A System and Method for Early Childhood Behavior Analysis Based on Image Recognition

Abstract

This invention relates to the field of image recognition technology, and in particular to a system and method for analyzing preschool children's behavior based on image recognition. The system includes a data acquisition module for acquiring images of preschool children and ambient light intensity; an edge blurring calculation module; a preschool child tracking module; and a preschool child behavior recognition module. This invention accurately identifies and analyzes preschool children's images to precisely identify target preschool children in multi-preschool environments and performs precise behavioral analysis on these children to alert supervisors. Specifically, the preschool child tracking module accurately identifies images of target preschool children in multi-preschool environments, thereby enabling precise monitoring of the target children.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a system and method for analyzing children's behavior based on image recognition. Background Technology

[0002] Young children have limited cognitive and coping abilities regarding their environment, making them vulnerable to various potential dangers, such as rolling, crawling, and touching dangerous objects. The presence of childcare can promptly identify and prevent these potential dangers, ensuring the safety and health of young children. Because childcare plays a crucial role in a child's life, it is essential to categorize and understand children's behaviors.

[0003] Chinese Patent Publication No. CN117373110A discloses a method, device, and equipment for infant behavior recognition using visible light-thermal infrared imaging. The method includes: acquiring real-time thermal infrared and visible light images in an infant care scenario; performing target detection on the visible light image, outputting the detection box position information and category information of a preset target related to infant care; acquiring a preset thermal infrared imaging projection area, and when the detection box of the preset target appears in the thermal infrared projection area, outputting thermal imaging morphological feature information and thermal imaging temperature change feature information according to preset rules; and classifying infant behavior and outputting the classification result. This invention achieves real-time, accurate, and non-contact infant behavior recognition, providing important support and assistance for infant care and safety. However, in multi-child environments, this invention struggles to accurately identify the target child in the image and analyze the target child's behavior, and is easily influenced by other children, leading to inaccurate behavior analysis. Summary of the Invention

[0004] Therefore, the present invention provides a system and method for analyzing children's behavior based on image recognition, in order to overcome the problem in the prior art that the analysis of children's behavior is inaccurate due to the influence of other children in a multi-child environment.

[0005] To achieve the above objectives, in one aspect, the present invention provides a system for analyzing preschool children's behavior based on image recognition, the system comprising:

[0006] The data acquisition module is used to collect images of children and the intensity of ambient light.

[0007] The edge blurring calculation module is used to blur the image of the child, obtain the gray value of the blurred image of the child and the current frame, output the binary difference between adjacent frames, judge the proportion of moving pixels, synthesize the spatiotemporal segment image, adjust the judgment process of the proportion of moving pixels, and revise the judgment process of the proportion of moving pixels.

[0008] The preschool tracking module is used to obtain labeled spatiotemporal segment images based on spatiotemporal segment images, and output the matching status of preschool children based on the labeled spatiotemporal segment images. It also performs a first fine-tuning on the output process of the preschool children matching status, and a second fine-tuning on the fine-tuning process of the first fine-tuning and the process of judging the proportion of motion pixels.

[0009] The preschool behavior recognition module is used to judge the degree of behavioral risk based on the matching of labeled spatiotemporal fragment images and preschool children, and to alert the supervisor based on the degree of behavioral risk. It is also used to modify the degree of behavioral risk based on the dangerous situation matched with the identity.

[0010] Furthermore, the edge blurring calculation module inputs the child's image into the face recognition model to obtain the child's image with face annotations output by the face recognition model. Then, it uses Gaussian blurring to blur the child's image with face annotations, resulting in a blurred child image. Finally, it calculates the edge blurring based on the red channel intensity R of the blurred child image. i Green channel intensity G i Blue channel intensity B i The first weighting coefficient α1, the second weighting coefficient α2, and the third weighting coefficient α3 affect the grayscale value Yh of the current frame. i Calculations are performed to obtain the grayscale value Yh of the current frame. i Set Yh i =α1×R i +α2×G i +α3×B i .

[0011] Furthermore, the edge blurring calculation module calculates the edge blurring based on the current frame grayscale value Yh. i And the grayscale value Yh of the previous frame i-1 The absolute difference Dc between adjacent frame pixels is calculated, and Dc = |Yh i -Yh i-1 The absolute difference Dc between adjacent frames is compared with the difference threshold Ty. Based on the comparison result, the binary difference Bt between adjacent frames is output, where:

[0012] When Dc≥Ty, the edge blurring calculation module outputs 1 as the binary difference Bt between adjacent frames;

[0013] When Dc < Ty, the edge blurring calculation module outputs 0 as the binary difference Bt between adjacent frames.

[0014] Furthermore, the edge blurring calculation module calculates the motion pixel ratio Rt based on the number of pixels Bt1 with a binary difference of 1 between adjacent frames and the frame resolution Wz×Hz, setting Rt=Bt1 / Wz×Hz×100%. The motion pixel ratio Rt is compared with a preset motion pixel ratio Rt0, and a judgment is made based on the comparison result. The frame attributes of the current frame are then output based on the judgment result, wherein:

[0015] When Rt≤Rt0, the edge blurring calculation module determines that the proportion of moving pixels is small, and outputs the non-key frame as the frame attribute of the current frame.

[0016] When Rt > Rt0, the edge blurring calculation module determines that the proportion of moving pixels is large, outputs the keyframe as the frame attribute of the current frame, and combines all the images whose frame attribute of the current frame is a keyframe into a spatiotemporal segment image.

[0017] Furthermore, the edge blurring calculation module calculates the average pixel difference Dp based on the frame resolution Wz×Hz and the sum of the absolute differences of pixels in all adjacent frames, Dch, and sets... The average pixel difference Dp is compared with the preset average pixel difference Dp0. Based on the comparison result, the average pixel difference is judged, and the judgment process for the proportion of moving pixels is adjusted according to the judgment result.

[0018] When Dp≤Dp0, the edge blur calculation module determines that the average pixel difference is small and does not adjust the judgment process for the proportion of moving pixels.

[0019] When Dp > Dp0, the edge blurring calculation module determines that the average pixel difference is large, and adjusts the judgment process for the moving pixel ratio. It adjusts the preset moving pixel ratio Rt0 using a deviation coefficient pcx, setting pcx = 1.57 - 0.23 × e -0.38×（Dp-Dp0） e is the base of the natural logarithm. The adjusted preset motion pixel ratio Rt01 is obtained. Rt01 is set to Rt0×pcx. The preset motion pixel ratio Rt0 is replaced with the adjusted preset motion pixel ratio Rt01. The motion pixel ratio Rt is then compared with the preset motion pixel ratio Rt0 again.

