CN109697830B - Method for detecting abnormal behaviors of people based on target distribution rule - Google Patents
Method for detecting abnormal behaviors of people based on target distribution rule Download PDFInfo
- Publication number
- CN109697830B CN109697830B CN201811569674.8A CN201811569674A CN109697830B CN 109697830 B CN109697830 B CN 109697830B CN 201811569674 A CN201811569674 A CN 201811569674A CN 109697830 B CN109697830 B CN 109697830B
- Authority
- CN
- China
- Prior art keywords
- curve
- personnel
- abnormal
- person
- distance
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B21/00—Alarms responsive to a single specified undesired or abnormal condition and not otherwise provided for
- G08B21/02—Alarms for ensuring the safety of persons
- G08B21/04—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons
- G08B21/0407—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons based on behaviour analysis
- G08B21/0423—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons based on behaviour analysis detecting deviation from an expected pattern of behaviour or schedule
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
-
- G—PHYSICS
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B21/00—Alarms responsive to a single specified undesired or abnormal condition and not otherwise provided for
- G08B21/02—Alarms for ensuring the safety of persons
- G08B21/04—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons
- G08B21/0407—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons based on behaviour analysis
- G08B21/043—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons based on behaviour analysis detecting an emergency event, e.g. a fall
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Abstract
The invention relates to a method for detecting abnormal behaviors of people based on a target distribution rule, which adopts a pre-trained deep learning network, target detection is carried out according to the target learning training fine tuning parameters of the deployment scene, the states of personnel and objects in the scene are detected, according to the set state abnormal alarm point, performing abnormal alarm when triggered, simultaneously acquiring the real-time positions of personnel and objects in the area, returning position information, setting the area target point, calculating the distance between the personnel position and the target point, calculating the distance between the object position and the target point, calculating the distance between the personnel and various objects, calculating the result, forming a curve, respectively forming a distribution activity curve of the object, a personnel state activity curve and a distance curve from the personnel to a preset mark object, acquiring a long-term iteration regression curve, and when the curve fluctuation exceeds the preset deviation amplitude and does not meet the long-term rule and special date rule curve, an abnormal alarm is triggered to remind.
Description
Technical Field
The invention relates to a method for detecting abnormal behaviors of people based on a target distribution rule, and belongs to the technical field of health protection.
Background
With the growing problem of aging of the population, the demand for private care personnel is increasing year by year. Under the conditions of the current social development level, the medical resource level and the limited number of guardians, it is difficult to provide all-around personalized monitoring service for each demander, for example, after the old or the patient has abnormal conditions, if the old or the patient cannot find the abnormal conditions in time and alarm processing is carried out, serious consequences can be caused.
The current monitoring system who needs to be attended personnel to old man and patient etc. adopts physiological parameters such as contact sensor collection old man's rhythm of the heart, pulse, motion condition mostly, adopts gesture such as pressure sensor collection lie, sit, the sensor is of a great variety and the system deployment is complicated, and contact wearing mode leads to user experience relatively poor.
In the prior art, most of position acquisition adopts active infrared and passive infrared sensors, which are easily influenced by shielding, inaccurate and need to be arranged in large quantity; in the prior art, some abnormal feedback uses that a Bayesian network is trained on the position of a person in a sensor feedback time sequence in big data, a Bayesian network model only depends on a position variable, a time variable and an abnormal probability variable are modeled, the model features are too few, the dependency relationship among the features is too large, the performance of the established Bayesian network model is poor, the overfitting of the model lacks generalization capability and individual uniqueness, and the judgment of whether a person under guardianship is abnormal is a complicated multi-dimensional problem and the judgment of whether the abnormal misjudgment rate is very large only depends on the position and the time.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a method for detecting abnormal behaviors of people based on a target distribution rule, which realizes the detection of the abnormal behaviors of the people by utilizing deep learning and counting based on the distribution rule of the people and the objects:
the invention aims to carry out statistics iterative learning on the daily life, work and rest rules and object distribution rules of individuals, automatically monitor in real time, alarm and remind in time when abnormality is found, and contact hospitals or families to carry out treatment in time.
The invention utilizes a deep learning method to obtain a target video by a monitoring camera, detect and position personnel and articles in real time, record the position and the state of target detection to form a 24-hour statistical curve, iterate the 24-hour statistical curve for multiple days to obtain a target distribution curve taking 24 hours as a period, namely a living habit and a work and rest rule of a person to be monitored, set a short-term rule threshold value and a long-term rule threshold value abnormal alarm point to alarm abnormal behaviors, and can timely alarm and contact related personnel to timely handle when an abnormal condition occurs in a detection area, thereby avoiding the occurrence of danger.
The invention can realize real-time monitoring of patients, disabled people and old people with poor physical conditions, can give an alarm in real time when the users are abnormal, avoids long-term in-hospital care, greatly liberates medical resources, reduces the cost of the person under guardianship and the inconvenience caused by in-hospital treatment, greatly liberates the time of family members of the person under guardianship, can continue to live at home after the person under guardianship is deployed, does not influence the daily life of the users, and brings great convenience to the society.
