accuracy of recognition from sensor information including false detection data.SOLUTION: An action recognition server 1 includes: a sensor information acquisition unit 11 for acquiring sensor information for each sensor 2 from an aggregate of sensors 2 detecting an observation target person 2u; a sensor information conversion unit 11T for converting the sensor information to a possibility density function of a time direction on the basis of a reaction time when the observation target person 2u of the sensor information in the time series is detected, the function being at its maximum at the reaction time; an action classifying unit 14 for classifying actions of the observation target person 2u at each time on the basis of the converted sensor information; and an action output unit 17 for making data from the classified actions of the observation target person 2u and outputting the data. The sensor information conversion unit 11T converts the sensor information to a possibility density function of a space direction on the basis of a reaction place where the observation target person 2u of the sensor information in the time series is detected, the function being at its maximum at the reaction time when the observation target person 2u is detected.SELECTED DRAWING: Figure 3

Description

The present invention relates to a behavior recognition server, a behavior recognition system, and a behavior recognition method.

In recent years, high-performance sensors connected to the Internet have become widespread as IoT (Internet of Things) devices. Then, an attempt is made to extract useful information by collecting a large amount of sensor information as big data from a large number of sensors installed in an environment such as a home and analyzing the big data.

As the measurement frequency of sensor information, it is generally possible to measure it in a much shorter time than the human action period. Therefore, if one reaction of the sensor is replaced with one human action as it is, an unrealistic action such as "standing up three times per second" may be recognized.
Therefore, Patent Document 1 describes a method of smoothing those reactions and associating them with one action even when the sensor reacts a plurality of times within one second according to the sensor reaction definition prepared in advance.

JP-A-2004-145820

Although various sensors actually have a human being to be detected, erroneous detection occurs in which measurement is performed as non-detection. For example, there is a factor of erroneous detection such that the infrared rays for detection emitted from the motion sensor are blocked by the amount of illumination light in the room. As a result, the motion sensor overlooks a stationary person, and even if a relaxed person exists, it may be mistakenly recognized as absent.

Therefore, when the elderly person of the care recipient living in the home is to be monitored, it is necessary to improve the accuracy of recognizing the behavior of the elderly person even if the sensor malfunctions momentarily. However, in the conventional technique such as Patent Document 1, the influence of the malfunction of the sensor is not considered.

Therefore, the main object of the present invention is to suppress the reduction of recognition accuracy from the sensor information including the false detection data.

In order to solve the above problems, the behavior recognition server of the present invention has the following features.
The present invention includes a sensor information acquisition unit that acquires sensor information indicating a detection result for each sensor from a set of sensors that detect an observed person.
A sensor information conversion unit that converts the sensor information into a probability density function in the time direction that maximizes the reaction time based on the reaction time detected by the observed person in the sensor information in the time series. ,
A behavior classification unit that classifies the behavior of the observed person at each time based on the converted sensor information,
It is characterized by having a behavior output unit that digitizes and outputs the classified behavior of the observed person.
Other means will be described later.

According to the present invention, it is possible to suppress a decrease in recognition accuracy from sensor information including false detection data.

It is a block diagram of the behavior recognition system which concerns on one Embodiment of this invention. It is a hardware block diagram of the action recognition system which concerns on one Embodiment of this invention. It is a block diagram which shows the detail of the behavior recognition server which concerns on one Embodiment of this invention. It is a flowchart which shows the process of the action recognition server which concerns on one Embodiment of this invention. It is a time series graph of the sensor information in the state which there is no detection omission about one Embodiment of this invention. From the time series graph of FIG. 5 relating to one embodiment of the present invention, it is a time series graph in a state where a part of detection omission has occurred. It is a time series graph when the probability density function other than the curve about one Embodiment of this invention is applied. It is a graph when the probability density function is applied to the space axis which concerns on one Embodiment of this invention. It is a top view which shows the specific example of the space to which the graph of FIG. 8 concerning one Embodiment of this invention is applied. It is explanatory drawing which shows an example of the layout data which concerns on one Embodiment of this invention. It is a table for explaining layout data which concerns on one Embodiment of this invention. It is explanatory drawing of the image data which concerns on one Embodiment of this invention. It is a time series graph which shows the processing content of the action correction part which concerns on one Embodiment of this invention.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a configuration diagram of an action recognition system.
The behavior recognition system is configured so that the observer 3u remotely monitors the living condition of the observer 2u living at home 2h using the observer terminal 3. The behavior recognition server 1 recognizes the living state of the observed person 2u based on the sensor information acquired from various sensors 2, and notifies the observer terminal 3 of the recognition result. As a result, the observer 3u who sees the display screen of the observer terminal 3 can grasp the living state of the observer 2u.
The observer 2u is, for example, a care recipient, and the observer 3u is, for example, the family of the care recipient. Alternatively, a behavior recognition system may be introduced in a hospital or a long-term care facility instead of the home 2h, in which case the observer 3u becomes a doctor or a care manager.

