CN108090458A

CN108090458A - Tumble detection method for human body and device

Info

Publication number: CN108090458A
Application number: CN201711468689.0A
Authority: CN
Inventors: 谢阳阳
Original assignee: Nanjing Science And Technology Ltd Of A Fanda Robot
Current assignee: Nanjing Science And Technology Ltd Of A Fanda Robot
Priority date: 2017-12-29
Filing date: 2017-12-29
Publication date: 2018-05-29
Anticipated expiration: 2037-12-29
Also published as: WO2019128304A1; CN108090458B

Abstract

The embodiment of the present application provides a kind of tumble detection method for human body and device, wherein, this method includes：Obtain target image；By target detection network, human testing is carried out to target image, to determine whether target image is the image comprising human body；In the case where definite target image is comprising the image of human body, pass through convolutional neural networks, tumble identification is carried out to target image, to determine whether the human body in target image is in tumble state, since the program is analyzed and processed by the target image rather than video flowing for obtaining single frames, and go out to include the image of human body using the target detection Network Recognition based on algorithm of target detection, again by carrying out Classification and Identification to the body state in target image based on the convolutional neural networks of sorting algorithm, with the state of human body in recognition target image, so as to solve the identification human body tumble poor accuracy of existing method, the technical issues of efficiency is low, reach accurate, efficiently identify the technique effect of tumble state.

Description

Tumble detection method for human body and device

Technical field

This application involves Human Detection field, more particularly to a kind of tumble detection method for human body and device.

Background technology

With the getting worse of social senilization's trend, people increasingly pay close attention to the daily life safety of old man.It is for example, uncommon Whether prestige can detect in time occurs tumble etc. surprisingly when old man is in alone.Therefore, in actual life, how effectively, accurately Ground detects whether to fall, so as in time to old man carry out relief become one it is important the problem of.

Currently in order to detection is fallen, existing method is to lay multiple cameras in physical activity region in advance mostly, To gather video stream data, then by analyzing in video stream data human body situation of change judge whether human body falls.On Method is stated when it is implemented, due to needing to handle video stream data, analyze, heavy workload, efficiency are low.In addition, pass through Analysis human body situation of change judges whether human body occurs that tumble deterministic process is complex, and error is relatively large.In summary, it is existing There is method when it is implemented, often there is technical issues that identify that tumble poor accuracy, error be big, efficiency.

In view of the above-mentioned problems, currently no effective solution has been proposed.

The content of the invention

The embodiment of the present application provides a kind of tumble detection method for human body and device, to solve to know present in existing method The technical issues of tumble poor accuracy, error be not big, efficiency is low reaches the technology effect for accurately and efficiently identifying tumble state Fruit.

The embodiment of the present application provides a kind of tumble detection method for human body, including：

Obtain target image；

By target detection network, human testing is carried out to the target image, with determine the target image whether be Image comprising human body；

In the case where the definite target image is comprising the image of human body, by convolutional neural networks, to the mesh Logo image carries out tumble identification, to determine whether the human body in the target image is in tumble state.

In one embodiment, the acquisition target image, including：

Gather the acoustic information in target area；

According to the acoustic information, target bearing is determined；

According to the target bearing, dollying head, to obtain the target image.

In one embodiment, the target detection network is established in the following way：

Human body image sample data is obtained, wherein, the human body image sample data includes multiple comprising body state Image；

Mark the human region in the image of the human body image sample data；

It is trained using the human body image sample data after mark, to obtain the target detection based on algorithm of target detection Network.

In one embodiment, the body state includes：State that state that human body stands, human body are seated, human body State, the state of human body on all fours of the couchant state of the state of recumbency, human body, human body at a slant.

In one embodiment, in the case where the definite target image is not comprising the image of human body, the side Method further includes：Reacquire target image.

In one embodiment, the convolutional neural networks are established in the following way：

Satisfactory image is extracted from the human body image sample data as pretreatment sample data；

According to the body state in the image of the pretreatment sample data, by the image in the pretreatment sample data Positive sample data and negative sample data are divided, wherein, the image in the positive sample data includes at least one of：Include The image for the state that human body stands, the image for including the state that human body is seated, the image for including the couchant state of human body, bag The image of state containing human body at a slant；Image in the negative sample data includes at least one of：Include human body The image of the state of recumbency, the image for including the state of human body on all fours；

It is trained using the positive sample data, the negative sample data, to establish to identify body state type The convolutional neural networks.

In one embodiment, the satisfactory image includes：Human region account for figure than be more than 80% figure Picture.

The embodiment of the present application additionally provides a kind of human body falling detection device, including：

Acquisition module, for obtaining target image；

Human detection module for passing through target detection network, human testing is carried out to the target image, to determine institute State whether target image is the image comprising human body；

Tumble identification module, in the case where the definite target image is comprising the image of human body, passing through convolution Neutral net carries out tumble identification to the target image, to determine whether the human body in the target image is in tumble shape State.

In one embodiment, the acquisition module includes：

Sound collector, for gathering the acoustic information in target area；

Locator, for according to the acoustic information, determining target bearing；

Mobile device and camera, wherein, the camera is arranged in the mobile device, and the mobile device is used for root According to the target bearing, the mobile camera, the camera is used to obtain target image.

In one embodiment, described device further includes alarm module, at the human body in target image is determined Sent in the case of tumble state alarm and/or, send information warning.

In the embodiment of the present application, analyzed and processed by the target image rather than video flowing that obtain single frames, and profit The image for including human body is first identified with the target detection network based on algorithm of target detection, then by based on sorting algorithm Convolutional neural networks carry out Classification and Identification to the body state in target image, to identify the specific shape of human body in target image State so as to solve the technical issues of tumble poor accuracy being identified present in existing method, error is big, efficiency is low, reaches Accurately, the technique effect of tumble state is efficiently identified.

Description of the drawings

It in order to illustrate the technical solutions in the embodiments of the present application or in the prior art more clearly, below will be to embodiment or existing There is attached drawing needed in technology description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this Some embodiments described in application, for those of ordinary skill in the art, in the premise of not making the creative labor property Under, it can also be obtained according to these attached drawings other attached drawings.