The invention adopts a deep learning target detection network, and can detect all objects needing to be detected in real time in the room in an off-line manner based on monitoring. The abnormity analysis of the invention is that one part of the two parts directly triggers abnormity alarm, for example, if a preset chair is fallen down, the old people lie on the ground, and triggers the direct alarm condition, and directly alarms, and the second part is the deviation long-term rule, and according to the curve iteration of the daily indoor object and human activity of the person under guardianship, the position of the object and the human body of the person under guardianship in each period of time can be predicted, and the deviation triggers abnormity.
Summary of the invention:
the method comprises the steps of adopting a pre-trained deep learning network to process a target area video acquired by a monitoring camera in real time, carrying out target detection, acquiring real-time positions of personnel and articles in the area, returning position information, setting an area target point, calculating the distance between the position of the article and the target point, calculating the distance between the person and various articles, calculating results, forming curves, respectively forming a distribution activity curve and a personnel state activity curve of the articles and a distance curve from the person to a preset mark article, iterating the curves acquired for multiple days in a 24-hour period, acquiring a long-term iteration curve, setting upper and lower threshold values of the curves, and alarming and reminding when the curve formed after the curve fluctuation exceeds a preset range or deviates from the long-term distribution rule iteration. And detecting the personnel and the objects by using a deep learning method, so as to count the positions of the targets to obtain a distribution rule, and finally detecting abnormal behaviors of the personnel.
The technical scheme of the invention is as follows:
a method for detecting the abnormal personnel by utilizing deep learning to detect a target based on a distribution rule comprises the following steps:
(1) target detection: adopting a deep learning target detection network model pre-trained by a public data set, carrying out fine tuning training according to a deployment scene to obtain a deployed target detection network for carrying out real-time personnel and object detection, forming a bounding box, namely a bounding box, for the detected personnel and objects, and returning the position information of the detected personnel and objects, wherein the position information comprises coordinate information of four vertexes of the bounding box; for example, model parameters of target detection network models such as fast-rcnn, Mask-rcnn and YOLO are pre-trained on an Imagenet data set, detection of daily articles and people can be achieved after preliminary pre-training, then scene video frame by frame image acquisition is carried out according to an actual deployment scene, the scene images are made into label files and sent to a target detection network for fine tuning training, verification and testing, and accuracy of scene object detection is improved.
Further preferably, in the step (1), a small target detection network is adopted to detect the persons and the objects in real time, and the position information of the detected persons and the detected objects is returned. For example, the Yolov3 network architecture can be deployed on an embedded platform NVIDIA TX2 by means of a darknet framework through experiments, NVIDIA TX2 only has a certain computing capacity with a bank card size, and after Tensorflow, OpenCV and CUDA are installed, the Yolov3 network architecture is deployed by means of the darknet framework for target detection, so that the deployment requirement is lower and lower in cost, the hardware requirement is lower, a networked background server is not required to operate a deep learning target detection network, offline deployment can be realized, a marker object is detected in real time, and the position information of the detected marker object is returned.
(2) And (4) position recording: for the position information of the returned detected personnel and each object, a position coordinate system is established by taking the lower left corner of one rectangular monitoring picture as the origin, the lower edge as the x axis and the left edge as the y axis, and four vertex coordinates (x) of each returned object bounding box are usedi1,yi1)、(xi2,yi2)、(xi3,yi3)、(xi4,yi4) I denotes different objects, and the coordinates (x) of the center point of each object are calculated and recordedicenter.yicenter)；
The coordinates (x ') of the four vertices of the human bounding box returned in the position coordinate system'i1,y′i1)、(x′i2,y′i2)、(x′i3,y′i3)、(x′i4,y′i4) Calculating and recording the coordinates (x) of the center point of the personperson-i,yperson-i)；
It is further preferred that the first and second liquid crystal compositions,
it is further preferred that the first and second liquid crystal compositions,
(3) counting according to time: counting the change condition of the coordinate of the center point of the object along with time, counting the change condition of the coordinate of the center point of the personnel along with time, and counting the distance H between the coordinate of the center point of the personnel and the coordinate of the center point of the mark object; marker objects, for example: beds, tables, chairs, toilets, balconies, kitchens, etc.,
it is further preferred that the first and second liquid crystal compositions,
preferably, step (3) includes:
statistics of object information: recording the coordinates (x) of the center point of each objecticenter.yicenter) Taking time as an X axis and the distance from the center point of each object to the origin in the position coordinate system as a Y axis to form a distance change curve LiCounting the change condition of the position information of each object;
for statistics of the person information: recording the coordinates of the center point of the person (x)person-i,yperson-i) Taking time as an X axis and the distance from the center point of the person to the origin in the position coordinate system as a Y axis to form a distance change curve LpCounting the change condition of the position information of the personnel;