At home 2h, various sensors 2 for monitoring the behavior of the observed person 2u are connected to the network. The sensor 2 may be, for example, a sensor incorporated in a home electric appliance such as a refrigerator 2a or an autonomous mobile vacuum cleaner 2b, or a single sensor such as a motion sensor 2c.
It is desirable that the sensor 2 such as the motion sensor 2c is installed in a direction in which the measurement area does not face the entrance of the room. By this installation, it is possible to prevent the motion sensor 2c from erroneously detecting a person different from the observed person 2u passing through the corridor outside the room.

FIG. 2 is a hardware configuration diagram of the behavior recognition system.
The sensor 2 notifies the observer 2u of a communication unit 121 that notifies other devices of sensor information detected by the detection unit 122, a detection unit 122 that detects the observer 2u, and a message from the observer 3u. It has a notification unit 123 and the like.
The action recognition server 1 controls the communication unit 111 that receives the sensor information from the sensor 2 and notifies the observer terminal 3 of the recognition result from the sensor information, the control unit 112 that recognizes the living state of the observer 2u, and the control unit 112. It has a storage unit 113 for storing data used for processing of the unit 112.
The observer terminal 3 inputs a communication unit 131 that receives the recognition result of the observer 2u, a notification unit 132 that notifies the observer 3u of the recognition result of the observer 2u, a message from the observer 2u, and the like. It has an input unit 133 to be operated.

The action recognition server 1 is configured as a computer having a CPU (Central Processing Unit) as an arithmetic unit (control unit 112), a memory as a main storage device, and a hard disk as an external storage device (storage unit 113).
In this computer, the CPU operates a control unit (control means) composed of each processing unit by executing a program (also called an application or an abbreviation for application) read in the memory.

FIG. 3 is a configuration diagram showing details of the action recognition server 1.
The control unit 112 (FIG. 2) of the action recognition server 1 includes a sensor information acquisition unit 11, a sensor information conversion unit 11T, a time information acquisition unit 12, an image conversion unit 13, an action classification unit 14, and an action correction unit. It has 15, a current action storage unit 16, and an action output unit 17.
The storage unit 113 (FIG. 2) of the action recognition server 1 stores the layout data 13L and the classification model 14m.
Hereinafter, the details of the components of FIG. 3 will be described with reference to the flowchart of FIG.

FIG. 4 is a flowchart showing the processing of the action recognition server 1.
The sensor information acquisition unit 11 acquires sensor information from the sensors 2 (refrigerator 2a, vacuum cleaner 2b, motion sensor 2c) installed in the home 2h (S101). The data format of the sensor information may differ depending on the type of the sensor 2.
The sensor information conversion unit 11T receives the sensor information in the discrete value data format of 0 or 1 from the sensor information acquisition unit 11, and converts the discrete value sensor information into the sensor information of the probability density function (S102, FIGS. 5 to 5). (See below in FIG. 9).

As the function value of the probability density function, the sensor information conversion unit 11T sets the function value at time t as the maximum value (for example, "1") from the input data of the discrete value "1" at the time t when the sensor reacts. Output data is created by adding a function value less than the maximum value in the time direction before and after that (Fig. 5). The function value less than the maximum value is calculated by the sensor information conversion unit 11T so that the function value becomes smaller as the time difference from the time t becomes larger.
On the other hand, the sensor information conversion unit 11T uses the input data, which is a data format other than the discrete value, as the output data as it is without conversion.

The image conversion unit 13 images a set of sensor information at a predetermined time based on the sensor information for each sensor 2 which is the output data of the sensor information conversion unit 11T (S103). In the layout data 13L referred to by the image conversion unit 13 at the time of conversion, information regarding the layout in the image, such as which part of the image the sensor information of which sensor 2 is to be arranged, is defined in advance (FIG. 10, FIG. FIG. 11).