Fig. 1 is the process flow schematic diagram of the tumble detection method for human body provided according to the application embodiment；

Fig. 2 is the composition structure diagram of the human body falling detection device provided according to the application embodiment；

Fig. 3 is that the electronic equipment of the tumble detection method for human body provided based on the application embodiment forms structural representation Figure；

Fig. 4 is that the tumble detection method for human body and device provided in a Sample Scenario using the application embodiment is set The composition structure diagram of the human body fall detection robot of meter；

Fig. 5 is the flow signal for carrying out human body fall detection using human body fall detection robot in a Sample Scenario Figure.

Specific embodiment

It is in order to make those skilled in the art better understand the technical solutions in the application, real below in conjunction with the application The attached drawing in example is applied, the technical solution in the embodiment of the present application is clearly and completely described, it is clear that described implementation Example is merely a part but not all of the embodiments of the present application.Based on the embodiment in the application, this field is common Technical staff's all other embodiments obtained without making creative work should all belong to the application protection Scope.

It is acquisition video stream data mostly when being embodied in view of existing method, and video stream data is analyzed, Processing since the data volume to be analyzed is big, causes occupancy resource more, and efficiency is low.In addition, existing method is to pass through analysis mostly Whether human body variation detection human body falls, and this identification method itself is complex, precision is poor, is susceptible to error.To sum up It understands, existing method often there is technical issues that identify tumble poor accuracy, efficiency when being embodied.It is above-mentioned for generating The basic reason of technical problem, the application consideration can obtain the image data of single frames rather than video stream data carries out specifically Analysis, to effectively reduce data processing amount；In addition, the characteristics of for image data and advantage, by analyzing in image Body state rather than human body change to judge whether human body falls, and it is accurate to solve identification tumble present in existing method The technical issues of exactness is poor, error is big, efficiency is low has reached technique effect that is accurate, efficiently identifying tumble state.

Based on above-mentioned thinking thinking, the embodiment of the present application provides a kind of tumble detection method for human body.Referring specifically to Fig. 1 The process flow schematic diagram of the shown tumble detection method for human body provided according to the application embodiment.The embodiment of the present application carries The tumble detection method for human body of confession, when it is implemented, may comprise steps of.

S11：Obtain target image.

In the present embodiment, in order to reduce calculation amount, the occupancy to computing resource is reduced, when it is implemented, can obtain The target image of single frames rather than the video flowing of existing method acquisition are taken, subsequently make a concrete analysis of, handle.Compared to video Stream, as long as analyzed an independent two field picture, detected, identified when target image subsequent analysis for single frames, processing, such as This, can efficiently reduce calculation amount, reduce calculating cost, improve recognition speed.

In one embodiment, in order to be further reduced the workload in follow-up human testing stage, it is avoided to obtain Image comprising human body repeatedly carries out image acquisition, during target image is obtained, when it is implemented, can be as far as possible excellent Effective image is first obtained as target image.Wherein, above-mentioned effective image specifically can be understood as including the image of human body.Phase It answers, can be invalid image by the image understanding not comprising human body.In this way, it can be avoided to obtain what can subsequently be used Image comprising human body is repeated several times the acquisition for carrying out target image, has helped to improve treatment effeciency.

In one embodiment, in order to efficiently obtain above-mentioned effective image, above-mentioned acquisition target image is specific real Shi Shi can include herein below：

S11-1：Gather the acoustic information in target area；

S11-2：According to the acoustic information, target bearing is determined；

S11-3：According to the target bearing, dollying head, to obtain the target image.

In the present embodiment, above-mentioned target bearing can be specifically the direction of sound source.Above-mentioned direction has larger Probability there are personnel's movements.Therefore, in above-mentioned target bearing, compared with other orientation, bag is got with larger probability Image containing human body, i.e. effective image.

In the present embodiment, when it is implemented, microphone array may be employed gathers target area as sound collector Acoustic information in domain；And pass through locator according to the acoustic information gathered, the direction of sound source is determined, by the above sound The direction in source is determined as above-mentioned target bearing.Certainly, it is necessary to which explanation, above-mentioned cited microphone array are intended merely to The application embodiment is better described.When it is implemented, it can also select to use other suitable sound as the case may be Collector.

In the present embodiment, when it is implemented, above-mentioned camera can be specifically provided in mobile device, that is, image It is not in fixed setting target area that head, which can move,.For example, camera can be arranged on by pulley and motor form Mobile device on.In this way, camera can neatly be moved by mobile device in target area, so as to have Effect ground expands the regional extent of acquisition target image, realizes and more target images are obtained in the detection range of bigger.I.e. originally Apply for that embodiment is different from the mode that camera is used in existing method using the mode of camera.Specifically, in existing method It is that camera is fixedly installed on some fixed position in the target area during using camera, to gather video stream data.It presses It is limited in scope according to what the single camera of mode using camera in existing method can be detected, in order to improve total detection range It then needs to lay camera respectively in multiple positions of target area.In this way, implementation cost can be increased again.And the embodiment of the present application The mode using camera of middle offer is then to set camera on the mobile device, and then according to circumstances can pass through shifting Dollying head is to obtain the target image of different position in target area in real time for dynamic device, so as to using one or few The camera of amount realizes the acquisition to target image interior in a big way, reduces implementation cost.Simultaneously as camera can be with It is mobile, the angle and distance of camera and human body can be adjusted according to the concrete condition of human body, so as to obtain quality higher Target image, subsequently can more accurately to carry out tumble identification.Certainly it is above-mentioned cited, it is necessary to explanation Mobile device is intended merely to that the application embodiment is better described.It when it is implemented, can also as the case may be and precision It is required that select to use other removable frames as mobile device, such as mobile robot, telecar etc., so as to spirit The position of ground living dollying head.In this regard, the application is not construed as limiting.

In the present embodiment, when it is implemented, the sound letter in microphone array acquisition target area can be first passed through Breath；The source direction of sound is determined by locator, and using the direction as the direction that there may be personnel activity, i.e. target Orientation；Again by mobile device, according to identified target bearing, camera is moved to the source position of sound, so as to To get the relatively high effective image of quality by common camera.

S12：By target detection network, human testing is carried out to the target image, is with the definite target image No is the image comprising human body.

In the present embodiment, first human testing is carried out to target image after target image is got, to determine institute Whether the target image to be analyzed is the image comprising human body, i.e. effective image.Subsequently can only to be carried out to effective image The tumble identification of next step.Using the image not comprising human body as invalid image, identified without the tumble of next step.So as to By excluding the image not comprising human body in advance, to avoid carrying out the image for not including human body insignificant identification of falling, drop The data processing amount of low identification of falling, further improves processing speed.