counting the activities of personnel in the monitoring area: selecting a marker object in a monitored area, and setting coordinates (x) of the center point of the marker objectmark-i,ymark-i) With time as the x-axis, the distance H from the person to the center point of each markerp-MFor the y-axis, form a graph Lp-M。
(4) Extracting short-term distribution rules: the 24-hour curve L of the object position information obtained by multi-day statistics i24 hours curve of person position LpGraph L of the distance of a person to a landmark object over 24 hoursp-MRespectively carrying out local weighted linear regression to obtain the distribution rule curve R of the object in each time period every dayiThe activity law curve R of the person in each time periodpAnd the distance law curve R of each time segment of the person relative to each marker objectp-M；
Further preferably, in the step (4), the local weighted linear regression process uses a loss function J (θ), as shown in formula (i):
in the formula (I), the compound is shown in the specification,
hθ(x(i)) As a function of the regression curve, x(i)Is an independent variable, y(i)As a dependent variable, w(i)The method comprises the steps of taking a weight function, taking x as a predicted value, taking k as a wavelength parameter to control the rate of weight distance reduction, setting the value of k in advance, obtaining a theta variable by obtaining the minimum value of J (theta), further obtaining a corresponding local weighted linear regression curve, and obtaining a distribution rule curve R of an object in each time period every day through local weighted linear regressioniThe activity law curve R of the person in each time periodpAnd the distance law curve R of each time segment of the person relative to each marker objectp-M。
(5) Extracting a long-term distribution rule: increasing the width of a time horizontal axis, taking a week as an iteration cycle, counting the distribution rules of object distribution, personnel activities and the daily distribution rule of the relative marker distance of the personnel in one week, and respectively forming a rule curve Wi，Wp，Wp-M；
Taking months as an iteration cycle, counting the distribution rule of the object distribution, the activity of personnel and the distribution rule of the relative marker distance of the personnel in one month, and respectively regressing to form a rule curve Mi，Mp，Mp-M；
(6) Abnormality detection: presetting the number of people under guardianship, turning off abnormal detection when detecting that family members exist in a guardianship area, and triggering abnormal detection when detecting that the number of scene personnel is less than or equal to the preset number of guardianship personnel, wherein the abnormal detection comprises scene abnormal triggering abnormal alarm and abnormal alarm deviating from a regular curve;
the scene exception triggers an exception: presetting abnormal alarm points in advance, setting scene abnormal alarm points by using data detected by pictures according to abnormal conditions under big data statistics, detecting whether a scene meets the abnormal alarm preset conditions or not, and triggering an alarm if the scene meets the abnormal alarm preset conditions; like elderly fall detection, seizure detection, etc., for example: detecting that the chair is in a falling abnormal state by detecting that the length-width ratio change of the chair exceeds a threshold value, and fusing the posture estimation result of the old people to obtain an alarm conclusion that the old people may fall down; the detection of immobility of the elderly at certain locations and the detection of abnormal rapid activity at key points of the limbs may estimate that the elderly may be in a tic state. According to the preset direct scene abnormity alarm point, abnormity is triggered, namely, the alarm is given to contact doctors or family members for checking, and danger can be avoided by timely handling.
Deviation from the regularity curve is abnormal: object distribution rule curve R obtained by iterationi、Wi、MiPersonnel activity law curve Rp、Wp、 MpAnd the distance distribution rule curve R of the person relative to the mark objectp-M、Wp-M、Mp-MWhen the trigger is abnormal, the short-term curve and the long-term curve need to be considered comprehensively, and objects and personnel deviate from the short-term and long-term distribution rule curves of the objects and the personnel respectively, so that the alarm is triggered to be abnormal. For example, the time length of the old people entering the toilet deviates from the statistical distribution rule, namely, an abnormal alarm is triggered; if the old people still do not get up after the statistical time of the distribution rule in a certain day, the old people trigger the abnormity but comprehensively consider the short-term and long-term distribution rule curves, for example, the statistical result shows that the user gets up at 7 o ' clock in weekdays but gets up at 9 o ' clock in weekends is normal, and if 8 o ' clock in a certain day are detected, whether the rule deviating from the long term is deviated in weekdays or holidays at the same time, the old people trigger the abnormity.
According to the optimization of the invention, the anomaly detection in the step (6) is realized by an anomaly detection system, and the anomaly detection system comprises a picture acquisition module, an operation module and an alarm reminding module;
the picture acquisition module comprises a plurality of cameras and is used for acquiring videos and transmitting the videos to the operation module for real-time video processing;
the operation module refers to a computer or an embedded device, such as NVIDIA TX2, and is used for processing transmitted data in real time to perform target identification and anomaly detection;
the alarm module is a GMS module and sends abnormal contact information and corresponding screen screenshots after receiving the alarm prompt of the operation module, the receiving time of feedback information of related contacts is set, and after the receiving time exceeds the time and processed information is not fed back, the standby contacts are contacted. So that the relevant personnel can timely view and process the abnormity to prevent the person under guardianship from being in danger.