Further, in addition to the set of sensor information, the image conversion unit 13 acquires time information indicating a predetermined time, which is the measurement time of the sensor information, via the time information acquisition unit 12, and sets the time information as an imaging target. May be included. If the sensor 2 includes a time stamp in the sensor information, the time information acquisition unit 12 acquires the time, and if there is no time stamp, the time information reception time is set as an image target.
Note that the behavior classification unit 14 may accept the sensor information and the time information that have not been imaged, omitting the image conversion process of the sensor information by the image conversion unit 13.

The behavior classification unit 14 classifies the behavior of the observed person 2u in the time information from the image data showing the sensor information (S104). For this classification process, a classification model 14m is prepared in which when image data is input in advance, the corresponding behavior is converted into data and output. The classification model 14m is trained by a machine learning algorithm such as deep learning.

The behavior correction unit 15 corrects an unnatural behavior that has occurred momentarily by referring to the behavior before and after the individual behavior output by the behavior classification unit 14 in time (FIG. 13). Later).
Therefore, when there is a local change from the action before and after the action (current action) to be focused on this time (S111, Yes), the action correction unit 15 changes the local action into the action before and after. After the correction is made so as to be consistent, the corrected action is accumulated in the current action storage unit 16 (S112). On the other hand, when there is no local change (S111, No), the natural behavior is accumulated in the current behavior storage unit 16 as it is (S113).
The action output unit 17 outputs the action recognition result currently accumulated in the action storage unit 16 to the outside (observer terminal 3). The output destination of the action recognition result is not limited to the customer environment (observer terminal 3), and may be output to another system such as a database system or a cloud system.

Hereinafter, a specific example of the processing (S102) of the sensor information conversion unit 11T will be described with reference to FIGS. 5 to 9.
FIG. 5 shows a time series graph of sensor information in a state where there is no detection omission.
FIG. 211 is sensor information input from the sensor information acquisition unit 11 to the sensor information conversion unit 11T. The discrete value “1” indicating the detection of the observed person 2u at the reaction times t1 to t5 is included in the graph 211.

The graph 212 is the result of the sensor information conversion unit 11T converting the discrete value sensor information into the probability density function using the graph 211 as input data. The sensor information conversion unit 11T receives the discrete value “1” of the reaction time t1 and converts it into a probability density function of the curve m1 having the reaction time t1 as a peak. Similarly, the sensor information conversion unit 11T creates a curve m2 at the reaction time t2, a curve m3 at the reaction time t3, a curve m4 at the reaction time t4, and a curve m5 at the reaction time t5, respectively.
The sensor information conversion unit 11T may apply, for example, a normal distribution, a student τ distribution, a U (Universal) distribution, and an arbitrary distribution used in other statistical fields as a distribution in which the sensor information is converted into a probability density function. can.

Graph 213 integrates the overlapping sections between the curves of Graph 212. Here, when a plurality of curves exist at the same time, the sensor information conversion unit 11T adopts the maximum value of those curves, but the sum of the curves may be adopted. As a result, the value of the probability density function at each time is uniquely obtained in the graph 213.
As described above, since the function value of each reaction time t1 to t5 is not "0" even after the conversion of the sensor information conversion unit 11T, the correct detection result is not deleted by the sensor information conversion unit 11T.

FIG. 6 shows a time-series graph in a state where a part of detection omission has occurred from the time-series graph of FIG.
Graph 221 is sensor information input from the sensor information acquisition unit 11 to the sensor information conversion unit 11T. Although the observed person 2u actually exists at home 2h at time t2 and t4, the discrete value becomes "0" due to the omission of detection. At the remaining reaction times t1, t3, and t5, the discrete value “1” is correctly detected as in FIG.
Graph 222 is the result of the sensor information conversion unit 11T converting the sensor information of discrete values into a probability density function using the graph 221 as input data. In this graph 222, the curve m2 at time t2 and the curve m4 at time t4 are missing from the graph 212 of FIG.

Graph 223 is a combination of overlapping sections between the curves of Graph 222, similar to Graph 213 of FIG. Here, focusing on the time t2, the sensor information (function value) at the time t2 is not "0" but is influenced by the probability density function (curves m1, m3) from the time t1 and t3 in the vicinity in time. .. Similarly, the function value at time t4 is also affected by the probability density function (curve m5) from time t5 in the vicinity in time.
In this way, even if a detection omission occurs at time t2 and t4, the detection omission can be relieved by using another signal in the vicinity in time as a probability density function.