In one embodiment, when it is implemented, can be the image not comprising human body in the definite target image In the case of, target image is reacquired, to be supervised in real time to the region there may be personnel activity in target area It surveys.

In one embodiment, it is contemplated that data acquired, to be analyzed are the images of single frames, it is contemplated that image point The specific feature of analysis, in order to quickly and accurately determine whether target image is the image comprising human body.Specific implementation When, can human testing be carried out to acquired target image by the target detection network based on algorithm of target detection, with true Whether the image that sets the goal is the image comprising human body.

In one embodiment, can step S12 execution before, beforehand through in the following manner establish it is above-mentioned be used for into The target detection network of row human testing：

S1：Human body image sample data is collected, the human body image sample data includes the human body image under different conditions；

S2：Human region is marked in the human body image sample data；

S3：It is trained using the human body image sample data after mark, to obtain the target based on algorithm of target detection Detect network.

In the present embodiment, above-mentioned algorithm of target detection can be specifically a kind of detection algorithm based on deep learning, Also referred to as SSD (Single Shot MultiBox Detector) algorithm.The core of the algorithm is that volume is being used on characteristic pattern Core is accumulated to predict the classification fraction of default bounding boxes a series of, offset, and then can quickly and accurately be examined Measure whether target image to be detected is the effective image for including human body.

In the present embodiment, in order to coordinate it is subsequent fall identification, it is desirable that the human body image sample data specifically may be used To include the multiple images of body state under different conditions.

In one embodiment, in order to comprehensively in view of a variety of different body state situations, above-mentioned human body State can specifically include：The couchant state of state that state that human body stands, human body are seated, the state of human body recumbency, human body, Human body state at a slant, the state of human body on all fours etc..It in this way, when it is implemented, can be by algorithm of target detection to more Kind learnt comprising the image of different body states, so as to establish can detect simultaneously, identify it is a variety of comprising difference Body state image.

In the present embodiment, when it is implemented, SSD target detections network can be utilized in human body image sample data Human region is calibrated in image, can subsequently to carry out and the relevant training of human region feature recognition.

It in one embodiment, can first structure before being trained using the human body image sample data after mark SSD target detection networks are built, that is, the initial model of target detection.It when it is implemented, can be in tensorflow frames The above-mentioned SSD target detections network of upper structure, and using inception_v2 characterized by extractor.

In one embodiment, the above-mentioned human body image sample data using after mark is trained, to be based on The target detection network of algorithm of target detection, when it is implemented, herein below can be included：It is decent using the human figure after mark Notebook data is as input data, and to above-mentioned SSD target detections network, i.e. the initial model of target detection is trained, to obtain Trained target detection network；Further according to human body image sample data and required precision, to above-mentioned trained target detection Network is adjusted optimization, to obtain the SSD networks for human testing, i.e., the described target detection based on algorithm of target detection Network.

S13：In the case where the definite target image is comprising the image of human body, by convolutional neural networks, to institute It states target image and carries out tumble identification, to determine whether the human body in the target image is in tumble state.

In one embodiment, in order to quickly and accurately identifying body state from the image comprising human body Corresponding body state is in tumble state or is not on tumble state, volume may be employed for example, telling human body Product neutral net carries out tumble identification to target image, to determine whether the human body in target image is in tumble state.

It in the present embodiment, can be by training when it is implemented, in view of the related thought of image classification algorithms (CNN) Good convolutional neural networks will determine the target image comprising human body as input data, by upper as tumble identification model It states tumble identification model and identifies whether the human body in above-mentioned target image is in tumble state, so as to according to single-frame images Judge whether human body falls.

In one embodiment, identified when it is implemented, can be established before S13 is carried out beforehand through in the following manner Tumble precision is higher, the faster convolutional neural networks of recognition speed：

S1：Human body image sample data is obtained, wherein, the human body image sample data includes the human body under different conditions Image；

S2：Satisfactory image is extracted from the human body image sample data as pretreatment sample data；

S3：It, will be in the pretreatment sample data according to the body state in the image of the pretreatment sample data Image divides positive sample data and negative sample data, wherein, the image in the positive sample data includes at least one of：Bag The image for the state that stand containing human body, the figure for including the image for the state that human body is seated, including the couchant state of human body Picture, the image for including the state of human body at a slant；Image in the negative sample data includes at least one of：Include The image of the state of human body recumbency, the image for including the state of human body on all fours；

S4：It is trained using the positive sample data, the negative sample data, to establish to identify body state class The convolutional neural networks of type.

In one embodiment, it is contemplated that in order to establish the more accurate target detection net based on algorithm of target detection Network has just included multiple images for including body state in above-mentioned human body image sample data.Therefore, in present embodiment In, human body image sample data can be based on and extract satisfactory image, as pretreatment sample data.

In the present embodiment, obtain pre-process sample data after, it is necessary to first according to fall and non-tumble two states, Classify to the image in pretreatment sample data.Specifically, it can will pre-process the figure that non-tumble is characterized in sample data Picture, including：Include the image for the state that human body stands, include the image for the state that human body is seated, to include human body couchant The image of state, the images such as image that include the state of human body at a slant be divided into positive sample data, i.e. positive image data Collection.The image that tumble is characterized in sample data will be pre-processed, including：Include the image of the state of human body recumbency, comprising someone The images such as the image of the state of body on all fours are divided into negative sample data, i.e. negative image data set.In this way, it can subsequently be directed to human body Tumble state and human body non-tumble two states identification, specifically instructed using corresponding above two sample data Practice study, to establish the higher convolutional neural networks of accuracy of identification.

In one embodiment, it is trained using the positive sample data, the negative sample data, is used for establishing The convolutional neural networks of body state type are identified, when it is implemented, herein below can be included：Build initial convolution Neutral net；By the use of above-mentioned positive sample data and negative sample data as input data to above-mentioned initial convolutional neural networks into Recognition training of the row on the non-tumble state of the tumble state and human body of human body, with arrive accuracy of identification is higher, recognition speed compared with Fast convolutional neural networks.And then the convolutional neural networks can be utilized to identify the body state pair in target image exactly Answer whether be human body tumble state.If identify that the body state in obtained target image corresponds to the tumble shape of human body State then may determine that human body is in tumble state；If identify that the body state in obtained target image corresponds to human body Non- tumble state then may determine that human body is not on tumble state.