The invention has the beneficial effects that:
1. the detection speed is high, the network structure is simple, the parameters are few, the requirement on hardware is low, and the method can be arranged on a common embedded platform such as an NVIDIA TX2 development board. The system can be deployed off-line, does not depend on a remote server, and protects personal privacy.
2. The system can realize real-time monitoring, immediately alarm and remind when abnormity occurs, and avoid danger of people under guardianship.
3. The wearable sensor does not need to be deployed, so that the user experience is better, the user experience is easy to accept, and the daily life of the user is not influenced.
4. Iterative evolution learning can be performed, different scenes can be adapted, and the precision can be gradually improved through training.
5. According to the obtained long-term article distribution rule and the life rule of people, the life rule suddenly deviates within a period of time, and the alarm reminding is carried out, so that a doctor can find hidden dangers in time.
6. The system can be deployed for monitoring the old and the patients, and can perform abnormal real-time alarm reminding according to preset scene abnormal points, contact related personnel to avoid danger, not only does not influence the daily life of the attended personnel, but also greatly saves medical resources and manpower resources of the guardians, and brings great convenience to the society.
Drawings
FIG. 1 is a schematic flow chart of a method for detecting abnormal personnel based on a distribution rule according to the present invention;
FIG. 2 is a graph showing a statistical guardian's law curve obtained by experiments with a daily regression cycle according to the present invention;
FIG. 3 is a graph illustrating a weekly activity rule curve of a statistical guardian obtained by experiments according to the present invention, wherein the weekly activity curve is a regression period;
FIG. 4 is a schematic flow chart of anomaly detection according to the present invention;
Detailed Description
The invention is further defined in the following, but not limited to, the figures and examples in the description.
Example 1
A method for detecting the abnormal condition of a person by utilizing deep learning to detect a target based on a distribution rule comprises the following steps as shown in figure 1:
(1) target detection: adopting a deep learning target detection network model pre-trained by a public data set, carrying out fine tuning training according to a deployment scene to obtain a deployed target detection network for carrying out real-time personnel and object detection, forming a bounding box, namely a bounding box, for the detected personnel and objects, and returning the position information of the detected personnel and objects, wherein the position information comprises coordinate information of four vertexes of the bounding box; for example, model parameters of target detection network models such as fast-rcnn, Mask-rcnn and YOLO are pre-trained on an Imagenet data set, detection of daily articles and people can be achieved after preliminary pre-training, then scene video frame by frame image acquisition is carried out according to an actual deployment scene, the scene images are made into label files and sent to a target detection network for fine tuning training, verification and testing, and accuracy of scene object detection is improved.
And detecting the personnel and each object in real time by adopting a micro target detection network, and returning the position information of the detected personnel and each object. For example, the Yolov3 network architecture can be deployed on an embedded platform NVIDIA TX2 by means of a darknet framework through experiments, NVIDIA TX2 only has a certain computing capacity with a bank card size, and after Tensorflow, OpenCV and CUDA are installed, the Yolov3 network architecture is deployed by means of the darknet framework for target detection, so that the deployment requirement is lower and lower in cost, the hardware requirement is lower, a networked background server is not required to operate a deep learning target detection network, offline deployment can be realized, a marker object is detected in real time, and the position information of the detected marker object is returned.
(2) And (4) position recording: for the position information of the returned detected personnel and each object, a position coordinate system is established by taking the lower left corner of one rectangular monitoring picture as the origin, the lower edge as the x axis and the left edge as the y axis, and four vertex coordinates (x) of each returned object bounding box are usedi1,yi1)、(xi2,yi2)、(xi3,yi3)、(xi4,yi4) I denotes different objects, and the coordinates (x) of the center point of each object are calculated and recordedicenter.yicenter)；
The coordinates (x ') of the four vertices of the human bounding box returned in the position coordinate system'i1,y′i1)、(x′i2,y′i2)、(x′i3,y′i3)、(x′i4,y′i4) Calculating and recording the coordinates (x) of the center point of the personperson-i,yperson-i)；
(3) Counting according to time: counting the change condition of the coordinate of the center point of the object along with time, counting the change condition of the coordinate of the center point of the personnel along with time, and counting the distance H between the coordinate of the center point of the personnel and the coordinate of the center point of the mark object; marker objects, for example: beds, tables, chairs, toilets, balconies, kitchens, etc.,
the step (3) includes:
statistics of object information: recording the coordinates (x) of the center point of each objecticenter.yicenter) Taking time as an X axis and the distance from the center point of each object to the origin in the position coordinate system as a Y axis to form a distance change curve LiCounting the change condition of the position information of each object;
for statistics of the person information: recording the coordinates of the center point of the person (x)person-i,yperson-i) Taking time as an X axis and the distance from the center point of the person to the origin in the position coordinate system as a Y axis to form a distance change curve LpCounting the change condition of the position information of the personnel;
counting the activities of personnel in the monitoring area: selecting a marker object in a monitored area, and setting coordinates (x) of the center point of the marker objectmark-i,ymark-i) With time as the x-axis, the distance H from the person to the center point of each markerp-MFor the y-axis, form a graph Lp-M。
(4) Extracting short-term distribution rules: the 24-hour curve L of the object position information obtained by multi-day statistics i24 hours curve of person position LpGraph L of the distance of a person to a landmark object over 24 hoursp-MRespectively carrying out local weighted linear regression to obtain the distribution rule curve R of the object in each time period every dayiThe activity law curve R of the person in each time periodpAnd the distance law curve R of each time segment of the person relative to each marker objectp-M；
In the step (4), the local weighted linear regression process adopts a loss function J (theta), which is shown in formula (I):
in the formula (I), the compound is shown in the specification,
hθ(x(i)) As a function of the regression curve, x(i)Is an independent variable, y(i)As a dependent variable, w(i)For the weight function, x is a predicted value, k is a rate of controlling the decrease of the distance of the weight by the wavelength parameter, the value of k needs to be set in advance, and the minimum value of J (theta) is obtainedObtaining a theta variable to further obtain a corresponding local weighted linear regression curve, and obtaining a distribution rule curve R of the object in each time period every day through local weighted linear regressioniThe activity law curve R of the person in each time periodpAnd the distance law curve R of each time segment of the person relative to each marker objectp-M。
(5) Extracting a long-term distribution rule: increasing the width of a time horizontal axis, taking a week as an iteration cycle, counting the distribution rules of object distribution, personnel activities and the daily distribution rule of the relative marker distance of the personnel in one week, and respectively forming a rule curve Wi，Wp，Wp-M；
Taking months as an iteration cycle, counting the distribution rule of the object distribution, the activity of personnel and the distribution rule of the relative marker distance of the personnel in one month, and respectively regressing to form a rule curve Mi，Mp，Mp-M；
(6) As shown in fig. 4, abnormality detection: presetting the number of people under guardianship, turning off abnormal detection when detecting that family members exist in a guardianship area, and triggering abnormal detection when detecting that the number of scene personnel is less than or equal to the preset number of guardianship personnel, wherein the abnormal detection comprises scene abnormal triggering abnormal alarm and abnormal alarm deviating from a regular curve;
the scene exception triggers an exception: presetting abnormal alarm points in advance, setting scene abnormal alarm points by using data detected by pictures according to abnormal conditions under big data statistics, detecting whether a scene meets the abnormal alarm preset conditions or not, and triggering an alarm if the scene meets the abnormal alarm preset conditions; like elderly fall detection, seizure detection, etc., for example: detecting that the chair is in a falling abnormal state by detecting that the length-width ratio change of the chair exceeds a threshold value, and fusing the posture estimation result of the old people to obtain an alarm conclusion that the old people may fall down; the detection of immobility of the elderly at certain locations and the detection of abnormal rapid activity at key points of the limbs may estimate that the elderly may be in a tic state. According to the preset direct scene abnormity alarm point, abnormity is triggered, namely, the alarm is given to contact doctors or family members for checking, and danger can be avoided by timely handling.
Deviation from the regularity curve is abnormal: object distribution rule curve R obtained by iterationi、Wi、MiPersonnel activity law curve Rp、Wp、 MpAnd the distance distribution rule curve R of the person relative to the mark objectp-M、Wp-M、Mp-MWhen the trigger is abnormal, the short-term curve and the long-term curve need to be considered comprehensively, and objects and personnel deviate from the short-term and long-term distribution rule curves of the objects and the personnel respectively, so that the alarm is triggered to be abnormal. For example, the time length of the old people entering the toilet deviates from the statistical distribution rule, namely, an abnormal alarm is triggered; if the old people still do not get up after the statistical time of the distribution rule in a certain day, the old people trigger the abnormity but comprehensively consider the short-term and long-term distribution rule curves, for example, the statistical result shows that the user gets up at 7 o ' clock in weekdays but gets up at 9 o ' clock in weekends is normal, and if 8 o ' clock in a certain day are detected, whether the rule deviating from the long term is deviated in weekdays or holidays at the same time, the old people trigger the abnormity.