FIG. 7 shows a time series graph when a probability density function other than a curve is applied from the same input data as the time series graph of FIG.
Similar to the graph 211, the graph 231 includes a discrete value “1” indicating the detection of the observed person 2u at times t1 to t5, respectively.
Graph 232 is the result of the sensor information conversion unit 11T converting the graph 231 as input data into a linear approximation probability density function having peaks at each time t1 to t5 of the discrete value “1”.
Straight line approximation requires less calculation. Further, in addition to the equation approximation of the linear approximation, the sensor information conversion unit 11T may use the curve approximation shown in FIG. 5, the polynomial approximation (not shown), or the like.

Graph 233 is the result of the sensor information conversion unit 11T converting the graph 231 into a predetermined range of random values using the graph 231 as input data. As shown below, the range in which the random value can be taken differs depending on whether the discrete value of the input data is "0" or "1".
-Discrete value of input data "0" → Output data "Random value in the range of 0 to 0.3"
-Discrete value of input data "1" → Output data "Random value in the range of 0.7 to 1.0"
As a result, it may be possible to relieve the detection omission even during the period when the discrete value "1" in the vicinity does not exist in time.

FIG. 8 is a graph when the probability density function is applied to the spatial axis.
In FIGS. 5 to 7, the sensor information conversion unit 11T artificially creates a detection signal around the time when the discrete value "1" of the input data is generated by applying the probability density function to the time axis. rice field.
Similarly, in FIG. 8, the sensor information conversion unit 11T applies a probability density function to the spatial axis, so that the sensor information conversion unit 11T is located around the place (living room) where the discrete value “1” of the input data is generated (bedroom, kitchen). ) May also create a pseudo detection signal.

FIG. 9 is a plan view showing a specific example of the space to which the graph of FIG. 8 is applied.
Assuming that the living room in which the discrete value "1" of the input data is generated is the existence probability "1 (100%)" of the observed person 2u, the sensor information conversion unit 11T spreads the existence probability to the room in the vicinity thereof.
For example, in the kitchen or bedroom, the discrete value "1" of the input data is not generated, and the observed person 2u is not detected. However, the sensor information conversion unit 11T spreads the kitchen (existence probability = 0.7) and the bedroom (existence probability = 0.5) in the order of proximity to the living room.

FIG. 10 is an explanatory diagram showing an example of layout data 13L used by the image conversion unit 13 for the imaging process. In the layout data 13L, the data contents to be written at each position in the square image data of 12 squares in the vertical direction and 12 squares in the horizontal direction are arranged as symbols in the figure such as "T" and "ACC1". The "mass" is the smallest unit obtained by subdividing the image area, and at least one writing area is assigned to the sensor information and the time information.

FIG. 11 is an explanatory table of the layout data 13L of FIG. For example, the top "T" in FIG. 10 corresponds to the symbol "T" in the figure of the first line "time" in FIG. The image data arranged at the position of the uppermost portion "T" in FIG. 10 is the time data acquired by the time information acquisition unit 12. That is, one image shown in FIG. 12 is visualized by aggregating a set of sensor information measured at the same measurement time (time data of "T") from the sensors 2 arranged at each location. The result.
The types of the sensor 2 that the sensor information conversion unit 11T converts into the probability density function in S102 include, for example, an acceleration sensor, a (door) open / close sensor, and the like that detect the operation of the observed person 2u, a motion sensor, and the like. Those that detect the presence of the observed person 2u can be mentioned.

The third column "number of cells" in the explanatory table indicates the size of the writing area. When the amount of data to be written is smaller than the amount of data that can be expressed in the write area, the write area is left over. At that time, the image conversion unit 13 fills the number of cells in the image by copying and writing the same data content to a plurality of locations.

The number of cells in the layout data 13L indicates the weight between the information to be written, and the larger the number of cells is assigned, the greater the influence on the behavior. The distribution of the number of squares is determined by, for example, the following policy.
・ Since it is customary for humans to take actions depending on the time of day, such as going out during the day and sleeping at night, the time information "T" allocates a larger number of cells (24 cells) than other sensor information. ..
-Since the actions that humans can take are narrowed down to some extent depending on where they are, the sensor information (location information) of the motion sensors "HM1 to HM5" allocates a larger number of cells (12 cells) than other sensor information.
・ Since human life has a habit of taking the same actions depending on the day of the week, such as going to work on weekdays and resting at home on holidays, the day of the week information "DoW" has more cells than the sensor information that measures the environment of home 2h. Allocate (12 squares).
-As sensor information for detecting human movements, the acceleration sensors "ACC1 to ACC4" and the open / close sensors "OC1 to OC3" allocate a larger number of cells (4 cells) than the sensor information for measuring the environment of the home 2h. ..