[0020] Furthermore, the edge blurring calculation module compares the ambient light intensity Gq with the preset ambient light intensity Gq0, judges the suitability of the light intensity based on the comparison result, and revises the judgment process of the moving pixel ratio based on the judgment result, wherein:

[0021] When Gq≤Gq0, the edge blurring calculation module determines that the illumination intensity is suitable and does not revise the judgment process for the proportion of moving pixels.

[0022] When Gq > Gq0, the edge blurring calculation module determines that the illumination intensity is unsuitable and revises the judgment process for the moving pixel ratio. It revises the preset moving pixel ratio Rt0 using the light intensity coefficient qd, setting qd = 1.38 - 0.21 × e -0.7×（Gq-Gq0） The revised preset motion pixel ratio Rt02 is obtained. Rt02 is set to Rt0 × qd. The preset motion pixel ratio Rt0 is replaced with the revised preset motion pixel ratio Rt02. The motion pixel ratio Rt is then compared with the preset motion pixel ratio Rt0 again.

[0023] Furthermore, when the preschool tracking module obtains the labeled spatiotemporal segment image based on the spatiotemporal segment image, it inputs the spatiotemporal segment image into the joint tracking model to obtain the labeled spatiotemporal segment image output by the joint tracking model. The labeled spatiotemporal segment image includes the preschool identity matching probability Sp of the previous frame, the number of occluded joints J, the set of coordinates of each joint Gj=[g1,g2,g3,...,gn], interactive behavior, and spatial relationship.

[0024] Furthermore, when the child tracking module outputs the child matching results based on the labeled spatiotemporal segment images, it compares the child identity matching probability Sp of the previous frame with the preset child identity matching probability Sp0, judges the child matching results of the previous frame based on the comparison results, and outputs the child matching results based on the judgment results, wherein:

[0025] When Sp < Sp0, the child tracking module determines that the child matching situation in the previous frame is not a match, and outputs the different children as the child matching situation;

[0026] When Sp≥Sp0, the child tracking module determines that the child matching situation in the previous frame is a match, and outputs the same child as the child matching situation.

[0027] Furthermore, when the child tracking module performs the first fine-tuning of the output process for child matching, it compares the child's behavioral complexity Fd with the preset child behavioral complexity Fd0, judges the child's behavioral complexity based on the comparison result, and performs the first fine-tuning of the output process for child matching based on the judgment result, wherein:

[0028] When Fd≤Fd0, the child tracking module determines that the complexity of the child's behavior is low and does not perform the first fine-tuning on the output process of the child matching situation;

[0029] When Fd > Fd0, the child tracking module determines that the child's behavior complexity is high, and performs a first fine-tuning on the output process of the child matching situation. This fine-tuning is done by using a complexity coefficient γ to adjust the preset child identity matching probability Sp0, where γ = 1.41 - 0.21 × e -1.23×（Fd-Fd0） After obtaining the refined preset child identity matching probability Sp0`, Sp0` is set to Sp0×γ. The preset child identity matching probability Sp0 is replaced with the refined preset child identity matching probability Sp0`, and the child identity matching probability Sp of the previous frame is compared with the preset child identity matching probability Sp0 again.

[0030] On the other hand, the present invention also provides a method for analyzing preschool children's behavior based on image recognition, the method comprising:

[0031] Step S1: Collect images of the child and the ambient light intensity;

[0032] Step S2: Blur the image of the child to obtain the grayscale value of the blurred image of the child and the current frame, output the binary difference between adjacent frames, judge the proportion of moving pixels, synthesize the spatiotemporal segment image, adjust the judgment process of the proportion of moving pixels, and revise the judgment process of the proportion of moving pixels.

[0033] Step S3: Obtain labeled spatiotemporal segment images based on spatiotemporal segment images, output the matching status of children based on labeled spatiotemporal segment images, perform a first fine-tuning on the output process of children's matching status, and perform a second fine-tuning on the fine-tuning process of the first fine-tuning and the judgment process of motion pixel ratio.

[0034] Step S4: The degree of behavioral risk is judged based on the labeled spatiotemporal segment images and the matching of children, and an alarm is alerted to the supervisor based on the degree of behavioral risk. The degree of behavioral risk is also modified based on the dangerous situation of identity matching.

[0035] Compared with existing technologies, the beneficial effects of this invention are as follows: The system collects images of children and ambient light intensity through a data acquisition module for subsequent analysis of children's identity and behavior. The system also uses an edge blurring calculation module to blur the children's images, reducing the computational burden and allowing computing power to be concentrated on children's identity recognition and behavior analysis, thereby improving the accuracy of child identification and the precision of behavior analysis. Furthermore, the system uses a child tracking module to accurately identify target children in multi-child environments for subsequent monitoring. The child tracking module also assesses the complexity of children's behavior and, based on the assessment results, [further actions are taken]. The output process of matching children's identities undergoes a first refinement to improve the accuracy of judging the matching of children in the previous frame. When the complexity of a child's behavior is high, excessively complex behavior may make the child unclear in the image, leading to an increase in the probability of matching the identity of the child in the previous frame output by the model. In this case, the impact of high complexity of a child's behavior on the accuracy of judging the matching of children in the previous frame is reduced by increasing the preset value of the child's identity matching probability, thereby improving the accuracy of judging the matching of children in the previous frame. The system also uses a child behavior recognition module to perform risk assessment on the child's behavior and alerts the supervisor based on the assessment results to achieve child monitoring and management. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the system structure of the image recognition-based preschool behavior analysis system in this embodiment;

[0037] Figure 2 This is a flowchart illustrating the method for analyzing preschool behavior based on image recognition in this embodiment. Detailed Implementation

[0038] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0039] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0040] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0041] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0042] Please see Figure 1 As shown, this is a schematic diagram of the system structure of the image recognition-based early childhood behavior analysis system in this embodiment. The system includes:

[0043] The data acquisition module is used to collect images of children and the intensity of ambient light.

[0044] The edge blurring calculation module is used to blur the image of the child, obtain the grayscale value of the blurred image of the child and the current frame, output the binary difference between adjacent frames, judge the proportion of moving pixels, synthesize the spatiotemporal segment image, adjust the judgment process of the proportion of moving pixels, and revise the judgment process of the proportion of moving pixels. The edge blurring calculation module is connected to the data acquisition module.

[0045] The child tracking module is used to obtain labeled spatiotemporal segment images based on spatiotemporal segment images, and output the child matching status based on the labeled spatiotemporal segment images. It also performs a first fine-tuning on the output process of the child matching status, and a second fine-tuning on the fine-tuning process of the first fine-tuning and the process of judging the proportion of moving pixels. The child tracking module is connected to the edge blur calculation module.

[0046] The preschool behavior recognition module is used to judge the degree of behavioral risk based on the matching of labeled spatiotemporal fragment images and preschool children, and to alert the supervisor based on the degree of behavioral risk. It is also used to modify the degree of behavioral risk based on the dangerous situation matched with the identity. The preschool behavior recognition module is connected to the preschool child tracking module.

[0047] Specifically, the image recognition-based early childhood behavior analysis system of the present invention is applied to early childhood supervision terminals, such as early childhood abnormal behavior monitoring terminals, kindergarten safety monitoring cameras, and amusement park safety monitoring equipment. Through precise recognition and analysis of early childhood images, it accurately identifies target children in multi-child environments and performs precise behavior analysis on the target children to alert supervisors. The system collects early childhood images and ambient light intensity through a data acquisition module for subsequent analysis of children's identity and behavior. The system also uses an edge blurring calculation module to blur early childhood images, reducing the computational burden on image processing and allowing computational power to be concentrated on early childhood identity recognition and behavior analysis, thereby improving the accuracy of early childhood identity recognition and the precision of behavior analysis. The system also uses a children tracking module to accurately identify children in multi-child environments. The system identifies target children for subsequent monitoring. A child tracking module assesses the complexity of the children's behavior and performs initial refinement on the matching output based on the assessment results. This improves the accuracy of matching the previous frame's child. When the child's behavior complexity is high, the child may be less clear in the image, increasing the probability of matching the previous frame's child. In this case, increasing the preset child identity matching probability reduces the impact of high behavioral complexity on the accuracy of matching the previous frame, thus improving the accuracy of matching the previous frame's child. The system also uses a child behavior recognition module to assess the risks of children's behavior and alerts supervisors based on the assessment results, enabling effective child monitoring and management.

[0048] Specifically, the data acquisition module collects images of children through multi-angle cameras. The multi-angle cameras refer to camera devices that can acquire image or video information from multiple different angles, such as panoramic cameras installed at the corners of walls in children's activity areas. The data acquisition module collects ambient light intensity through light intensity sensors.

[0049] Specifically, the data acquisition module collects images of children and the intensity of ambient light to enable precise monitoring and behavioral analysis of the children in the future.

[0050] Specifically, the edge blurring calculation module inputs the child's image into the face recognition model to obtain a child's image with face annotations output by the face recognition model. Then, it uses Gaussian blurring to blur the child's image with face annotations, resulting in a blurred child image. Finally, it calculates the blurred child image based on the intensity R of the red channel of the blurred child image. i Green channel intensity G i Blue channel intensity B i The first weighting coefficient α1, the second weighting coefficient α2, and the third weighting coefficient α3 affect the grayscale value Yh of the current frame.i Calculations are performed to obtain the grayscale value Yh of the current frame. i Set Yh i =α1×R i +α2×G i +α3×B i .

[0051] Specifically, the face recognition model refers to the YOLOx5 model, which takes images of young children as input and outputs images of young children with labeled faces. The YOLOx5 model is a single-stage object detection model based on the YOLO series. YOLO stands for You Only Look Once. The edge blurring calculation module constructs the face recognition model using a face recognition model construction method. This embodiment does not limit the specific method of face recognition model construction; those skilled in the art can set it according to actual needs. For example, the YOLOx5 model can be trained, validated, and tested by dividing 70% of the historical recognition data into a recognition training set, 20% into a recognition validation set, and 10% into a recognition test set. The historical recognition data refers to the dataset used to construct the YOLOx5 model, including historically acquired images of young children and historically acquired images of young children. The image refers to a child's image with labeled facial features. This labeled image is obtained by annotating the facial regions of the child's image. This embodiment does not limit the specific method of annotating the facial regions; those skilled in the art can set it according to actual needs, such as circling the facial regions in the child's image with black lines and labeling them with the letter F. The recognition training set refers to the dataset used to train the YOLOx5 model. The recognition validation set refers to the dataset used to validate the results of the YOLOx5 model. The recognition test set refers to the dataset used to test the positive and negative results of the YOLOx5 model. The accuracy test dataset uses Gaussian blurring, an image processing technique primarily aimed at reducing image detail and noise to create a smooth, soft blur effect. This embodiment does not limit the specific processing method for blurring the face-annotated children's image using Gaussian blurring. Those skilled in the art can set it according to actual needs, such as establishing a Gaussian kernel conforming to a Gaussian distribution, sequentially covering each pixel in the facial region of the face-annotated children's image, calculating the weighted average of the pixel and its surrounding pixels using the Gaussian kernel, and replacing the original pixel value with this average to obtain the blurred children's image. The red channel intensity refers to the value of the red primary color in the facial area of ​​the blurred child's image. The green channel intensity refers to the value of the green primary color in the facial area of ​​the blurred child's image. The blue channel intensity refers to the value of the blue primary color in the facial area of ​​the blurred child's image. The red, blue, and green primary colors refer to the red, blue, and green primary colors of light, which are the three primary colors of light: red, green, and blue. This embodiment does not limit the specific method of obtaining the red, green, and blue channel intensities of the blurred child's image; those skilled in the art can set them according to actual needs.The intensity of the red, green, and blue channels in the blurred image of the child can be directly viewed using image processing software. The first weighting coefficient α1 refers to the weighting coefficient corresponding to the red channel intensity, α1 = 0.299; the second weighting coefficient α2 refers to the weighting coefficient corresponding to the green channel intensity, α2 = 0.587; and the third weighting coefficient α3 refers to the weighting coefficient corresponding to the blue channel intensity, α3 = 0.114. Since the human eye perceives the brightness contribution ratio of the RGB three channels as approximately red: 29.9%, green: 58.7%, and blue: 11.4%, α1 = 0.299, α2 = 0.587, and α3 = 0.114 are set. i represents the frame order of the blurred image of the child, and i is a positive integer.

[0052] Specifically, the edge blurring calculation module blurs the images of young children to achieve the effect of protecting privacy while preserving body movements in subsequent analysis.

[0053] Specifically, the edge blur calculation module calculates the edge blur based on the current frame grayscale value Yh. i And the grayscale value Yh of the previous frame i-1 The absolute difference Dc between adjacent frame pixels is calculated, and Dc = |Yh i -Yh i-1 The absolute difference between adjacent frames (Dc) is compared with a difference threshold (Ty), where 15 ≤ Ty ≤ 30. Based on the comparison result, the binary difference between adjacent frames (Bt) is output, where:

[0054] When Dc≥Ty, the edge blurring calculation module outputs 1 as the binary difference Bt between adjacent frames;

[0055] When Dc < Ty, the edge blurring calculation module outputs 0 as the binary difference Bt between adjacent frames.

[0056] Specifically, the grayscale value of the previous frame refers to the grayscale value of the image of the previous frame based on the current frame, and the difference threshold refers to the preset value for judging the binary difference Bt between adjacent frames. This embodiment does not limit the specific value of the difference threshold. For example, the specific value of the difference threshold can be set according to the frame picture. When the frame picture is a children's activity area, it is set to 10≤Ty≤15. The binary difference between adjacent frames refers to converting the absolute difference of pixels between adjacent frames into binary values.

[0057] Specifically, the edge blurring calculation module compares the absolute difference between pixels in adjacent frames with a difference threshold and converts the absolute difference between pixels in adjacent frames into binary values ​​for subsequent analysis and statistics of children's images.

[0058] Specifically, the edge blurring calculation module calculates the motion pixel ratio Rt based on the number of pixels Bt1 with a binary difference of 1 between adjacent frames and the frame resolution Wz×Hz, setting Rt=Bt1 / Wz×Hz×100%. The motion pixel ratio Rt is compared with a preset motion pixel ratio Rt0, and a judgment is made based on the comparison result. The frame attributes of the current frame are then output based on the judgment result.

[0059] When Rt≤Rt0, the edge blurring calculation module determines that the proportion of moving pixels is small, and outputs the non-key frame as the frame attribute of the current frame.

[0060] When Rt > Rt0, the edge blurring calculation module determines that the proportion of moving pixels is large, outputs the keyframe as the frame attribute of the current frame, and combines the images of all the current frame's frame attributes as keyframes into a spatiotemporal segment image.

[0061] Specifically, the number of pixels with a binary difference of 1 between adjacent frames refers to the total number of pixels with a binary difference of 1 between adjacent frames. The frame resolution refers to the resolution of the current frame. Wz refers to the height of the current frame image. Hz refers to the width of the current frame image. This embodiment does not limit the method of obtaining the height and width of the current frame image. For example, the height and width of the current frame image can be obtained through image processing software. The preset motion pixel ratio refers to a preset value used to judge the motion pixel ratio. This embodiment does not limit the specific value of the preset motion pixel ratio. Those skilled in the art can set it according to actual needs. For example, the specific value of the preset motion pixel ratio can be set according to background noise, such as 5%≤Rt0≤10%. The motion pixel ratio situation refers to the magnitude of the motion pixel ratio judged based on the motion pixel ratio and the preset motion pixel ratio. The motion pixel ratio situation includes a large motion pixel ratio and a small motion pixel ratio. This embodiment does not limit the specific synthesis method of synthesizing all images with keyframe attributes of the current frame into a spatiotemporal segment image. For example, all images with keyframe attributes of the current frame can be arranged in chronological order to synthesize a spatiotemporal segment image.

[0062] Specifically, the edge blurring calculation module eliminates non-key frame images by identifying key frames, thereby obtaining key images, reducing interference from non-key frame images when judging children's behavior, and thus reducing the amount of processing.

[0063] Specifically, the edge blur calculation module calculates the average pixel difference Dp based on the frame resolution Wz×Hz and the sum of the absolute differences of pixels in all adjacent frames, Dch, and sets... The average pixel difference Dp is compared with the preset average pixel difference Dp0. Based on the comparison result, the average pixel difference is judged, and the judgment process for the proportion of moving pixels is adjusted according to the judgment result.

[0064] When Dp≤Dp0, the edge blur calculation module determines that the average pixel difference is small and does not adjust the judgment process for the proportion of moving pixels.

[0065] When Dp > Dp0, the edge blurring calculation module determines that the average pixel difference is large, and adjusts the judgment process for the moving pixel ratio. It adjusts the preset moving pixel ratio Rt0 using a deviation coefficient pcx, setting pcx = 1.57 - 0.23 × e -0.38×（Dp-Dp0） e is the base of the natural logarithm. The adjusted preset motion pixel ratio Rt01 is obtained. Rt01 is set to Rt0×pcx. The preset motion pixel ratio Rt0 is replaced with the adjusted preset motion pixel ratio Rt01. The motion pixel ratio Rt is then compared with the preset motion pixel ratio Rt0 again.

[0066] Specifically, the sum of absolute differences of pixels in all adjacent frames refers to the sum obtained by adding the absolute differences of pixels in adjacent frames. The preset average pixel difference refers to a preset value for judging the average pixel difference. This embodiment does not limit the specific value of the preset average pixel difference. Those skilled in the art can set it according to actual conditions, such as setting the specific value of the preset average pixel difference based on expert experience. The average pixel difference situation refers to the magnitude of the average pixel difference judged based on the average pixel difference and the preset average pixel difference. The average pixel difference situation includes small average pixel difference and large average pixel difference.

[0067] Specifically, the edge blur calculation module judges the average pixel difference and adjusts the judgment process of the motion pixel ratio based on the judgment result. When the average pixel difference is large, the scene of the child's image shooting changes from static to dynamic. By setting a deviation coefficient that increases with the increase of the average pixel difference and multiplying it with the preset motion pixel ratio, the value of the preset motion pixel ratio is increased, thereby improving the accuracy of the edge blur calculation module.

[0068] Specifically, the edge blurring calculation module compares the ambient light intensity Gq with the preset ambient light intensity Gq0, determines the suitability of the light intensity based on the comparison result, and revises the judgment process for the proportion of moving pixels based on the judgment result, wherein:

[0069] When Gq≤Gq0, the edge blurring calculation module determines that the illumination intensity is suitable and does not revise the judgment process for the proportion of moving pixels.

[0070] When Gq > Gq0, the edge blurring calculation module determines that the illumination intensity is unsuitable and revises the judgment process for the moving pixel ratio. It revises the preset moving pixel ratio Rt0 using the light intensity coefficient qd, setting qd = 1.38 - 0.21 × e -0.7×（Gq-Gq0） The revised preset motion pixel ratio Rt02 is obtained. Rt02 is set to Rt0 × qd. The preset motion pixel ratio Rt0 is replaced with the revised preset motion pixel ratio Rt02. The motion pixel ratio Rt is then compared with the preset motion pixel ratio Rt0 again.

[0071] Specifically, the ambient light intensity refers to the light intensity in the children's activity space, and the preset ambient light intensity refers to a preset value for judging whether the light intensity is suitable. This embodiment does not limit the specific value of the preset ambient light intensity. For example, the specific value of the preset ambient light intensity can be set according to the surrounding environment. The light intensity suitability refers to whether the light intensity is suitable based on the ambient light intensity and the preset ambient light intensity. The light intensity suitability includes unsuitable and suitable.

[0072] Specifically, the edge blurring calculation module determines whether the lighting intensity is suitable and revises the judgment process of the motion pixel ratio based on the judgment result. When the lighting intensity is not suitable, the excessive light will cause the child's image to be overexposed, resulting in excessively high pixel values ​​and causing non-key frames to be misjudged as key frames. By setting a light intensity coefficient that increases with the increase of ambient light intensity and multiplying it by the preset motion pixel ratio, the value of the preset motion pixel ratio is increased, thereby reducing the probability of non-key frames being misjudged as key frames and improving the accuracy of the edge blurring calculation module.

[0073] Specifically, when the preschool tracking module obtains labeled spatiotemporal segment images based on spatiotemporal segment images, it inputs the spatiotemporal segment images into the joint tracking model to obtain labeled spatiotemporal segment images output by the joint tracking model. The labeled spatiotemporal segment images include the preschooler identity matching probability Sp of the previous frame, the number of occluded joints J, the set of coordinates of each joint Gj=[g1,g2,g3,...,gn], interactive behavior, and spatial relationships.

[0074] Specifically, the keypoint tracking model refers to a convolutional neural network model that takes spatiotemporal segment images as input and labeled spatiotemporal segment images as output. The child tracking module constructs the keypoint tracking model using a keypoint tracking model construction method. This embodiment does not limit the specific method of keypoint tracking model construction; those skilled in the art can set it according to actual needs. For example, the convolutional neural network model can be trained using a tracking training dataset to obtain the keypoint tracking model. The tracking training dataset refers to the training dataset used to construct the keypoint tracking model. The tracking training dataset includes historically acquired spatiotemporal segment images and labeled spatiotemporal segment images corresponding to the historically acquired spatiotemporal segment images. The previous frame child identity matching probability refers to the probability that the child in the current frame is the same child as the child in the previous frame. The occluded joint number refers to the labeled spatiotemporal segment. The number of joints obscured in the image; the set of joint coordinates refers to the set of coordinates of each joint marked in the labeled spatiotemporal segment image; the coordinates of each joint point include the coordinates of the center of the face, the coordinates of the inner side of the left eye socket, the coordinates of the inner side of the right eye socket, the coordinates of the upper part of the left earlobe, the coordinates of the upper part of the right earlobe, the coordinates of the acromion of the left scapula, the coordinates of the acromion of the right scapula, the coordinates of the left elbow, the coordinates of the right elbow, the coordinates of the left wrist, the coordinates of the right wrist, the coordinates of the left ankle, the coordinates of the right ankle, the coordinates of the left knee, the coordinates of the right knee, the coordinates of the left ankle, the coordinates of the right ankle, the coordinates of the left palm root, the coordinates of the right palm root, the coordinates of the left heel, and the coordinates of the right heel; the interactive behavior refers to the interactive behavior between children and other children in the labeled spatiotemporal segment image; the interactive behavior includes playing together and chasing; the spatial relationship refers to the spatial relationship between children and other children in the labeled spatiotemporal segment image; the spatial relationship includes approaching and moving away.

[0075] Specifically, the child tracking module obtains the child's identity matching probability Sp, the number of occluded joints J, the set of coordinates of each joint Gj=[g1,g2,g3,...,gn], interactive behavior and spatial relationship in the previous frame through the joint tracking model, so as to analyze the child's behavior and continuously track the target child.

[0076] Specifically, when the child tracking module outputs the child matching status based on the labeled spatiotemporal segment images, it compares the child identity matching probability Sp of the previous frame with the preset child identity matching probability Sp0, where 60%≤Sp0≤70%. Based on the comparison result, it judges the child matching status of the previous frame and outputs the child matching status based on the judgment result, wherein:

[0077] When Sp < Sp0, the child tracking module determines that the child matching situation in the previous frame is not a match, and outputs the different children as the child matching situation;

[0078] When Sp≥Sp0, the child tracking module determines that the child matching situation in the previous frame is a match, and outputs the same child as the child matching situation.

[0079] Specifically, the preset child identity matching probability refers to a preset value used to judge the matching probability of the child in the previous frame. Since the probability of a true match is mostly concentrated between 60% and 80%, and the probability of a non-match is mostly below 50%, setting 60%≤Sp0≤70% can effectively reduce false judgments. The matching situation of the child in the previous frame refers to the matching situation between the child in the previous frame and the child in the current frame, judged based on the matching probability of the child in the previous frame and the preset child identity matching probability. The matching situation of the child in the previous frame includes a non-match and a match. In this embodiment, the target child refers to the child that the user needs to monitor.

[0080] Specifically, the child tracking module identifies the target child by judging the matching situation of children in the previous frame, and excludes the monitoring of other children, so as to clearly and explicitly track the behavior of the target child.

[0081] Specifically, when the child tracking module performs the first fine-tuning of the output process for child matching, it compares the child's behavioral complexity Fd with the preset child behavioral complexity Fd0, judges the child's behavioral complexity based on the comparison result, and performs the first fine-tuning of the output process for child matching based on the judgment result, wherein:

[0082] When Fd≤Fd0, the child tracking module determines that the complexity of the child's behavior is low and does not perform the first fine-tuning on the output process of the child matching situation;

[0083] When Fd > Fd0, the child tracking module determines that the child's behavior complexity is high, and performs a first fine-tuning on the output process of the child matching situation. This fine-tuning is done by using a complexity coefficient γ to adjust the preset child identity matching probability Sp0, where γ = 1.41 - 0.21 × e -1.23×（Fd-Fd0） After obtaining the refined preset child identity matching probability Sp0`, Sp0` is set to Sp0×γ. The preset child identity matching probability Sp0 is replaced with the refined preset child identity matching probability Sp0`, and the child identity matching probability Sp of the previous frame is compared with the preset child identity matching probability Sp0 again.

[0084] Specifically, the so-called child behavior complexity refers to a specific numerical value of the degree of behavioral complexity mapped based on interactive behavior and spatial relationships. This embodiment does not limit the mapping method of the specific numerical value of behavioral complexity mapped based on interactive behavior and spatial relationships. For example, the child behavior complexity can be obtained by using expert experience to analyze interactive behavior and spatial relationships and mapping the child behavior complexity in the range of 0-1. The preset child behavior complexity refers to a preset value for judging the child behavior complexity. This embodiment does not limit the specific value of the preset child behavior complexity. For example, the specific value of the preset child behavior complexity can be set after sampling the daily behavior of 1000 children. The child behavior complexity situation refers to the situation of the child behavior complexity being high or low based on the judgment of the child behavior complexity and the preset child behavior complexity. The child behavior complexity situation includes the child behavior complexity situation being low and the child behavior complexity situation being high.

[0085] Specifically, the child tracking module improves the accuracy of matching the previous frame by judging the complexity of the child's behavior and performing a first refinement on the output process of the child matching based on the judgment result. When the complexity of the child's behavior is high, the excessive complexity may make the child unclear in the image, resulting in an increased probability of matching the child's identity in the previous frame output by the model. By setting a complexity coefficient that increases with the increase of the child's behavior complexity and multiplying it by the preset probability of matching the child's identity, the value of the probability of matching the child's identity is increased, thereby reducing the impact of high complexity of the child's behavior on the accuracy of matching the previous frame, and thus improving the accuracy of matching the child's identity in the previous frame.

[0086] Specifically, when the preschool tracking module performs a second fine-tuning process during the first fine-tuning process and the judgment process of motion pixel ratio, it calculates the joint occlusion rate Zd based on the number of occluded joints J and the total number of joints JZ, sets Zd=J / JZ, and compares the joint occlusion rate Zd with the preset joint occlusion rate Zd0. Based on the comparison result, it judges the joint occlusion situation, and performs a second fine-tuning process during the first fine-tuning process and the judgment process of motion pixel ratio, wherein:

[0087] When Zd≤Zd0, the child tracking module determines that the joint occlusion is of low probability and does not perform the second fine-tuning process of the first fine-tuning process and the judgment process of the motion pixel ratio.

[0088] When Zd > Zd0, the child tracking module determines that the joint occlusion situation has a high probability of occlusion. A second refinement is then performed on the first refinement process and the process of judging the motion pixel ratio. The preset child behavior complexity Fd0 is then refined using an occlusion coefficient xzd, and xzd is set to = The refined preset child behavior complexity Fd01 is obtained, and Fd01 = Fd0 × xzd is set. The preset child behavior complexity Fd0 is replaced with the refined preset child behavior complexity Fd01, and the child behavior complexity Fd is compared with the preset child behavior complexity again. The preset motion pixel ratio Rt0 is refined a second time using the occlusion coefficient xzd to obtain the refined preset motion pixel ratio Rt03, and Rt03 = xzd × Rt0 is set. The preset motion pixel ratio Rt0 is replaced with the refined preset preset motion pixel ratio Rt03, and the motion pixel ratio Rt is compared with the preset motion pixel ratio Rt0 again.

[0089] Specifically, the total number of joints refers to the total number of joints that can be identified in the joint tracking model, which is 21 joints. The preset joint occlusion rate refers to a preset value for judging the joint occlusion situation. This embodiment does not limit the specific value of the preset joint occlusion rate. For example, the preset joint occlusion rate can be set according to the child's body shape. The joint occlusion situation refers to the level of joint occlusion probability judged based on the joint occlusion rate and the preset joint occlusion rate. The joint occlusion situation includes joint occlusion situation with low occlusion probability and joint occlusion situation with high occlusion probability.

[0090] Specifically, the child tracking module improves the accuracy of judging the child's matching status in the previous frame by judging the joint occlusion status and then performing a second fine-tuning process based on the judgment result of the first fine-tuning process and the judgment process of the motion pixel ratio. When the joint occlusion status is high, the judgment of the complexity of the child's behavior will be inaccurate due to the excessively high joint occlusion probability, and complex behaviors will be easily judged as simple behaviors. At this time, by setting an occlusion coefficient that decreases with the increase of the joint occlusion rate and multiplying it with the preset child behavior complexity, the preset child behavior complexity is reduced, thereby reducing the impact of the excessively high joint occlusion probability on the judgment of the complexity of the child's behavior, and thus improving the accuracy of judging the child's matching status in the previous frame. At the same time, since the excessively high joint occlusion probability will affect the judgment of key frames, the joints will be misjudged as non-key frames after being occluded. At this time, by reducing the value of the preset motion pixel ratio, the impact of the excessively high joint occlusion probability on the judgment of key frames is reduced, thereby improving the accuracy of the judgment of key frames.

[0091] Specifically, the preschool behavior recognition module acquires the preschool matching information output by the preschool tracking module. If the matching information indicates the same preschooler, the images of the same preschooler in the labeled spatiotemporal segment images are recombined into a target preschooler image set. This target preschooler image set is then input into the behavior risk analysis model to obtain a risk-labeled target preschooler image set output by the behavior risk analysis model. This risk-labeled target preschooler image set includes a behavior risk index Rf. The behavior risk index Rf is compared with a preset minimum risk index Rfmin and a preset maximum risk index Rfmax. Based on the comparison results, the degree of behavior risk is determined, and an alert is sent to the supervisor based on the degree of behavior risk.

[0092] When Rf≤Rfmin, the child behavior recognition module determines the level of behavioral risk as low risk and does not alert the supervisor.

[0093] When Rfmin < Rf ≤ Rfmax, the child behavior recognition module determines the behavioral risk level to be medium risk, does not alert the supervisor, and calibrates the preset identity matching probability Sp0 using a risk coefficient fx, set to fx = 1.28 - 0.25 × e -(Rf-Rfmin) e is the base of the natural logarithm. The calibrated preset identity matching probability Sp02 is obtained. Sp02 is set to fx × Sp0. The preset identity matching probability Sp0 is replaced with the calibrated preset identity matching probability Sp02. The identity matching probability Sp is then compared with the preset identity matching probability Sp0 again.

[0094] When Rf > Rfmax, the child behavior recognition module determines the level of behavioral risk to be high and alerts the supervisor.

[0095] Specifically, this embodiment does not limit the specific implementation method of recombinizing images of the same child in labeled spatiotemporal segment images into a target child image set. Those skilled in the art can set it according to actual needs, such as extracting frame images containing the same child from labeled spatiotemporal segment images and sorting them chronologically to obtain the target child image set. The behavioral risk analysis model refers to a convolutional neural network model that takes the target child image set as input and outputs a risk-labeled target child image set. The behavioral risk index refers to the numerical mapping of child behavioral risk output by the behavioral risk analysis model. The child behavior recognition module constructs the behavioral risk analysis model using a behavioral risk analysis model construction method. This embodiment does not limit the specific method of constructing the behavioral risk analysis model; those skilled in the art can set it according to actual needs, such as training the convolutional neural network model using a risk analysis dataset to obtain the behavioral risk analysis model. The risk analysis dataset refers to the training dataset used to construct the behavioral risk analysis model. The dataset includes a historically acquired set of target children's images and a risk-labeled set of target children's images corresponding to the historically acquired set of target children's images. The preset minimum risk index refers to the minimum preset value for judging the degree of behavioral risk, and the preset maximum risk index refers to the maximum preset value for judging the degree of behavioral risk. This embodiment does not limit the specific values ​​of the preset minimum risk index and the preset maximum risk index. Those skilled in the art can set them according to the actual situation. For example, the specific values ​​of the preset minimum risk index and the preset maximum risk index can be set according to the number of people and obstacles in the scene. The degree of behavioral risk refers to the risk level of the child's behavior judged based on the behavioral risk index and the preset minimum risk index and the preset maximum risk index. The degree of behavioral risk includes low risk, medium risk, and high risk. This embodiment does not limit the specific method of alerting the supervisor. Those skilled in the art can set it according to the actual situation. For example, information about the child engaging in risky behavior can be sent to the supervisor's mobile device.

[0096] Specifically, the child behavior recognition module judges the level of behavioral risk and alerts the supervisor based on the judgment result, enabling the supervisor to promptly detect when the target child's behavior is at risk and to control the target child to ensure the safety of the child's behavior. When the level of behavioral risk is medium risk, the module increases the value of the preset identity matching probability by multiplying the risk coefficient that increases with the increase of the behavioral risk index, thereby strengthening the monitoring of the target child and enabling accurate identification of the child's behavior in the future.

[0097] Specifically, the preschool behavior recognition module compares the preschool identity matching probability Sp of the previous frame with the lowest identity matching probability Spmin, and judges the risk of identity matching based on the comparison result. Based on the judgment result, it modifies the degree of behavioral risk, wherein:

[0098] When Sp≥Spmin, the child behavior recognition module determines that the identity matching danger situation is not dangerous and does not modify the degree of behavioral risk;

[0099] When Sp < Spmin, the child behavior recognition module determines that the identity matching is dangerous and modifies the behavior risk level, changing the behavior risk level from low risk to medium risk, and from medium risk to high risk.

[0100] Specifically, the minimum identity matching probability refers to a preset value for judging the dangerous situation of identity matching. This embodiment does not limit the specific value of the minimum identity matching probability. For example, the specific value of the minimum identity matching probability can be set according to the number of people in the scene. The dangerous situation of identity matching refers to whether the identity matching is dangerous based on the identity matching probability and the minimum identity matching probability. The dangerous situation of identity matching includes the dangerous situation of identity matching being not dangerous and the dangerous situation of identity matching being dangerous.

[0101] Specifically, the child behavior recognition module judges the dangerous situation of identity matching and modifies the level of behavioral risk according to the judgment result, thereby improving the supervision of children's behavioral risks. When the child behavior recognition module determines that the dangerous situation of identity matching is dangerous, it is impossible to judge the child's behavioral risk because the identity matching probability is too low. At this time, the level of behavioral risk is modified to increase the level of behavioral risk and alert the supervisor so that the target child can be found and controlled in time.

[0102] Please see Figure 1 As shown, this is a flowchart illustrating the method for analyzing preschool children's behavior based on image recognition in this embodiment. The method includes:

[0103] Step S1: Collect images of the child and the ambient light intensity;

[0104] Step S2: Blur the image of the child to obtain the grayscale value of the blurred image of the child and the current frame, output the binary difference between adjacent frames, judge the proportion of moving pixels, synthesize the spatiotemporal segment image, adjust the judgment process of the proportion of moving pixels, and revise the judgment process of the proportion of moving pixels.

[0105] Step S3: Obtain labeled spatiotemporal segment images based on spatiotemporal segment images, output the matching status of children based on labeled spatiotemporal segment images, perform a first fine-tuning on the output process of children's matching status, and perform a second fine-tuning on the fine-tuning process of the first fine-tuning and the judgment process of motion pixel ratio.

[0106] Step S4: The degree of behavioral risk is judged based on the labeled spatiotemporal segment images and the matching of children, and an alarm is alerted to the supervisor based on the degree of behavioral risk. The degree of behavioral risk is also modified based on the dangerous situation of identity matching.

[0107] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A system for analyzing preschool children's behavior based on image recognition, characterized in that, The system comprises: a data acquisition module for acquiring infant images and ambient light intensity; an edge blur calculation module for performing blur processing on the infant images to obtain blurred infant images and current frame gray values, outputting binary difference values of adjacent frames, judging motion pixel proportion, synthesizing a spatiotemporal segment image, adjusting the judgment process of the motion pixel proportion, and revising the judgment process of the motion pixel proportion; an infant tracking module for obtaining a labeled spatiotemporal segment image from the spatiotemporal segment image, outputting infant matching conditions according to the labeled spatiotemporal segment image, first fine-tuning the output process of the infant matching conditions, and second fine-tuning the fine-tuning process and the judgment process of the motion pixel proportion; an infant behavior recognition module for judging a behavior risk degree according to the labeled spatiotemporal segment image and the infant matching conditions, alarming a supervisor according to the behavior risk degree, and revising the behavior risk degree according to identity matching danger conditions; The edge blur calculation module calculates the average pixel difference Dp according to the frame resolution Wz x Hz and the sum Dch of absolute difference values of all adjacent frame pixels, sets and compares the average pixel difference Dp with a preset average pixel difference Dp0, judges the average pixel difference condition according to the comparison result, and adjusts the judgment process of the motion pixel proportion condition according to the judgment result. The edge blur calculation module compares the ambient light intensity Gq with a preset ambient light intensity Gq0, judges light intensity suitability according to the comparison result, and revises the judgment process of the motion pixel proportion according to the judgment result. The infant tracking module inputs the spatiotemporal segment image into a joint tracking model to obtain a labeled spatiotemporal segment image output by the joint tracking model, wherein the labeled spatiotemporal segment image includes a previous frame infant identity matching probability Sp, a number of occluded joints J, a set of joint coordinates Gj = [g1, g2, g3, …, gn], an interactive behavior, and a spatial relationship. The infant tracking module calculates a joint occlusion rate Zd according to the number of occluded joints J and a total number of joints JZ, sets Zd = J / JZ, compares the joint occlusion rate Zd with a preset joint occlusion rate Zd0, judges joint occlusion conditions according to the comparison result, and second fine-tunes the fine-tuning process and the judgment process of the motion pixel proportion according to the judgment result. 2.The system for analyzing infant behavior based on image recognition according to claim 1, wherein, The edge blurring calculation module inputs the child's image into the face recognition model, obtaining a child's image with face annotations output by the face recognition model. It then uses Gaussian blurring to blur the child's image, resulting in a blurred child image. Finally, it calculates the edge blurring based on the red channel intensity R of the blurred child image. i Green channel intensity G i Blue channel intensity B i The first weighting coefficient α1, the second weighting coefficient α2, and the third weighting coefficient α3 affect the grayscale value Yh of the current frame. i Calculations are performed to obtain the grayscale value Yh of the current frame. i Set Yh i =α1×R i +α2×G i +α3×B i . 3.The system for analyzing infant behavior based on image recognition according to claim 2, wherein, The edge blur calculation module calculates the absolute difference value Dc of the adjacent frame pixels according to the current frame gray value Yh i and the last frame gray value Yh i-1 , sets Dc = |Yh i -Yh i-1 |, compares the absolute difference value Dc of the adjacent frame pixels with the difference value threshold Ty, and outputs the binary difference value Bt of the adjacent frame according to the comparison result, wherein: When Dc ≥ Ty, the edge blur calculation module outputs 1 as the binary difference value Bt of the adjacent frames. When Dc < Ty, the edge blur calculation module outputs 0 as the binary difference value Bt of the adjacent frames.

4. The system for analyzing infant behavior based on image recognition according to claim 3, wherein, The edge blur calculation module calculates a motion pixel proportion Rt according to the number of pixels with a binary difference value of 1 Bt1 and a frame resolution Wz × Hz, sets Rt = Bt1 / Wz × Hz × 100%, compares the motion pixel proportion Rt with a preset motion pixel proportion Rt0, judges the motion pixel proportion according to the comparison result, and outputs a frame attribute of the current frame according to the judgment result, wherein: When Rt ≤ Rt0, the edge blur calculation module determines that the motion pixel proportion is small, and outputs a non-key frame as the frame attribute of the current frame. When Rt>Rt0, the edge blur calculation module determines that the motion pixel proportion condition is large, outputs the key frame as the frame attribute of the current frame, and combines all the frame attributes of the current frame into a spatio-temporal segment image. 5.The system for analyzing infant behavior based on image recognition according to claim 4, wherein, The edge blur calculation module calculates the average pixel difference Dp according to the frame resolution Wz x Hz and the sum Dch of absolute difference values of all adjacent frame pixels, sets and compares the average pixel difference Dp with a preset average pixel difference Dp0, judges the average pixel difference condition according to the comparison result, and adjusts the judgment process of the motion pixel proportion condition according to the judgment result, wherein: When DpDp0, the edge blur calculation module determines that the average pixel difference condition is small, and does not adjust the judgment process of the motion pixel proportion condition; When Dp > Dp0, the edge blurring calculation module determines that the average pixel difference is large, and adjusts the judgment process for the moving pixel ratio. It adjusts the preset moving pixel ratio Rt0 using a deviation coefficient pcx, setting pcx = 1.57 - 0.23 × e -0.38×（Dp-Dp0） e is the base of the natural logarithm. The adjusted preset motion pixel ratio Rt01 is obtained. Rt01 is set to Rt0×pcx. The preset motion pixel ratio Rt0 is replaced with the adjusted preset motion pixel ratio Rt01. The motion pixel ratio Rt is then compared with the preset motion pixel ratio Rt0 again. 6.The system for analyzing infant behavior based on image recognition according to claim 5, wherein, The edge blur calculation module compares the ambient light intensity Gq with the preset ambient light intensity Gq0, judges the light intensity suitability condition according to the comparison result, and revises the judgment process of the motion pixel proportion condition according to the judgment result, wherein: When GqGq0, the edge blur calculation module determines that the light intensity suitability condition is suitable, and does not revise the judgment process of the motion pixel proportion condition; When Gq> Gq0, the edge blur calculation module determines that the light intensity suitable condition is unsuitable, revises the judgment process of the motion pixel proportion condition, revises the preset motion pixel proportion Rt0 through the light intensity coefficient qd, sets qd = 1.38-0.21×e -0.7×（Gq-Gq0） , obtains the revised preset motion pixel proportion Rt02, sets Rt02 = Rt0×qd, replaces the preset motion pixel proportion Rt0 with the revised preset motion pixel proportion Rt02, and compares the motion pixel proportion Rt with the preset motion pixel proportion Rt0 again. 7.The system for analyzing infant behavior based on image recognition of claim 1, wherein, When Sp<Sp0, the infant tracking module determines that the previous frame infant matching condition is not matched, and outputs a non-same infant as the infant matching condition; When Sp≥Sp0, the infant tracking module determines that the previous frame infant matching condition is matched, and outputs the same infant as the infant matching condition. The infant tracking module performs first fine calibration on the output process of the infant matching condition, compares the infant behavior complexity Fd with the preset infant behavior complexity Fd0, judges the infant behavior complexity condition according to the comparison result, and performs first fine calibration on the output process of the infant matching condition according to the judgment result, wherein:

8. The system for analyzing infant behavior based on image recognition according to claim 7, wherein, When FdFd0, the infant tracking module determines that the infant behavior complexity condition is low, and does not perform first fine calibration on the output process of the infant matching condition; The method comprises: When Fd>Fd0, the infant tracking module determines that the infant behavior complexity condition is high complexity, performs first precision correction on the output process of the infant matching condition, performs first precision correction on the preset infant identity matching probability Sp0 through a complexity coefficient γ, sets γ=1.41-0.21×e -1.23×（Fd-Fd0） , obtains the corrected preset infant identity matching probability Sp0`, sets Sp0`=Sp0×γ, replaces the preset infant identity matching probability Sp0 with the corrected preset infant identity matching probability Sp0`, and compares the previous frame infant identity matching probability Sp with the preset infant identity matching probability Sp0 again.

9. A method applied to the system for analyzing infant behavior based on image recognition according to any one of claims 1-8, characterized in that, Step S1, collecting infant images and ambient light intensity; Step S2, performing blur processing according to the infant images to obtain blurred infant images and current frame gray values, outputting binary difference values of adjacent frames, judging the motion pixel proportion condition, combining spatio-temporal segment images, revising the judgment process of the motion pixel proportion condition, and adjusting the judgment process of the motion pixel proportion condition; Step S3, obtaining labeled spatio-temporal segment images according to the spatio-temporal segment images, outputting the infant matching condition according to the labeled spatio-temporal segment images, performing first fine calibration on the output process of the infant matching condition, and performing second fine calibration on the fine calibration process of the first fine calibration and the judgment process of the motion pixel proportion condition; Step S4, judging the behavior risk degree according to the labeled spatio-temporal segment images and the infant matching condition, alarming the supervisor according to the behavior risk degree, and modifying the behavior risk degree according to the identity matching risk condition. ​

Images