Example 2
The method for detecting the abnormal personnel based on the distribution rule by utilizing the deep learning to detect the target is characterized by comprising the following steps of:
the anomaly detection in the step (6) is realized through an anomaly detection system, and the anomaly detection system comprises a picture acquisition module, an operation module and an alarm reminding module;
the image acquisition module comprises four groups of Haikang high-definition cameras and is used for acquiring videos and transmitting the videos to the operation module for real-time video processing; the camera collects pictures, simple inter-frame difference and background difference detection is carried out, and the pictures are transmitted to the operation module after being changed, so that the design can reduce the pressure of the operation module;
the operation module is NVIDIA TX2 and is used for processing transmitted data in real time to perform target identification and anomaly detection; the NVIDIA TX2 is only as large as one bank card, rapid deployment can be achieved, a YOLOv3 target recognition network architecture is built, the target recognition network is pre-trained through ImageNet, fine tuning training is conducted according to pictures collected by scenes to adjust network parameters, and recognition of people, tables, chairs, televisions, beds and the like in the scenes is achieved. The scene abnormity detection conjectures the state of a person through the detection of key points of the human body and the feedback of the length-width ratio of a boundary frame of the person and an object, sets the abnormal falling alarm to the detection feedback of the key points of the human body to be in a lying state or the ratio of the boundary frame of the person under guardianship to be in the lying state, assists the object to detect the abnormal falling state and improves the abnormal confidence coefficient, the lying position is not positioned in a daily position which can be laid down, such as a sofa and a bed, and the abnormal falling alarm is triggered when the person under guardianship is in the abnormal lying state for. The position information of the object and the person which are fed back and detected by the image data target detection network is collected, a daily human activity curve is formed by taking 24 hours as a period, time as an x axis and distance as a y axis, and local weighted regression is carried out on the daily curve or the daily human rule curve after the regression is obtained, such as an experimental chart 2. The curve counted in 24 hours is divided into Monday, Tuesday and Sunday in a week period, local weighted regression is respectively carried out to form a rule curve counted in the week period, such as an experiment or a figure 3 obtained, and the daily life rule of a person in one week is collected. And (3) with a month as a period, firstly carrying out image similarity clustering on the curves every day in one month, acquiring data of several months, carrying out local weighted regression according to the clustering labels to form a monthly rule curve, and also acquiring daily curves, weekly curves and monthly curves of the object distribution rule and the relative personnel marker object distance distribution rule. Presetting collection amount, respectively regressing after the collection data amount is reached to form a new curve, reminding to check after the deviation of the new and old regular curves is large, updating the regular curves when the deviation is within the preset range, and iteratively regressing the curves according to the updating of the collected data amount ceaselessly.
The alarm module is a GMS module and sends abnormal contact information and corresponding screen screenshots after receiving the alarm prompt of the operation module, the receiving time of feedback information of related contacts is set, and the standby contacts are contacted after the processing information is not fed back after the receiving time exceeds the time. So that the relevant personnel can timely view and process the abnormity to prevent the person under guardianship from being in danger. The method comprises the steps of collecting scene videos in real time daily, presetting the number of people under guardianship, triggering abnormal detection if the number of people detected in the same scene exceeds the number of people under guardianship, triggering abnormal detection if the number of people accords with the number of people under guardianship, triggering abnormal alarm if the number of people and object position information fed back deviates from a regular curve in time sequence, and contacting related personnel to check and solve the abnormality in time to prevent the people under guardianship from danger.
Claims (8)
1. A method for detecting abnormal behaviors of people based on a target distribution rule is characterized by comprising the following steps:
(1) target detection: adopting a deep learning target detection network model pre-trained by a public data set, carrying out fine tuning training according to a deployment scene to obtain a deployed target detection network for carrying out real-time personnel and object detection, forming a bounding box, namely a bounding box, for the detected personnel and objects, and returning the position information of the detected personnel and objects, wherein the position information comprises coordinate information of four vertexes of the bounding box;
(2) and (4) position recording: for the position information of the returned detected personnel and each object, a position coordinate system is established by taking the lower left corner of one rectangular monitoring picture as the origin, the lower edge as the x axis and the left edge as the y axis, and four vertex coordinates (x) of each returned object bounding box are usedi1,yi1)、(xi2,yi2)、(xi3,yi3)、(xi4,yi4) I denotes different objects, and the coordinates (x) of the center point of each object are calculated and recordedicenter.yicenter)；
The coordinates (x ') of the four vertices of the human bounding box returned in the position coordinate system'i1,y′i1)、(x′i2,y′i2)、(x′i3,y′i3)、(x′i4,y′i4) Calculating and recording the coordinates (x) of the center point of the personperson-i,yperson-i)；
(3) Counting according to time: counting the change condition of the coordinate of the center point of the object along with time, counting the change condition of the coordinate of the center point of the personnel along with time, and counting the distance H between the coordinate of the center point of the personnel and the coordinate of the center point of the mark object;
(4) extracting short-term distribution rules: the 24-hour curve L of the object position information obtained by multi-day statisticsi24 hours curve of person position LpDistance of person to object 24 hoursGraph Lp-MRespectively carrying out local weighted linear regression to obtain the distribution rule curve R of the object in each time period every dayiThe activity law curve R of the person in each time periodpAnd the distance law curve R of each time segment of the person relative to each marker objectp-M；
(5) Extracting a long-term distribution rule: increasing the width of a time horizontal axis, taking a week as an iteration cycle, counting the distribution rules of object distribution, personnel activities and the daily distribution rule of the relative marker distance of the personnel in one week, and respectively forming a rule curve Wi，Wp，Wp-M；
Taking months as an iteration cycle, counting the distribution rule of the object distribution, the activity of personnel and the distribution rule of the relative marker distance of the personnel in one month, and respectively regressing to form a rule curve Mi，Mp，Mp-M；
(6) Abnormality detection: presetting the number of people under guardianship, turning off abnormal detection when detecting that family members exist in a guardianship area, and triggering abnormal detection when detecting that the number of scene personnel is less than or equal to the preset number of guardianship personnel, wherein the abnormal detection comprises scene abnormal triggering abnormal alarm and abnormal alarm deviating from a regular curve;
the scene exception triggers an exception: presetting abnormal alarm points in advance, setting scene abnormal alarm points by using data detected by pictures according to abnormal conditions under big data statistics, detecting whether a scene meets the abnormal alarm preset conditions or not, and triggering an alarm if the scene meets the abnormal alarm preset conditions;
deviation from the regularity curve is abnormal: object distribution rule curve R obtained by iterationi、Wi、MiPersonnel activity law curve Rp、Wp、MpAnd the distance distribution rule curve R of the person relative to the mark objectp-M、Wp-M、Mp-MWhen the trigger is abnormal, the short-term curve and the long-term curve need to be considered comprehensively, and objects and personnel deviate from the short-term and long-term distribution rule curves of the objects and the personnel respectively, so that the alarm is triggered to be abnormal.