The fourth column "value" of the explanatory table indicates the data content to be written in the writing area. For example, when the color depth of the image data is 8 bits of gray scale, the amount of data that can be expressed by the writing area is 2 to the 8th power = 256 numerical values. Since the color depth of the image data can be set arbitrarily, the amount of data that can be expressed is not limited to 256. Therefore, for example, 8-bit grayscale and 16-bit color may be converted to different values and accuracy even if the reaction value of the same sensor is used. In this embodiment, the range of 0.00 to 1.00 is described with an accuracy of 0.01.
For example, the value "0.31" of the time "T" indicates 7:40 am when 0:00 is the value "0.00" and 23:59 is the value "1.00". On the other hand, the day of the week is selected from seven ways when Monday is set to the value "0.00" and Sunday is set to the value "1.00".
The above-mentioned "value" is a value in an arbitrary range based on the value of each sensor information. In addition to the case of referring to the color corresponding to the value of each sensor information as described above, the case of referring to the value of each sensor information as it is is also included.

In the line "HUM" of "humidity" in FIG. 11, the number of cells "5 (= 1x5)" means that the size of the writing area of the humidity sensor is one cell for one sensor and the number of sensors is five. It means a total of 5 squares. The "humidity" value "0.66,0.57,0.64,0.58,0.7" is the value of the first humidity sensor "0.66", the value of the second humidity sensor "0.57", ..., The value of the fifth humidity sensor in order from the left. Indicates "0.7".
The layout data 13L described above has described an example of arranging the same type of sensor information in close proximity in an image. On the other hand, sensor information having the same sensor installation location (room) may be arranged close to each other in the image.

FIG. 12 is an explanatory diagram of image data as a result of writing the “value” of FIG. 11 with respect to the layout data 13L of FIG. In FIG. 12, symbols in the figure such as “T” and “ACC1” are also shown for the sake of clarity, but in reality, the symbols in the figure are omitted from the image.
For example, the image conversion unit 13 writes black indicating the value “0” in the writing area of “ACC1”. On the other hand, the image conversion unit 13 writes white indicating the value "1" in the writing area of "HM4". That is, the larger the value to be written, the closer to white.
Further, the classification model 14m is defined by associating the image data created by the image conversion unit 13 with the behavior "returning home" of the observer 2u indicating the situation indicated by the image data.

The behavior classification unit 14 refers to the classification model 14m registered in the past, and when image data matching or similar to the image data of the classification model 14m is detected from the current observer 2u, the classification model The corresponding action "returning home" at 14 m is output as a classification result (S104).
In addition, in the definition of the classification model 14m, a person such as the observer 3u may teach a meaningful action label such as "going home" or "resting". On the other hand, an action label such as "behavior A" or "behavior B" automatically classified by machine learning may be used, which is simply a group of similar actions that have no meaning.

FIG. 13 is a time series graph showing the processing contents of the action correction unit 15.
Graph 241 shows the output data of the action classification unit 14 before correction. In Graph 241 it is basically detected that the observer 2u is out, but the bathing behavior for 5 minutes (ΔT1) at 10:00 and the cleaning for 3 minutes (ΔT2) at 15:00. Suppose an action is detected.
Graph 242 shows the output data of the behavior correction unit 15 after correction. When an action different from the previous / next action is suddenly detected, the action correction unit 15 corrects the different action so that the action is the same as the previous / next action (S112).
Therefore, when the period (ΔT1, ΔT2) of each action is shorter than the predetermined period Th = 10 minutes, the action correction unit 15 determines that the action is an unnatural action to be corrected. As a result, the bathing behavior at 10:00 and the cleaning behavior at 15:00 are corrected to the same behaviors as "going out" before and after, respectively.

Further, the action correction unit 15 may refer not only to the period of action but also to the type of action as a method of detecting an unnatural action to be corrected. For example, the behavior correction unit 15 may make a correction target for an action (going out) that is unnatural to occur immediately after (1 minute later) with respect to the previous action (relaxation).
The behavior correction unit 15 may change the predetermined period Th to be compared depending on the type of different behavior as to whether or not to correct the behavior different from the front and back. For example, bathing behavior is corrected as unnatural behavior if it is less than 20 minutes (predetermined period Th1), while cleaning behavior is corrected as unnatural behavior if it is less than 5 minutes (predetermined period Th2).
On the other hand, as a comparative example, there is a method of improving the accuracy of action recognition by shortening the time interval of action detection, but this method causes complicated control.

