Method for body fat detection

Abstract

This invention relates to a method for detecting body fat. It utilizes an RGB camera and a thermal imager to capture images of the subject from multiple angles. After segmenting the human body contour using machine learning, the thermal imaging contour is mapped and aligned with the RGB image. A binary mask is generated based on the thermal imaging temperature distribution to identify the subject's skin and clothing. The clothing area is then removed to restore the true human body contour. A closed model is generated using 3D point clouds. The volume of the geometric body is calculated and accumulated using the polyhedral volume formula to obtain the subject's volume value. Finally, the subject's volume value is compared with magnetic resonance proton density fat fraction data. A deep learning model is used to separate fat and non-fat portions, generating a body fat estimation result.

Claims

1. A method for detecting body fat, comprising the following steps: using an RGB camera and a thermal imager to capture images of a subject from multiple angles, acquiring RGB image data and thermal image data respectively; using a machine learning model to perform human body contour segmentation on the RGB image data and the thermal image data; performing a mapping alignment procedure on the thermal image contour of the thermal image data and the RGB image data, and generating a binary mask based on a thermal imaging temperature distribution, using the binary mask to identify the subject's skin and clothing, removing the area of ​​the clothing in the RGB image to restore a realistic human body contour; generating a closed three-dimensional model based on the realistic human body contour data and multiple three-dimensional point cloud data; decomposing the three-dimensional model into multiple geometric bodies using a polyhedral volume formula, calculating the volume of each of the geometric bodies and accumulating them to obtain a subject's volume value; and comparing the subject's volume value with magnetic resonance proton density fat fraction data, and using a deep learning model combined with the spatial features of the three-dimensional model to separate the fat and non-fat portions of the subject's volume value, generating a body fat estimation result.

2. The body fat detection method as described in claim 1, wherein in the step of using an RGB camera and a thermal imager to capture images of a subject from multiple angles and acquire RGB image data and thermal image data respectively, the angles include the front, side and back of the subject.

3. The body fat detection method as described in claim 1, wherein in the step of using an RGB camera and a thermal imager to capture images of a subject from multiple angles to acquire RGB image data and thermal image data respectively, the RGB camera is used to record a high-resolution 2D color image to acquire one of the subject's appearance features and a texture information.

4. The body fat detection method as described in claim 1, wherein in the step of segmenting the human body contour using a machine learning model on the RGB image data and the thermal image data, the machine learning model is a convolutional neural network or a fully convolutional network.

5. The body fat detection method as described in claim 1, wherein in the steps of performing a mapping alignment procedure between a thermal imaging contour of the thermal imaging image data and the RGB image data, generating a binary mask based on a thermal imaging temperature distribution, identifying a subject's skin and clothing using the binary mask, and removing the area of ​​the clothing in the RGB image to restore a real human body contour data, the thermal imager has a high thermal sensitivity of less than 50 mK, and the high thermal sensitivity is used to acquire a real human body contour under the clothing.

6. The body fat detection method as described in claim 1, wherein in the step of generating a closed three-dimensional model based on the real human body contour data and a plurality of three-dimensional point cloud data, the three-dimensional model generates a pair of point cloud data for a repair procedure and a smoothing procedure, so that the three-dimensional model has integrity and continuity.

7. The body fat detection method as claimed in claim 1, wherein in the step of decomposing the three-dimensional model into a plurality of geometric bodies using a polyhedral volume formula, calculating the volume of each of the geometric bodies and summing them to obtain a subject's volume value, the volume is a tetrahedron formed by a plurality of triangles on the surface of the three-dimensional model and a reference point, and the volume of each tetrahedron is calculated and summed using a mixed product formula.

8. The body fat detection method as described in claim 1, wherein in the steps of comparing the subject's volume value with magnetic resonance proton density fat fraction data, and separating the fat and non-fat portions of the subject's volume value through a deep learning model combined with a spatial feature of the three-dimensional model to generate a body fat estimation result, the deep learning model is an iterative nearest-point algorithm or a rigid body transformation algorithm.

Images