2. The method for detecting the abnormal behavior of the people based on the target distribution rule as claimed in claim 1, wherein in the step (1), a tiny target detection network is adopted to detect the people and the objects in real time, and the position information of the detected people and the detected objects is returned.
3. The method for detecting abnormal behavior of people based on target distribution rule as claimed in claim 1,
4. the method for detecting abnormal behavior of people based on target distribution rule as claimed in claim 1,
5. the method for detecting abnormal behavior of people based on target distribution rule as claimed in claim 1,
6. the method for detecting abnormal behavior of people based on the target distribution rule as claimed in claim 1, wherein the step (3) comprises:
statistics of object information: recording the coordinates (x) of the center point of each objecticenter.yicenter) Taking time as an X axis and the distance from the center point of each object to the origin in the position coordinate system as a Y axis to form a distance change curve LiCounting the change condition of the position information of each object;
for statistics of the person information: recording the coordinates of the center point of the person (x)person-i,yperson-i) Taking time as an X axis and the distance from the center point of the person to the origin in the position coordinate system as a Y axis to form a distance change curve LpCounting the change of the position information of the personnel；
Counting the activities of personnel in the monitoring area: selecting a marker object in a monitored area, and setting coordinates (x) of the center point of the marker objectmark-i,ymark-i) With time as the x-axis, the distance H from the person to the center point of each markerp-MFor the y-axis, form a graph Lp-M。
7. The method for detecting abnormal human behavior based on the target distribution rule of claim 1, wherein in the step (4), the local weighted linear regression process employs a loss function J (θ), as shown in formula (I):
in the formula (I), the compound is shown in the specification,
hθ(x(i)) As a function of the regression curve, x(i)Is an independent variable, y(i)As a dependent variable, w(i)The method comprises the steps of taking a weight function, taking x as a predicted value, taking k as a wavelength parameter to control the rate of weight distance reduction, obtaining a theta variable by obtaining the minimum value of J (theta), further obtaining a corresponding local weighted linear regression curve, and obtaining a distribution rule curve R of an object in each time period every day through local weighted linear regressioniThe activity law curve R of the person in each time periodpAnd the distance law curve R of each time segment of the person relative to each marker objectp-M。
8. The method for detecting the abnormal behavior of the personnel based on the target distribution rule is characterized in that the abnormal detection is realized by an abnormal detection system, and the abnormal detection system comprises a picture acquisition module, an operation module and an alarm reminding module;
the picture acquisition module comprises a plurality of cameras and is used for acquiring videos and transmitting the videos to the operation module for real-time video processing;
the operation module is a computer or an embedded device and is used for processing the transmitted data in real time to perform target identification and abnormality detection;
the alarm reminding module is a GMS module and is used for sending abnormal contact information and corresponding screen screenshots after receiving the alarm reminding of the operation module, setting the receiving time of feedback information of related contacts, and contacting standby contacts after the receiving time exceeds the time and processing information is not fed back.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811569674.8A CN109697830B (en) | 2018-12-21 | 2018-12-21 | Method for detecting abnormal behaviors of people based on target distribution rule |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811569674.8A CN109697830B (en) | 2018-12-21 | 2018-12-21 | Method for detecting abnormal behaviors of people based on target distribution rule |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109697830A CN109697830A (en) | 2019-04-30 |
| CN109697830B true CN109697830B (en) | 2020-10-20 |
Family
ID=66231883
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811569674.8A Active CN109697830B (en) | 2018-12-21 | 2018-12-21 | Method for detecting abnormal behaviors of people based on target distribution rule |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109697830B (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110443408A (en) * | 2019-07-04 | 2019-11-12 | 特斯联（北京）科技有限公司 | Travel forecasting approaches and device |