In the present embodiment described above, even if the sensor information acquisition unit 11 acquires the sensor information and the sensor information misses the observed person 2u due to a momentary malfunction, the sensor information conversion unit 11T is on the time axis or Detection omission can be relieved by using a probability density function based on sensor information in the vicinity on the spatial axis. As a result, it is possible to suppress a decrease in recognition accuracy from sensor information including false detection data.

The present invention is not limited to the above-described embodiment, and includes various modifications. For example, the above-described embodiment has been described in detail in order to explain the present invention in an easy-to-understand manner, and is not necessarily limited to those having all the described configurations.
Further, it is possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment.
Further, it is possible to add / delete / replace a part of the configuration of each embodiment with another configuration. Further, each of the above configurations, functions, processing units, processing means and the like may be realized by hardware by designing a part or all of them by, for example, an integrated circuit.
Further, each of the above configurations, functions, and the like may be realized by software by the processor interpreting and executing a program that realizes each function.

Information such as programs, tables, and files that realize each function can be stored in memory, hard disks, recording devices such as SSDs (Solid State Drives), IC (Integrated Circuit) cards, SD cards, DVDs (Digital Versatile Discs), etc. Can be placed on the recording medium of.
In addition, the control lines and information lines indicate those that are considered necessary for explanation, and do not necessarily indicate all the control lines and information lines in the product. In practice, it can be considered that almost all configurations are interconnected.
Further, the communication means for connecting each device is not limited to the wireless LAN, and may be changed to a wired LAN or other communication means.

1 Action recognition server 2 Sensor 2u Observer 3 Observer terminal 11 Sensor information acquisition unit 11T Sensor information conversion unit 12 Time information acquisition unit 13 Image conversion unit 13L Layout data 14 Behavior classification unit 14m Classification model 15 Behavior correction unit 16 Current behavior Accumulation unit 17 Action output unit

Claims (6)

A sensor information acquisition unit that acquires sensor information indicating the detection result for each sensor from a set of sensors that detect the observed person, and a sensor information acquisition unit.
A sensor information conversion unit that converts the sensor information into a probability density function in the time direction that maximizes the reaction time based on the reaction time detected by the observed person in the sensor information in the time series. ,
A behavior classification unit that classifies the behavior of the observed person at each time based on the converted sensor information,
A behavior recognition server characterized by having a behavior output unit that converts the classified behavior of the observed person into data and outputs the data. The sensor information conversion unit converts the sensor information into a probability density function in the spatial direction having the reaction location as the maximum value based on the reaction location detected by the observed person in the sensor information. The behavior recognition server according to claim 1. The action recognition server further includes an image conversion unit.
The image conversion unit images the sensor information by writing a value in an arbitrary range based on the value of the sensor information according to the arrangement of the sensor information in the image defined by the layout data. The behavior recognition server according to claim 1, wherein an image is used as the sensor information to be input to the behavior classification unit. The behavior recognition server further includes a behavior correction unit.
The behavior correction unit makes the behavior of the observed person classified by the behavior classification unit the same behavior as the behavior before and after when a behavior different from the behavior before and after is suddenly detected. The behavior recognition server according to claim 1, further comprising correcting different behaviors. A set of the sensors for detecting the observer, including a sensor installed in a direction not facing the entrance of the room in which the observer lives, and a set of the sensors.
A behavior recognition system having a behavior recognition server that recognizes the behavior of the observed person.
The action recognition server is
A sensor information acquisition unit that acquires sensor information indicating a detection result for each sensor from the set of sensors that detect the observed person, and a sensor information acquisition unit.
A sensor information conversion unit that converts the sensor information into a probability density function in the time direction that maximizes the reaction time based on the reaction time detected by the observed person in the sensor information in the time series. ,
A behavior classification unit that classifies the behavior of the observed person at each time based on the converted sensor information,
A behavior recognition system characterized by having a behavior output unit that converts the classified behavior of the observed person into data and outputs the data. The action recognition server has a sensor information acquisition unit, a sensor information conversion unit, an action classification unit, and an action output unit.
The sensor information acquisition unit acquires sensor information indicating a detection result for each sensor from a set of sensors that detect an observed person.
Based on the reaction time when the observed person is detected in the sensor information in the time series, the sensor information conversion unit converts the sensor information into a probability density function in the time direction having the reaction time as the maximum value. Converted,
The behavior classification unit classifies the behavior of the observed person at each time based on the converted sensor information.
The behavior output unit is a behavior recognition method characterized in that the classified behaviors of the observed person are converted into data and output.

Images