In one embodiment, during above-mentioned convolutional neural networks are established, when it is implemented, can also include Herein below：

S1：Obtain the image sample data not comprising human body；

S2：Using the image sample data not comprising human body, error detection training is carried out to the convolutional neural networks.

In the present embodiment, trained by above-mentioned error detection, can first identify and filter out the mesh not comprising human body Logo image improves treatment effeciency of the convolutional neural networks when carrying out fall detection.

In the embodiment of the present application, compared to the prior art, flowed by the target image rather than video that obtain single frames Row analyzing and processing, and first identify the image for including human body using the target detection network based on algorithm of target detection, then lead to It crosses and Classification and Identification is carried out to the body state in target image based on the convolutional neural networks of sorting algorithm, to identify target figure The particular state of human body as in, so as to solve identification tumble poor accuracy present in existing method, error is big, efficiency is low Technical problem has reached technique effect that is accurate, efficiently identifying tumble state.

In one embodiment, tumble recognition training is carried out in order to extract to be applicable in from human body image sample data Pretreatment sample data, the satisfactory image can specifically include：Human region account for figure than be more than 80% figure Picture.In this way, the sample data for being applicable in tumble recognition training can be extracted from human body image sample data, avoid and resurvey The sample data of tumble identification is carried out, trained cost is reduced, improves learning efficiency.

In one embodiment, above-mentioned initial convolutional neural networks can be specifically inception_v3 networks.Its In, above-mentioned inception_v3 networks are specifically a kind of convolutional neural networks suitable for image identification.Certainly, it is necessary to explanation It is that above-mentioned cited convolutional neural networks are intended merely to that the application embodiment is better described.It when it is implemented, can also As the case may be other suitable convolutional neural networks are used with the selection of the specific features of identification.In this regard, the application does not limit It is fixed.

In one embodiment, it is trained using the positive sample data, the negative sample data, to establish use Before the convolutional neural networks of identification body state type, the method further includes, according to initial convolutional Neural net Network pre-processes the image in the positive sample data, negative sample data, so that the positive sample data, negative sample Image in data matches with initial convolutional neural networks.Specifically, for example, it is in initial convolutional neural networks Inception_v3 networks, above-mentioned pretreatment can specifically include：To the image in the positive sample data, negative sample data into Row image transforms to specified size, for example, transforming to the size of 299 × 299 pixels.

In one embodiment, further contemplate when carrying out tumble identification using convolutional neural networks, actually Only need the non-tumble state of differentiation two types, i.e. the tumble state of human body and human body.Therefore, according to convolutional neural networks institute The complexity of Classification and Identification is wanted, takes into account to improve treatment effeciency, reduces the occupancy to computing resource and waste, is being established just During the convolutional neural networks of beginning, first the convolutional neural networks can be carried out to simplify improvement.Wherein, above-mentioned simplified improvement is specific It can include：Reduce convolutional neural networks the number of plies and/or, reduce the convolution kernel numbers of convolutional neural networks.It can pass through The number of plies for individually reducing convolutional neural networks either individually reduces the convolution kernel number of convolutional neural networks or rolls up less simultaneously The number of plies of product neutral net and the convolution kernel number of reduction convolutional neural networks carry out above-mentioned convolutional neural networks to simplify improvement, While taking into account accuracy of identification so as to reach, the occupancy to computing resource is reduced, improves treatment effeciency.

In one embodiment, it is above-mentioned right in the case where convolutional neural networks is inception_v3 networks Simplified improve of inception_v3 networks can specifically include：By the number of plies of inception_v3 networks by 11 layers (or structures) Delete as 6 layers or 5 layers and/or, delete the convolution kernel number in inception_v3 networks, and then the volume that can be simplified Product neutral net.

In one embodiment, can be built in the following way during the convolutional neural networks specific implementation of above-mentioned simplification It is vertical：

S1：Existing inception_V3 networks are carried out to simplify processing.

In the present embodiment, specifically, the last 5 inception structures of inception_V3 networks can be deleted, obtain Inception_v3 networks after to simplification.

S2：Using the sample data training of pretreatment into the inception_v3 networks after simplification, obtain can be used for falling The parameter model Fa1 of detection.

S3：The convolution kernel number of all convolutional layers of the inception_v3 networks after simplification is reduced to successively original 2/3rds, while parameter model Fa1 is changed, it adapts it to reduce the network after convolution kernel number.

S4：The sample data for being continuing with pretreatment is trained amended parameter model Fa1, and to amended Fa1 is finely adjusted, and obtains the parameter model Fa2 available for fall detection.

S5：Verify above-mentioned parameter model Fa2, according to check results to above-mentioned parameter model Fa2 according to S4 included by instruction Practice and fine tuning operation is adjusted, with the convolutional neural networks being simplified.

In the present embodiment, above-mentioned verification can specifically include：Compare after convolution kernel is reduced and fall with the network before reduction The accuracy rate detected if the accuracy rate of fall detection is not decreased obviously, can continue above-mentioned reduction convolution kernel, And training accordingly and the operation finely tuned are carried out, with the convolutional neural networks more simplified；If the accuracy rate of fall detection It is decreased obviously, then can be with deconditioning and the operation of fine tuning, and last network and parameter model are determined for falling Detection, i.e., as the convolutional neural networks for fall detection.

In one embodiment, after determining that the human body in target image is in tumble state, it can be determined that target area In human body fall, and then alarm can be sent, someone in target area to be prompted to fall.Wherein, it is above-mentioned to send alarm It can specifically include sounding the alarm that someone is reminded to fall by buzzer；It can also be by communication apparatus to target area Director or the medical staff on periphery send warning message (for example, alarm short message), ask treatment etc. in time.Certainly, on Cited a variety of modes for sending alarm are stated to be intended merely to that the application embodiment is better described.When it is implemented, also may be used It is alarmed in a manner of as the case may be other being selected suitably to send alarm.In this regard, the application is not construed as limiting.

It can be seen from the above description that tumble detection method for human body provided by the embodiments of the present application, single by obtaining The target image rather than video flowing of frame are analyzed and processed, and are first known using the target detection network based on algorithm of target detection Do not go out to include the image of human body, then by based on the convolutional neural networks of sorting algorithm to the body state in target image into Row Classification and Identification to identify the particular state of human body in target image, is fallen so as to solve identification present in existing method The technical issues of poor accuracy, error are big, efficiency is low has reached technique effect that is accurate, efficiently identifying tumble state； Further through acquisition acoustic information to determine target bearing, and according to target bearing dollying head to gather effective target figure Picture effectively expands the detection range of fall detection, improves the accuracy for obtaining effective target image, improves detection effect Fruit improves user experience；Also it is used as sample data by obtaining the image comprising a variety of body states, to establish target detection Network, convolutional neural networks improve the precision that human body tumble is identified according to single-frame images；Also by according to the shape to be identified The complexity of state type has carried out convolutional neural networks corresponding simplified improvement, has improved efficiency of the practice, reduces to computing The occupancy of resource.

Based on same inventive concept, a kind of human body falling detection device is additionally provided in the embodiment of the present invention, as following Described in embodiment.Since the principle that human body falling detection device solves the problems, such as is similar to tumble detection method for human body, device Implementation may refer to the implementation of tumble detection method for human body, overlaps will not be repeated.It is used below, term " unit " Or " module " can realize the combination of the software and/or hardware of predetermined function.Although the described device of following embodiment compared with It is realized goodly with software, but the realization of the combination of hardware or software and hardware is also what may and be contemplated.It refers to Fig. 2, is a kind of composition structure diagram of human body falling detection device provided by the embodiments of the present application, which can specifically wrap It includes：Acquisition module 21, human detection module 22, tumble identification module 23, are below specifically described the structure.

Acquisition module 21 specifically can be used for obtaining target image；

Human detection module 22 specifically can be used for through target detection network, and human body inspection is carried out to the target image It surveys, to determine whether the target image is the image comprising human body；

Tumble identification module 23 specifically can be used in the situation for determining that the target image is the image comprising human body Under, by convolutional neural networks, tumble identification is carried out to the target image, whether to determine the human body in the target image In tumble state.

In the present embodiment, it is necessary to which explanation is that above-mentioned human body falling detection device can be specifically that one kind can be realized The human body fall detection robot of human body fall detection.Above-mentioned human body fall detection robot specifically can be applied to family, doctor A variety of places such as institute, market to detect above-mentioned place in real time, find that the personnel in place fall, to carry out in time in time Alarm carries out related relief in time.

In one embodiment, in order to expand detection range, effective target image, the acquisition are efficiently obtained Module 21 can specifically include following structural unit：

Sound collector specifically can be used for gathering the acoustic information in target area；

Locator specifically can be used for, according to the acoustic information, determining target bearing；

Mobile device and camera, wherein, the camera can be specifically arranged in the mobile device, the mobile dress It puts and specifically can be used for according to the target bearing, the mobile camera, the camera specifically can be used for obtaining target Image.

In the present embodiment, above-mentioned mobile device can specifically include pulley and motor.It in this way, when it is implemented, can With the mobile device by carrying pulley and motor, camera is driven to be moved as target bearing, preferably to obtain effective mesh Logo image.Certainly, it is necessary to which explanation, above-mentioned cited mobile device are intended merely to better illustrate the application embodiment party Formula.When it is implemented, above-mentioned mobile device can also be other kinds of movable equipment, for example, movable machine people, distant Control automobile etc..In this regard, the application is not construed as limiting.

In the present embodiment, above-mentioned effective target image can be specifically the image for including human body.By can be with Above-mentioned mobile device can obtain effective target image as much as possible, so as to reduce according to target bearing dollying head The workload of human detection module 22 improves work efficiency.

In one embodiment, in order to detect human body fall after alarm in time with to tumble personnel carry out and When give treatment to, described device specifically can also include alarm module, for sending alarm.

In one embodiment, above-mentioned alarm module can specifically include buzzer, in this way, the alarm module is specific During implementation, can by buzzer determine target image in be considered at tumble state in the case of send alarm.

In one embodiment, above-mentioned alarm module specifically can also including sender unit etc. communication apparatus, in this way, The alarm module by communication apparatus such as sender units in target image is determined when it is implemented, can be considered at In the case of tumble state alarm is sent to relevant person in charge (such as guardian either ensure public security by market) or periphery medical staff Information relevant person in charge or periphery medical staff someone to be prompted to fall, is given treatment to as early as possible.

In one embodiment, described device specifically can also establish module, target detection including target detection network Network can perform when establishing module specific implementation according to following procedure：Human body image sample data is obtained, wherein, the human body Image sample data includes multiple images for including body state；Mark the human body in the image of the human body image sample data Region；It is trained using the human body image sample data after mark, to obtain the target detection net based on algorithm of target detection Network.

In one embodiment, the body state can specifically include：The shape that state that human body stands, human body are seated State, the state of human body on all fours of the couchant state of state, the state of human body recumbency, human body, human body at a slant etc..Certainly, it is necessary to Illustrate, above-mentioned cited body state is intended merely to that the application embodiment is better described.When it is implemented, also may be used With as the case may be and requirement, introduce other states in addition to above-mentioned cited state as body state.In this regard, this Application is not construed as limiting.

In one embodiment, above-mentioned human detection module 22 is connected with acquisition module 21, when it is implemented, human body is examined Acquisition module can be sent information in the case where the definite target image is not comprising the image of human body by surveying module 22 21, target image is reacquired by acquisition module 21.

In one embodiment, described device specifically can also establish module including convolutional neural networks, for establishing For identifying the convolutional neural networks of body state type, wherein, the convolutional neural networks, which establish module, can specifically include：

Acquiring unit specifically can be used for obtaining human body image sample data, wherein, the human body image sample data bag Include multiple images for including body state；

Extraction unit specifically can be used for from the human body image sample data extracting satisfactory image as pre- Handle sample data；

Division unit specifically can be used for the body state in the image according to the pretreatment sample data, by described in Image division positive sample data and the negative sample data in sample data are pre-processed, wherein, the image in the positive sample data Including at least one of：Include the image for the state that human body stands, include the image for the state that human body is seated, include The image of the couchant state of human body, the image for including the state of human body at a slant；Image in the negative sample data includes At least one of：Include the image of the state of human body recumbency, the image for including the state of human body on all fours；

Unit is established, specifically can be used for being trained using the positive sample data, the negative sample data, to establish For identifying the convolutional neural networks of body state type.

In one embodiment, the convolutional neural networks are established module and can also specifically be included：

Error detection training unit specifically can be used for obtaining the image sample data not comprising human body；And described in not Image sample data comprising human body carries out error detection training to the convolutional neural networks.

In the present embodiment, it is described to meet the requirements in order to establish and train the higher convolutional neural networks of accuracy Image can specifically include：Human region accounts for the image of figure than being more than 80% etc..

Each embodiment in this specification is described by the way of progressive, identical similar portion between each embodiment Point just to refer each other, and the highlights of each of the examples are difference from other examples.It is real especially for system For applying example, since it is substantially similar to embodiment of the method, so description is fairly simple, related part is referring to embodiment of the method Part explanation.

It should be noted that system, device, module or unit that the above embodiment illustrates, it specifically can be by computer Chip or entity are realized or realized by having the function of certain product.For convenience of description, in the present specification, retouch It is divided into various units when stating apparatus above with function to describe respectively.It certainly, when implementing the application can be the function of each unit It realizes in the same or multiple software and or hardware.

In addition, in the present specification, adjective can be only used for an element or dynamic such as first and second Make to distinguish with another element or action, without requiring or implying any actual this relation or order.Permit in environment Perhaps in the case of, one in only element, component or step is should not be interpreted as limited to reference to element or component or step (s) It is a, and can be one or more of element, component or step etc..

It can be seen from the above description that human body falling detection device provided by the embodiments of the present application, single by obtaining The target image rather than video flowing of frame are analyzed and processed, and are known first with the target detection network based on algorithm of target detection Do not go out to include the image of human body, then by based on the convolutional neural networks of sorting algorithm to the body state in target image into Row classification, it is accurate so as to solve identification tumble present in existing method to identify the particular state of human body in target image The technical issues of exactness is poor, efficiency is low has reached technique effect that is accurate, efficiently identifying tumble state；Further through acquisition Acoustic information is effectively expanded with determining target bearing according to target bearing dollying head with gathering effective target image The big detection range of fall detection, improves the accuracy for obtaining effective target image, improves detection result.

The application can specifically refer to shown in Fig. 3 real based on the application embodiment further provides a kind of electronic equipment The electronic equipment composition structure diagram for the tumble detection method for human body that the mode of applying provides, the electronic equipment can specifically include Input equipment 31, processor 32, memory 33.Wherein, the input equipment 31 specifically can be used for receiving acquired target Image.The processor 32 specifically can be used for through target detection network, human testing be carried out to the target image, with true Whether the fixed target image is the image comprising human body；In the situation for determining that the target image is the image comprising human body Under, by convolutional neural networks, tumble identification is carried out to the target image, whether to determine the human body in the target image In tumble state.The memory 33 specifically can be used for storing the target image, the target detection network, the volume Intermediate data generated in product neutral net and detection process etc..

In the present embodiment, the input equipment can be specifically that information exchange is carried out between user and computer system One of main device.The input equipment can include keyboard, mouse, camera, scanner, light pen, writing input board, language Sound input unit etc.；Input equipment is used to initial data be input to the programs for handling these numbers in computer.The input Equipment, which can also obtain, receives the data that other modules, unit, equipment transmit.The processor can be by any appropriate Mode is realized.For example, processor can take such as microprocessor or processor and storage that can be performed by (micro-) processor Computer readable program code (such as software or firmware) computer-readable medium, logic gate, switch, application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), programmable logic controller (PLC) and embedded microcontroller Form etc..The storage implement body can be used to protect stored memory device in modern information technologies.The storage Device can include many levels, in digital display circuit, as long as can preserve binary data can be memory；In integrated electricity Lu Zhong, a circuit with store function without physical form are also memory, such as RAM, FIFO；In systems, have The storage device for having physical form is also memory, such as memory bar, TF card.

In the present embodiment, the function and effect of electronic equipment specific implementation, can compare with other embodiment It explains, details are not described herein.

A kind of computer storage media based on tumble detection method for human body, institute are additionally provided in this theory application embodiment It states computer storage media and is stored with computer program instructions, realization is performed in the computer program instructions：Obtain mesh Logo image；By target detection network, human testing is carried out to the target image, to determine whether the target image is bag Image containing human body；In the case where the definite target image is comprising the image of human body, by convolutional neural networks, to institute It states target image and carries out tumble identification, to determine whether the human body in the target image is in tumble state.

In the present embodiment, above-mentioned storage medium includes but not limited to random access memory (Random Access Memory, RAM), read-only memory (Read-Only Memory, ROM), caching (Cache), hard disk (Hard Disk Drive, HDD) or storage card (Memory Card).The memory can be used for storing computer program instructions.Network leads to Believe that unit can be according to standard setting as defined in communication protocol, for carrying out the interface of network connection communication.

In the present embodiment, the function and effect of the program instruction specific implementation of computer storage media storage, can To compare explanation with other embodiment, details are not described herein.

It is embodied at one in Sample Scenario, provides tumble detection method for human body using the application and device designs accordingly Human body fall detection robot, and the application human body fall detection robot carries out specific human body fall detection.It is specific real The process of applying can refer to herein below.

In the present embodiment, above-mentioned human body fall detection detection robot can specifically refer to shown in Fig. 4 at one The human body fall detection machine that the tumble detection method for human body and device provided in Sample Scenario using the application embodiment designs The composition structure diagram of device people.The robot can specifically use auditory localization module to position human body general orientation (i.e. target Orientation), camera gathered data (i.e. target image) is recycled, the human body based on single-frame images is realized by deep learning algorithm Fall detection.Wherein, the fall detection robot include specifically can with mobile machine human body 12, camera module 13, Multiple work(such as alarm module 14 (optional), auditory localization module 15 (optional), human detection module 16 and tumble identification module 17 It can module.

When it is implemented, above-mentioned auditory localization module 15 specifically can be used for judging human body general orientation, and utilize camera shooting Head module 13 shoots single-frame images, human detection module 16 and the specific image that can be used for according to shooting of tumble identification module 17 Judge whether people falls, and result is transferred to packaged type robot body 12；The packaged type robot body if falling 12 can be by controlling alarm module 14 to alarm.

Wherein, the packaged type robot body 12 includes at least：The structures such as robot body, motor and pulley. The camera module 13 is specific to be can be used for gathering single image, and is sent into human detection module 16 to judge whether to deposit At human body (judging whether image is the image that includes human body).The alarm module 14 can at least include mobile communication work( It can be with 110 warning functions.In this way, when it is implemented, transmission and the picture of mobile communication function realization tumble information can be utilized Information is sent, and realizes 110 alarms by 110 warning functions to succour in time.The auditory localization module 15 specifically can be with The source direction of sound is judged by microphone array, easily to find people.The human detection module 16 specifically may be used To realize human testing by SSD algorithm of target detection in deep learning.The tumble identification module 17 passes through deep learning Middle convolutional neural networks realize tumble state recognition.

In the present embodiment, it is necessary to which explanation, above-mentioned human body fall detection robot may be considered a kind of specific Human body falling detection device, implement cardinal principle it is identical with human body falling detection device.

When it is implemented, can refer to Fig. 5 shown in a Sample Scenario using human body fall detection robot into The flow diagram of pedestrian's body fall detection carries out human body fall detection using human body fall detection robot.When it is implemented, It may comprise steps of：

S1：Optionally, the general direction of people is found by packaged type robot combination auditory localization module；

S2：Single-frame images, and incoming packaged type robot are gathered by camera module；

S3：The single-frame images collected is passed to by human detection module by packaged type robot body；

S4：By human detection module judge acquisition image in whether presence of people.If any then continuing 5；If no, Then return to 1；

S5：The human region detected is sent into tumble identification module, judges whether human body falls；

S6：It will identify that obtained result information is transferred to packaged type robot body；

S7：If falling, continue 8；If not falling, 2 are returned；

S8：Alarm is performed, it will be on the information of tumble and image transmitting to the mobile phone or other-end of connection.

In the present embodiment, above-mentioned human detection module is realized based on SSD algorithm of target detection in deep learning. Detection module can carry out SSD Algorithm for Training before image detection is carried out according to following flow：

S1：Collect the image data (ratio that people accounts for picture is unlimited) (i.e. human body image sample data) for including human body.Cause It to need to detect human region, and needs to detect the human body under any state, therefore the image data collected can specifically include Human body under different conditions such as stands, is couchant, recumbency, human body at a slant.

S2：The image data being collected into is labeled.SSD target detections network can calibrate human body in human testing Region, therefore in training need first to provide the region of human body in image data.

S3：Build SSD target detection networks.When it is implemented, the inspection of SSD targets can be built on tensorflow frames Survey grid network, and the extractor characterized by inception_v2.

S4：SSD target detection networks are trained with the image data handled well, and utilize existing trained parameter mould Type is finely adjusted it, obtains the SSD networks (i.e. target detection network) for human testing.

In the present embodiment, above-mentioned fall detection module can specifically include the convolutional Neural net in a kind of deep learning Network.Tumble identification module specifically can carry out convolutional neural networks training before image identification is carried out by following flow：

S1：Comprising human body image data, (ratio that people accounts for picture is more than 80%, i.e. human detection module detects for collection Human region picture) (pre-processing sample data).

S2：Build positive and negative image data sample.Positive sample (i.e. positive sample data) includes the human figure of all non-tumbles Piece, i.e. body state to stand, holding, at a slant etc.；The picture that negative sample (i.e. negative sample data) is included all is people's tumble Picture afterwards, i.e. body state for recumbency, on all fours etc..

S3：Image in preprocessing image data sample.It specifically, can be big to specifying by all image data transformations It is small, such as 299 × 299 pixel sizes.

S4：Build convolutional neural networks.Specifically, inception_v3 networks may be employed in above-mentioned tumble identification module.

In the present embodiment, it is necessary to supplement, for the demand for identification of falling, usually used inception_v3 Network in computing resource exist waste.Therefore when building inception_v3 networks, simplified modification has been carried out to it, has been had Body, which simplifies to improve, includes herein below：

S4-1：While recognition accuracy is ensured, inception structures, such as the number of plies are reduced.Reach and simplify network Structure improves recognition speed, has saved the effect of computing resource.

S4-2：Ensure recognition accuracy simultaneously, reduce convolution kernel number.Reaching reduces network size, improves knowledge Other speed, has saved the effect of computing resource.

S5：Image data sample input inception_v3 networks after pretreatment are trained, obtain identification net of falling Network (i.e. convolutional neural networks).

In the present embodiment, specifically human body fall detection is carried out using above-mentioned human detection module and fall detection module When can specifically include herein below：

S1：The picture collected is inputted into SSD target detection networks, the region where human body is detected, and result is preserved.

S2：All human regions detected are transformed into specified size, such as 299 × 299 pixel sizes.

S3：To be obtained in S2 in the inception_v3 models that result is input to, in a manner of multithreading simultaneously into Row prediction, provides recognition result.

S4：It is recorded a demerit according to the identification, display fall detection is as a result, determine whether human body falls.

After carrying out multiple fall detection test to above-mentioned human body fall detection robot, analysis is found：Above-mentioned human body is fallen Robot is detected due to the use of algorithm of target detection SSD and image classification algorithms CNN, single frames can be passed through under complex scene Image realizes the fall detection of degree of precision, and can implement alert process.Human testing is inaccurate in the existing method overcome The problem of；Simultaneously because fall detection only need not be can be achieved with single-frame images by the analyzing and processing to video flowing, reduce Calculation amount, improves detection efficiency；And with the artificial carrier of packaged type machine, it can be achieved that comprehensive monitoring.

By above-mentioned Sample Scenario, tumble detection method for human body provided by the embodiments of the present application and device are demonstrated, is passed through The target image rather than video flowing for obtaining single frames are analyzed and processed, and first with the target detection based on algorithm of target detection Network Recognition goes out to include the image of human body, then by based on the convolutional neural networks of sorting algorithm to the human body in target image State is classified, and to identify the particular state of human body in target image, is solved present in existing method and is identified really Tumble poor accuracy, the technical issues of efficiency is low, have reached technique effect that is accurate, efficiently identifying tumble state.

Although mentioning different specific embodiments in teachings herein, the application is not limited to be industry Standard or the described situation of embodiment etc., some professional standards or the implementation base described using self-defined mode or embodiment On plinth embodiment amended slightly can also realize above-described embodiment it is identical, it is equivalent or it is close or deformation after it is anticipated that Implementation result.It, still can be with using these modifications or the embodiment of deformed data acquisition, processing, output, judgment mode etc. Belong within the scope of the optional embodiment of the application.

Although this application provides the method operating procedure as described in embodiment or flow chart, based on conventional or noninvasive The means for the property made can include more or less operating procedures.The step of being enumerated in embodiment order is only numerous steps A kind of mode in execution sequence, does not represent unique execution sequence.It, can when device or client production in practice performs With according to embodiment, either method shown in the drawings order is performed or parallel performed (such as at parallel processor or multithreading The environment of reason, even distributed data processing environment).Term " comprising ", "comprising" or its any other variant are intended to contain Lid non-exclusive inclusion, so that process, method, product or equipment including a series of elements not only will including those Element, but also including other elements that are not explicitly listed or further include as this process, method, product or equipment Intrinsic element.In the absence of more restrictions, be not precluded from the process including the element, method, product or Also there are other identical or equivalent elements in person's equipment.

Device that above-described embodiment illustrates or module etc. can specifically be realized or by computer chip or entity by having There is the product of certain function to realize.For convenience of description, it is divided into various modules during description apparatus above with function to retouch respectively It states.Certainly, the function of each module is realized can in the same or multiple software and or hardware when implementing the application, The module for realizing same function can be realized by the combination of multiple submodule etc..Device embodiment described above is only Schematically, for example, the division of the module, is only a kind of division of logic function, there can be other draw in actual implementation The mode of dividing, such as multiple module or components may be combined or can be integrated into another system or some features can be ignored, Or it does not perform.

It is also known in the art that in addition to realizing controller in a manner of pure computer readable program code, it is complete Entirely can by by method and step carry out programming in logic come controller with logic gate, switch, application-specific integrated circuit, may be programmed The form of logic controller and embedded microcontroller etc. realizes identical function.Therefore this controller is considered one kind Hardware component, and the structure that can also be considered as to the device for being used to implement various functions that its inside includes in hardware component.Or The device for being used to implement various functions even, can be considered as either the software module of implementation method can be hardware again by person Structure in component.

The application can be described in the general context of computer executable instructions, such as program Module.Usually, program module includes routines performing specific tasks or implementing specific abstract data types, program, object, group Part, data structure, class etc..The application can also be put into practice in a distributed computing environment, in these distributed computing environment, By performing task by communication network and connected remote processing devices.In a distributed computing environment, program module can To be located in the local and remote computer storage media including storage device.

As seen through the above description of the embodiments, those skilled in the art can be understood that the application can It is realized by the mode of software plus required general hardware platform.Based on such understanding, the technical solution essence of the application On the part that the prior art contributes can be embodied in the form of software product in other words, the computer software product It can be stored in storage medium, such as ROM/RAM, magnetic disc, CD, it is used including some instructions so that a computer equipment (can be personal computer, mobile terminal, server either network equipment etc.) perform each embodiment of the application or implementation Method described in some parts of example.

Each embodiment in this specification is described by the way of progressive, the same or similar portion between each embodiment Point just to refer each other, and the highlights of each of the examples are difference from other examples.The application can be used for crowd In mostly general or special purpose computing system environments or configuration.Such as：Personal computer, server computer, handheld device or Portable device, laptop device, multicomputer system, the system based on microprocessor, set top box, programmable electronics are set Standby, network PC, minicomputer, mainframe computer, distributed computing environment including any of the above system or equipment etc..

Although depicting the application by embodiment, it will be appreciated by the skilled addressee that the application there are many deformation and Variation is without departing from spirit herein, it is desirable to which appended embodiment includes these deformations and changes without departing from the application.

Claims

1. a kind of tumble detection method for human body, which is characterized in that including：

Obtain target image；

By target detection network, human testing is carried out to the target image, with determine the target image whether be comprising The image of human body；

In the case where the definite target image is comprising the image of human body, by convolutional neural networks, to the target figure As carrying out tumble identification, to determine whether the human body in the target image is in tumble state.

2. according to the method described in claim 1, it is characterized in that, the acquisition target image, including：

Gather the acoustic information in target area；

According to the acoustic information, target bearing is determined；

According to the target bearing, dollying head, to obtain the target image.

3. according to the method described in claim 1, it is characterized in that, the convolutional neural networks are established in the following way：

Human body image sample data is obtained, wherein, the human body image sample data includes multiple images for including body state；

According to the body state in the image of the pretreatment sample data, the image in the pretreatment sample data is divided Positive sample data and negative sample data, wherein, the image in the positive sample data includes at least one of：Include human body The image for the state that stand, the image for including the state that human body is seated, include the image for including the couchant state of human body The image of the state of human body at a slant；Image in the negative sample data includes at least one of：Include human body recumbency State image, include the image of the state of human body on all fours；

It is trained using the positive sample data, the negative sample data, to establish to identify the volume of body state type Product neutral net.

4. according to the method described in claim 3, it is characterized in that, during establishing the convolutional neural networks, the side Method further includes：

Obtain the image sample data not comprising human body；

Using the image sample data not comprising human body, error detection training is carried out to the convolutional neural networks.

5. a kind of human body falling detection device, which is characterized in that including：

Acquisition module, for obtaining target image；

Human detection module for passing through target detection network, human testing is carried out to the target image, to determine the mesh Whether logo image is the image comprising human body；

Tumble identification module, in the case where the definite target image is comprising the image of human body, passing through convolutional Neural Network carries out tumble identification to the target image, to determine whether the human body in the target image is in tumble state.

6. device according to claim 5, which is characterized in that the acquisition module includes：

Sound collector, for gathering the acoustic information in target area；

Mobile device and camera, wherein, the camera is arranged in the mobile device, and the mobile device is used for according to institute State target bearing, the mobile camera；The camera is used to obtain target image.

7. device according to claim 5, which is characterized in that described device further includes convolutional neural networks and establishes module, For establishing to identify the convolutional neural networks of body state type, wherein, the convolutional neural networks, which establish module, to be included：

Acquiring unit, for obtaining human body image sample data, wherein, the human body image sample data includes multiple comprising people The image of body state；

Extraction unit, for extracting satisfactory image from the human body image sample data as pretreatment sample number According to；

Division unit, for the body state in the image according to the pretreatment sample data, by the pretreatment sample number Image division positive sample data and negative sample data in, wherein, the image in the positive sample data include it is following at least One of：Include the image for the state that human body stands, include the image for the state that human body is seated, include the couchant shape of human body The image of state, the image for including the state of human body at a slant；Image in the negative sample data includes at least one of： Include the image of the state of human body recumbency, the image for including the state of human body on all fours；

Unit is established, for being trained using the positive sample data, the negative sample data, human body is identified to establish The convolutional neural networks of Status Type.

8. device according to claim 7, which is characterized in that the convolutional neural networks are established module and further included：

Error detection training unit, for obtaining the image sample data not comprising human body；And utilize the figure for not including human body Decent notebook data carries out error detection training to the convolutional neural networks.