| CN110633652A (en) * | 2019-08-27 | 2019-12-31 | 宇龙计算机通信科技(深圳)有限公司 | Unexpected situation determination method and device and electronic equipment |
| CN110633680A (en) * | 2019-09-19 | 2019-12-31 | 天津天地伟业机器人技术有限公司 | Video-based people number abnormity detection method |
| CN110675586A (en) * | 2019-09-25 | 2020-01-10 | 捻果科技(深圳)有限公司 | Airport enclosure intrusion monitoring method based on video analysis and deep learning |
| CN110991461B (en) * | 2019-10-29 | 2021-05-25 | 重庆特斯联智慧科技股份有限公司 | Intelligent extraction key security target excavation method and system |
| CN111723633B (en) * | 2019-12-09 | 2022-05-20 | 深圳市鸿逸达科技有限公司 | Personnel behavior pattern analysis method and system based on depth data |
| CN111428703B (en) * | 2020-06-15 | 2020-09-08 | 西南交通大学 | Method for detecting pit leaning behavior of electric power operation and inspection personnel |
| CN112578726A (en) * | 2021-01-06 | 2021-03-30 | 常州百芝龙智慧科技有限公司 | Human body abnormity monitoring equipment for independently learning human body behavior habits |
| CN113658273B (en) * | 2021-08-19 | 2024-04-26 | 上海新氦类脑智能科技有限公司 | Scene self-adaptive target positioning method and system based on space perception |
| CN114419528B (en) * | 2022-04-01 | 2022-07-08 | 浙江口碑网络技术有限公司 | Anomaly identification method and device, computer equipment and computer readable storage medium |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104680525B (en) * | 2015-02-12 | 2017-05-10 | 南通大学 | Automatic human body fall-down detection method based on Kinect depth image |
| JP6812772B2 (en) * | 2016-12-09 | 2021-01-13 | 富士ゼロックス株式会社 | Monitoring equipment and programs |
| CN108664887A (en) * | 2018-03-20 | 2018-10-16 | 重庆邮电大学 | Prior-warning device and method are fallen down in a kind of virtual reality experience |
| CN108629946B (en) * | 2018-06-14 | 2020-09-04 | 清华大学深圳研究生院 | Human body falling detection method based on RGBD sensor |
| CN108876822B (en) * | 2018-07-09 | 2020-11-20 | 山东大学 | Behavior risk degree evaluation method and home security nursing system |
-
2018
- 2018-12-21 CN CN201811569674.8A patent/CN109697830B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN109697830A (en) | 2019-04-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109697830B (en) | Method for detecting abnormal behaviors of people based on target distribution rule | |
| CN110139598B (en) | Monitoring and tracking system, method, article and apparatus | |
| US11688265B1 (en) | System and methods for safety, security, and well-being of individuals | |
| US20190029569A1 (en) | Activity analysis, fall detection and risk assessment systems and methods | |
| JP5143212B2 (en) | Method and apparatus for modeling behavior using probability distribution functions | |
| JP6483928B2 (en) | Monitoring human daily activities | |
| US20190328278A1 (en) | System and method for monitoring respiration | |
| Jansen et al. | 3D human pose recognition for home monitoring of elderly | |
| US6238337B1 (en) | Medical non-intrusive prevention based on network of embedded systems | |
| Wang et al. | Activity density map visualization and dissimilarity comparison for eldercare monitoring | |
| Heydarzadeh et al. | In-bed posture classification using deep autoencoders | |
| Zouba et al. | An activity monitoring system for real elderly at home: Validation study | |
| CN112842258A (en) | Methods, processors, and computer-readable media for event prevention and prediction | |
| JP2005531845A5 (en) | ||
| WO2019013257A1 (en) | Monitoring assistance system and method for controlling same, and program | |
| CN106796746A (en) | Activity monitoring approach and system | |
| CN108882853B (en) | Triggering measurement of physiological parameters in time using visual context | |
| CN112036267A (en) | Target detection method, device, equipment and computer readable storage medium | |
| Tan et al. | Indoor activity monitoring system for elderly using RFID and Fitbit Flex wristband | |
| Vildjiounaite et al. | Unsupervised illness recognition via in-home monitoring by depth cameras | |
| CN114140997A (en) | System and method for monitoring and early warning of retention and physical condition of old people in toilet of nursing home | |
| Bauer et al. | Modeling bed exit likelihood in a camera-based automated video monitoring application | |
| CN116110116A (en) | Human action recognition, storage, and retrieval to maintain privacy through joint edge and cloud computing | |
| Eldib et al. | Behavior analysis for aging-in-place using similarity heatmaps | |
| O'Brien et al. | Design and implementation of an embedded system for monitoring at-home solitary Alzheimer's patients |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |