CN115578792A - Method and system for early warning and detection of indoor personnel falls based on machine vision - Google Patents

Abstract

The invention relates to a machine vision-based indoor personnel falling early warning detection method and system, wherein the method comprises the following steps: the method comprises the steps of obtaining field images shot by machine vision, collecting the field images to form a field data set, inputting the field images into a deep learning network model based on a target detection algorithm, carrying out image preprocessing on the field images, obtaining human body images and posture characteristics of the human body images, storing the human body images and the field images in the field data set after the human body images are associated, carrying out posture characteristic analysis on the human body images through the deep learning network model and outputting a detection result, wherein the deep learning network model is trained by a training set and a verification set which are formed by selecting and collecting behavior and posture characteristics when the human body falls down. The invention can meet certain requirements on real-time performance and accuracy, and can detect the falling of the old people and send an alarm in time according to the behavior and posture characteristics when the old people fall on the premise of hardly hindering the daily life of the old people in practical application.

Description

Method and system for early warning and detection of indoor personnel falls based on machine vision

technical field

The invention relates to a machine vision-based indoor occupant fall warning detection method and system, and belongs to the field of image processing and recognition.

Background technique

In the traditional fall detection research field, the environment-based fall detection system has the defect that it is greatly affected by the external environment; the wearable fall detection system has the defect of interfering with the normal activities of the elderly to a certain extent; and the technology of the visual early warning fall detection system is relatively Advanced, its high target recognition rate and accuracy rate, less impact on the elderly, has become one of the popular research fields. However, in the field of fall warning detection research, compared with post-fall behavior detection, its research is more difficult, especially the use of machine vision methods to achieve fall behavior early warning detection, the target behavior classification and detection content is also more complicated.

Contents of the invention

The present invention provides a machine vision-based indoor occupant fall warning detection method and system, aiming at at least solving one of the technical problems existing in the prior art.

The technical solution of the present invention relates to a machine vision-based indoor personnel fall warning detection method, the method according to the present invention includes the following steps:

S10. Acquire on-site images captured by machine vision and collect them to form an on-site data set, and input the on-site images into a deep learning network model based on a target detection algorithm;

S20. Perform image preprocessing on the on-site image, acquire the human body image and its posture features, and store the human body image in the on-site data set after being associated with the on-site image;

S30. Analyzing the body features of the human body image through the deep learning network model and outputting a detection result;

Wherein, the deep learning network model is trained based on a training set and a verification set formed by selecting and collecting the characteristics of behavior and posture during a fall.

Further, wherein, the structure of the deep learning network model includes an input terminal, a reference network, a Neck network and an output terminal for outputting detection results; wherein, the step S20 includes:

performing image preprocessing on the input live image at the input terminal;

The reference network includes a Focus module and a CSPNet module, wherein the operation of the Focus module includes cutting the preprocessed live image and combining multiple feature information channels of the live image; the operation of the CSPNet module includes passing through the cross-stage The hierarchy merges the two parts of the base layer separated by feature maps;

The Neck network connects the reference network and the output terminal, and forms a complete feature pyramid network through FPN and PAN.

Further, wherein, in the step S20:

The image preprocessing operations include mosaic enhancement, adaptive anchor box calculation and adaptive image scaling.

Further, wherein, for the training of the deep learning network model, the formation of the training set and the verification set includes the following steps:

S40. Formulate a fall warning detection standard based on the behavior and posture characteristics when falling, classify and select images from the training database based on the fall warning detection standard and collect them to form a data set, mark the images of the data set and divide them into training in proportion Set and verification set; Wherein, described training database comprises UR Fall Detection Dataset database material, the video material of Le2i data set and described field data set.

Further, wherein, in the step S40:

The behavior and posture characteristics of the fall on the basis of the early warning detection standard include stretching out the arms when there is no supportable object around, showing the posture of slightly bending the elbows with the hands on the ground, crossing or overlapping steps, and changes in the angle of inclination of the human body. And detect the change of the aspect ratio of the rectangular box.

Further, wherein, in the step S40:

S41. Based on the training database, perform frame extraction processing on a given fall video, and save the training image of each frame;

S42. According to the fall warning detection standard, classify and select the training images according to the behavior before the fall and the result after the fall, and collect and form a data set according to the ratio of the number of the two images being 8:1;

S43. Label the images of the data set, and then divide the data set into a training set and a verification set at a ratio of 4:1.

Further, wherein, the training of the deep learning network model comprises the following steps:

S50. Representing a loss function by the error weighted sum of the prediction frame, confidence degree and class probability, and adjusting the parameters of the fall warning model according to the loss value calculated by the loss function;

Wherein, the loss value of the prediction frame is calculated as follows:

Among them, CIOU represents the loss function, and IOU represents the regression loss function; ρ is the straight-line distance between the center points of the two rectangular frames, c is the diagonal length of the smallest rectangle surrounding the two rectangular frames, and v is the aspect ratio similarity of the two rectangular frames , α is the impact factor of v.

Further, wherein, in the step S50:

The loss value for the confidence level is calculated as follows:

loss _BCE (z, x, y)=-L(z, x, y)×logp(z, x, y)-(1-L(z, x, y))

Among them, loss _BCE represents the confidence value of the classification loss function, the matrix L represents the confidence label, and the matrix P represents the prediction confidence. The value ranges of z, x, and y are as follows:

Then the confidence loss value of the overall network is calculated as follows:

loss _obj80 = a×l _obj +(1-a)×l _noobj

Among them, the value ranges of z, x, and y are as above; mask represents the mask value of each rectangular box.

The technical solution of the present invention also relates to a computer-readable storage medium on which program instructions are stored, and the above-mentioned method is implemented when the program instructions are executed by a processor.

The technical solution of the present invention also relates to an indoor occupant fall warning and detection system based on machine vision. The system includes a computer device, and the computer device includes the above-mentioned computer-readable storage medium.

The beneficial effects of the present invention are as follows:

The present invention is based on the machine vision-based indoor occupant fall early warning detection method and system. Starting from the direction of the visual early warning fall detection system, a machine vision-based indoor elderly fall early warning detection algorithm is designed and implemented to meet certain real-time and accuracy requirements. , so that in practical application, the algorithm can detect the fall of the elderly in time and issue an alarm under the premise of hardly hindering the daily life of the elderly, which is conducive to protecting the health of the elderly, and to a certain extent solves the problem of It solves the current lack of data sets related to fall warning detection; realizes the fall warning function in indoor environments, and provides a set of feasible solutions for the selection of target detection algorithms and the construction of related data sets; for future specific applications It provides a possibility for the daily indoor life monitoring and physical safety protection of the elderly.

Description of drawings

FIG. 1 is a basic flow chart of a machine vision-based indoor occupant fall warning detection method according to the present invention.

Fig. 2 is an algorithmic network structure diagram according to the method of the present invention.

Fig. 3 is a schematic block diagram of YOLOV5 black border filling according to an embodiment of the present invention.

Fig. 4 is a flow chart of creating a data set according to an embodiment of the present invention.

Fig. 5 is an image of a portion of a data set according to an embodiment of the present invention.

Fig. 6 is a partial pre-fall warning action image selected according to body characteristics according to an embodiment of the present invention.

Fig. 7 is a schematic diagram of an annotation interface and an XML data file according to an embodiment of the present invention.

Fig. 8 is a schematic diagram of a confusion matrix for model training according to an embodiment of the present invention.

FIG. 9 is a graph showing changes in evaluation indicators for model training according to an embodiment of the present invention.

Fig. 10 is a schematic diagram of some experimental results of the embodiment of the present invention.

detailed description

The idea, specific structure and technical effects of the present invention will be clearly and completely described below in conjunction with the embodiments and accompanying drawings, so as to fully understand the purpose, scheme and effect of the present invention.

It should be noted that, unless otherwise specified, when a feature is called "fixed" or "connected" to another feature, it can be directly fixed and connected to another feature, or indirectly fixed and connected to another feature. on a feature. As used herein, the singular forms "a", "the" and "the" are also intended to include the plural unless the context clearly dictates otherwise. Also, unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. The terms used in the specification herein are for describing specific embodiments only, and are not intended to limit the present invention. As used herein, the term "and/or" includes any combination of one or more of the associated listed items.

It should be understood that although the terms first, second, third etc. may be used in the present disclosure to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish elements of the same type from one another. For example, a first element could be termed a second element, and, similarly, a second element could be termed a first element, without departing from the scope of the present disclosure. The use of any and all examples, or exemplary language ("such as," "such as," etc.) provided herein is intended merely to better illuminate embodiments of the invention and does not pose a limitation on the scope of the invention unless otherwise claimed .

Referring to Figures 1 and 2, in some embodiments, the machine vision-based indoor occupant fall warning detection method according to the present invention at least includes the following steps:

S10. Obtain on-site images captured by machine vision and collect them to form an on-site data set, and input the on-site images into a deep learning network model based on a target detection algorithm;

S20. Perform image preprocessing on the on-site image, acquire the human body image and its body features, and store the human body image in the on-site data set after being associated with the on-site image;

S30, analyze the posture characteristics of the human body image through the deep learning network model and output the detection result;

Among them, the deep learning network model is trained based on the training set and verification set formed by selecting and collecting the characteristics of the behavior and posture during the fall.

The specific implementation manner of step S20

The embodiment of the present invention uses the YOLOV5 algorithm to detect the trend behavior of the human body before the fall, wherein the network structure of the YOLOV5 mainly includes four main modules, which are the input terminal, the backbone network, the Neck network and the Head output terminal.

Input the live image through the input terminal, and the size of the input live image is required to be 640*640. The on-site image is preprocessed, and the preprocessing operations include image flipping, mosaic enhancement, adaptive anchor frame calculation, and adaptive image scaling.

Among them, the mosaic data enhancement method is an image stitching technology, which uses random scaling, cropping and layout methods to stitch a group of images together, and integrates multiple photos into one photo for detection and learning, thus greatly improving the network quality. The training performance of the model reduces the memory requirements of the model and enriches the specific content of the data set.

In terms of adaptive anchor box calculation, YOLOV5 adds the calculation function to the program code, and automatically infers the most suitable anchor box according to the provided data set labels in each training. In terms of adaptive image scaling, the YOLOV5 algorithm adaptively adds the least black borders to the scaled image of the original image (see Figure 3), which improves the inference speed to a certain extent, reduces the amount of calculation and information redundancy, and makes target detection The speed will be improved to a certain extent.

After the on-site image is pre-processed, it enters the benchmark network (also known as BackBone network, backbone network) for image slicing operations and cross-level parallel network operations. The baseline network improves the level of feature information extraction by reducing the width and height information of the image and increasing the number of channels. Among them, modules such as Focus and CSPNet are combined through the YOLOV5 algorithm to form the main body of the network. Before entering the benchmark network, the Focus module cuts the preprocessed image at the input end, and combines the feature information such as the height and width of the image into the channel; the CSPNet module divides the feature map of the basic layer into two parts, and then uses the cross-stage layer The structure is combined, so that the learning ability of the CNN deep learning network model is enhanced, the learning efficiency is improved, and the calculation bottleneck and storage cost are reduced.

Furthermore, a Neck network is inserted between the reference network and the Head output terminal, and a complete feature pyramid network is formed by FPN and PAN. in. The Neck network serves as the connection hub between the trunk network and the top-level network, and is inserted between the backbone network and the final Head output as the neck of the network. Specifically, the Neck network of YOLOV5 is mainly composed of FPN and PAN. FPN uses a top-down approach to transfer information from high-level parts to the lower part and output related feature maps. It is mainly used to solve detection problems at multiple scales, thereby improving the performance of detecting tiny objects. PAN forms a complete feature pyramid network in a bottom-up manner. FPN and PAN combine and interact to convey strong semantic features through the top layer and strong positioning features through the bottom layer, so as to obtain more feature information.

Finally, the output end, also known as the Head network, is responsible for outputting the target detection results.

The specific implementation manner of step S40

Referring to Fig. 4 and Fig. 5, the data set in the embodiment of the present invention includes the result DetectionDataset database material after UR falls, the video material of Le2i data set and described scene image, to improve the quality and the quantity that data set contains image, improve model Detection effect. Frame extraction processing is performed on the fall video in the above database, and each frame of image is saved to a folder to form a preliminary data set.

After the preliminary data set is established, the fall warning detection standard is formulated according to the body characteristics to select the images of the above preliminary data set. According to different triggers and initial positions, fall behaviors are commonly classified as: slipping, tripping, fainting, slipping and falling, falling from standing to sitting, falling when getting up from a seat, and when supporting something fall and so on. Through the research on the images and videos of falling behavior, it is concluded that the fall warning and detection standards that conform to the behavior and posture characteristics of the fall include stretching out the arm when there is no object around, showing a posture of slightly bending the elbow with the hand on the ground, and walking in a straight line. Intersecting or overlapping shapes, changes in the angle of inclination of the human body, and changes in the aspect ratio of the detected rectangular frame (see Figure 6).

According to the fall warning detection standard, images are selected according to the two categories of pre-fall behavior (Warning) and post-fall result (fall). The image (fall) is the image of the result of the fall, and the number of images of the two categories is selected at a ratio of 1:8 to form the final data set.

After completing the selection and collection of the preliminary data set, label each image. Specifically, use Labelimg, an open source data labeling tool, and select the VOC label format, save it as an XML file, and label the interface diagram and XML data file ( See Figure 7).

Classify the marked images and label files, divide them into training set and verification set, where the training set: verification set is divided according to the ratio of 4:1, and generate corresponding TXT label files. The training set is used as the data sample for model fitting, and the model ability and the generalization function of the test model are initially evaluated through the verification set. The training error is optimized during the training process by minimizing the gradient.

The specific implementation manner of step S50

The YOLOV5 target recognition algorithm implemented in the present invention withdraws the evaluation index of the result through the CIOU loss function to measure the gap between the predicted information and the expected information of the deep learning network model. Its loss function can be divided into three parts: prediction frame error (lossrect), confidence error (lossobj) and category probability error (lossclc). Among them, the rectangular frame, that is, the prediction frame, is used to represent the size and precise position of the target. Confidence represents the credibility of the predicted rectangular frame, and the value ranges from 0 to 1. The larger the value, the more likely there is an expected detection target in the rectangular frame. The category probability characterizes the degree of prediction of the target category, so the loss function can be described as the weighted sum of the three errors of the prediction box, confidence degree and category probability, which is defined by the following formula.

loss＝A×loss _bj +b×loss _rect +c×loss _clc

The CIOU loss function evaluates the detection effect by comparing the predicted frame with the actual frame based on three factors: the aspect ratio of the rectangular frame, the centerline distance, and the minimum circumscribed rectangle. At the same time, the weighted Nms non-maximum value suppression method is added. This method uses the method of calculating the local maximum value, that is, looking for a way close to the expected output bounding box, so as to select the detection result with the highest probability and display its bounding box and category. . Among them, the CIOU loss function is used to calculate the prediction frame error, and the BCE_loss classification loss function is used to calculate the confidence error and category probability error.

The regression loss function of Complete-IOU (CIOU) can improve the stability and convergence of training. It mainly represents the overlapping part of the rectangular frame in IOU, and DIOU mainly represents the center point distance of the rectangular frame, the overlapping area of the two rectangular frames and the aspect ratio of the rectangular frame, and the three are added to the calculation formula in relative proportion. The regression loss function calculation formula is as follows:

In the above formula, ρ is the straight-line distance between the center points of the two rectangular frames, c is the diagonal length of the smallest rectangle enclosing the two rectangular frames, υ is the similarity of the aspect ratio of the two rectangular frames, and α is the influencing factor of υ. Since the value range of the arctan function is 0～π/2, the value range of υ is 0～1. When the aspect ratio of the two rectangular frames is equal, υ is 0. When the aspect ratio of the two rectangular frames is infinite , υ takes 1. When the distance between the two rectangular frames is infinite and the difference in aspect ratio is infinite, DIOU takes -1, υ takes 1, and α takes 0.5. At this time, CIOU takes -1.5; when the two rectangular boxes completely overlap, DIOU takes 1 , υ takes 0, α takes 0, then CIOU takes 1. Therefore, the value range of CIOU is -1.5~1.

Through the above formula, it can be obtained that when the IOU is larger, that is, the overlapping area of the two rectangular frames is larger, then the α is larger, and thus the influence of v is greater; otherwise, the smaller the IOU, that is, the smaller the overlapping area of the two rectangular frames, The smaller α is, the smaller the influence of v is. Therefore, the optimization direction of the rectangular frame can be determined according to the size of the overlapping area, so as to achieve the purpose of improving the detection progress.

For an image, the deep learning network model will divide it into a fixed number of grids with a fixed size before processing. For each grid, the deep learning network model will predict its adjacent three rectangular boxes, and Given the width, height, classification probability, confidence and other information of the rectangular frame, a corresponding confidence value will be output for each rectangular frame, and its label dimension should be consistent with the output dimension of the deep learning network model. The confidence label indicates the reliability of the prediction frame, and its value range corresponds to the CIOU loss function, which is 0 to 1 (when the CIOU value is negative, it will be truncated, that is, 0 will be used as the label). The larger the value of the confidence label, the closer the effect of the detection target is to the actual result, and the more in line with the expected effect of the experiment.

The BCE_loss classification loss function is based on the confidence label. Let the matrix L represent the confidence label, and the matrix P represent the prediction confidence. Then the calculation formula and value range of BCE loss for each value in the matrix are as follows:

loss _BCE (z, x, y)=-L(z, x, y)×logp(z, x, y)-(1-L(z, x, y))

Among them, lossBCE represents the confidence value of the classification loss function, the matrix L represents the confidence label, and the matrix P represents the prediction confidence. The value ranges of z, x, and y are as follows:

Then the confidence loss value of the overall network is calculated as follows:

loss _obj80 = a×l _obj +(1-a)×l _noobj

Among them, the value ranges of z, x, and y are as above; mask represents the mask value of each rectangular frame, and the value is used to judge whether there is a detection target in the rectangular frame. If it exists, it is represented by true, and if it does not exist, it is represented by false. The prediction box whose mask is true is used as a positive sample, and the prediction box whose mask is false is taken as a negative sample, and the total loss value is the weighted average of the two.

The data set that has been produced and classified is used for model training. The parameter settings during training include: the learning rate is 0.01, the learning rate momentum is 0.943; the weight file is YOLOV5s, and the network model parameters in the file are modified; the weight decay coefficient is 0.0005; The training batch is determined according to the size of the video memory. The batch used in this training is 2; the number of data set training epochs is 100. After the training is completed, the files named best.pt and last.pt will be obtained, which respectively represent the optimal training model file and the last training model file.

Through the confusion matrix obtained after training (see Figure 8), the recall rate is 1.00 and the precision rate is about 0.9524 for the pre-fall behavior class, and both the recall rate and the precision rate for the post-fall result are 0.95.

The change curve of evaluation indicators obtained during training (see Figure 9), from left to right, and from top to bottom are: training set position loss, training set confidence loss, training set classification loss, precision rate, recall rate, test Set position loss, test set confidence loss, test set classification loss, map (0.5 and 0.95) mean average precision, it can be seen that in the training process, with the increase of training times, the precision rate and recall rate and other indicators have gradually reached more ideal level.

Use the training model file, randomly select images and videos, and turn on the real-time camera for verification. Some screenshots of the detection results are shown in Figure 10. The fall warning system designed in this experiment can more accurately identify the various directions and postures of the indoor human body. Fall results, and early warning detection of possible fall risks based on the change trend of human body posture within a certain period of time before the fall, and assign a higher level of confidence to behaviors that clearly have a tendency to fall, so as to achieve the effect of fall warning.

It should be recognized that the method steps in the embodiments of the present invention may be implemented or implemented by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable memory. The methods can use standard programming techniques. Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system. However, the programs can be implemented in assembly or machine language, if desired. In any case, the language may be a compiled or interpreted language. Furthermore, the program can be run on an application specific integrated circuit programmed for this purpose.

In addition, operations of processes described herein may be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by context. The processes described herein (or variations and/or combinations thereof) can be performed under the control of one or more computer systems configured with executable instructions, and as code that collectively executes on one or more processors (e.g. , executable instructions, one or more computer programs or one or more applications), hardware or a combination thereof. The computer program comprises a plurality of instructions executable by one or more processors.

Further, the method can be implemented in any type of computing platform operably connected to a suitable one, including but not limited to personal computer, minicomputer, main frame, workstation, network or distributed computing environment, stand-alone or integrated computer platform, or communicate with charged particle tools or other imaging devices, etc. Aspects of the invention can be implemented as machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optically read and/or written storage medium, RSIM, ROM, etc., such that they can be read by a programmable computer, when the storage medium or device is read by the computer, can be used to configure and operate the computer to perform the processes described herein. Additionally, the machine-readable code, or portions thereof, may be transmitted over a wired or wireless network. The invention described herein includes these and other various types of non-transitory computer-readable storage media when such media include instructions or programs that implement the steps described above in conjunction with a microprocessor or other data processor. The invention may also include the computer itself when programmed according to the methods and techniques described herein.

Computer programs can be applied to input data to perform the functions described herein, thereby transforming the input data to generate output data stored to non-volatile memory. Output information may also be applied to one or more output devices such as a display. In a preferred embodiment of the invention, the transformed data represents physical and tangible objects, including specific visual depictions of physical and tangible objects produced on a display.

The above is only a preferred embodiment of the present invention, and the present invention is not limited to the above-mentioned implementation, as long as it achieves the technical effect of the present invention by the same means, within the spirit and principles of the present invention, any Any modification, equivalent replacement, improvement, etc., shall be included within the protection scope of the present invention. Various modifications and changes may be made to its technical solutions and/or implementations within the protection scope of the present invention.

Claims (10)

1. An indoor person falling early warning detection method based on machine vision is characterized by comprising the following steps:

s10, acquiring field images shot by mechanical vision, converging the field images to form a field data set, and inputting the field images into a deep learning network model based on a target detection algorithm;

s20, carrying out image preprocessing on the field image, acquiring a human body image and the posture characteristics thereof, associating the human body image with the field image and storing the human body image and the field image in the field data set;

s30, performing posture characteristic analysis on the human body image through the deep learning network model and outputting a detection result;

wherein the deep learning network model is trained by a training set and a verification set which are selected and collected based on behavior and posture characteristics when falling.

2. The method of claim 1, wherein the structure of the deep learning network model comprises an input, a reference network, a hack network, and an output for outputting a detection result; wherein the step S20 includes:

performing image preprocessing on the field image input at the input end;

the reference network comprises a Focus module and a CSPNet module, wherein the Focus module is operated by cutting the preprocessed live image and combining a plurality of characteristic information of the live image into a channel; the CSPNet module operates to merge two portions of the base layer into which the feature map is divided by a cross-phase hierarchy;

the Neck network is connected with the reference network and the output end, and forms a complete characteristic pyramid network through the FPN and the PAN.

3. The method according to claim 2, wherein in step S20:

the image pre-processing operations include mosaic enhancement, adaptive anchor frame calculation, and adaptive image scaling.

4. The method of claim 1, wherein for training of the deep-learning network model, the forming of the training set and the validation set comprises the steps of:

s40, making a falling early warning detection standard based on behavior and posture characteristics during falling, classifying and selecting images from a training database based on the falling early warning detection standard, collecting the images to form a data set, labeling the images of the data set, and dividing the images into a training set and a verification set according to a proportion; wherein the training database comprises UR FallDetection Dataset database data, video data of Le2i Dataset and the field Dataset.

5. The method of claim 4, wherein in step S40:

the behavior and posture characteristics of the early warning detection standard when falling down comprise the posture that an arm is stretched out when no supportable object exists in the periphery, the slightly bent elbow is supported by a hand, the step is in a cross shape or an overlapping shape, the change condition of the inclination angle of the human body and the change condition of the aspect ratio of the detection rectangular frame.

6. The method of claim 5, wherein in step S40:

s41, based on a training database, performing frame extraction processing on a given falling video, and storing a training image of each frame;

s42, according to the falling early warning detection standard, classifying and selecting the training images according to the results before falling and after falling, and collecting and forming a data set according to the ratio that the number of the two images is 8;

s43, labeling the images of the data set, and dividing the data set into a training set and a verification set according to the proportion of 4.

7. The method of claim 1, wherein the training of the deep learning network model comprises the steps of:

s50, representing a loss function through a prediction frame, a confidence coefficient and an error weighted sum of class probabilities, and adjusting parameters of the fall early warning model according to a loss value obtained through calculation of the loss function;

wherein the loss value of the prediction box is calculated as follows:

wherein, CIOU represents a loss function, IOU represents a regression loss function; rho is the straight-line distance between the center points of the two rectangular frames, c is the length of the diagonal line of the rectangle which minimally surrounds the two rectangular frames, upsilon is the similarity of the width-height ratio of the two rectangular frames, and alpha is the influence factor of upsilon.

8. The method of claim 7, wherein in step S50:

the loss value of confidence is calculated as follows:

loss _BCE (z，x，y)＝-L(z，x，y)×logp(z，x，y)-(1-L(z，x，y))

therein, loss _BCE The confidence value of the classification loss function is represented, the matrix L represents a confidence label, the matrix P represents the prediction confidence, and the value ranges of z, x and y are as follows:

the confidence loss value for the overall network is calculated as follows:

loss _obj80 ＝a×l _obj +(1-a)×l _noobj

wherein the mask represents a mask value for each rectangular box.

9. A computer readable storage medium having stored thereon program instructions which, when executed by a processor, implement the method of any one of claims 1 to 8.

10. The utility model provides an indoor personnel early warning detecting system that tumbles based on machine vision which characterized in that includes:

computer arrangement comprising a computer readable storage medium according to claim 9.

Images