Overview / Usage

Excess of body fat often leads to obesity. Obesity is typically associated with serious medical diseases like heart disease, and diabetes. Accordingly, knowing the body fat is an extremely important issue since it affects everyone’s health. However, accurate methods of assessing body fat are expensive, inconvenient, and require immobile equipment. Hence , the goal of the project is to provide people a cost effective and easy way of calculating body fat from the ease of their homes. It captures the surface of the human body using using camera . The developed systems can be implemented in clinical or personal settings and be utilised as a public health research tool and deployed widely given the low-cost of the hardware required. In addition to the immediate impact that the system will have on managing obesity, the project will have a broad impact on a number of areas. A large database of such shapes captured over time may lead to ways to predict how an individual's body shape will change given a particular intervention.

Methodology / Approach

Convolutional Neural Networks are quite popular Pattern Recognition tools that demonstrate an impressive ability of predicting categories and numerical values of imaged scene parameters directly from the pixel array input. The idea of this work was to build a deep network linking image layers with convolutional filters adding nonlinear units. The purpose of this CNN-based fat estimator was to provide a method to automatically measure whole body fat percentage from a inexpensive and quick range scan.

• A data set of depth images rendered from 3D scans was used.
• Depth images have been created simulating a front and a back acquisition from a distance of about
• Images were re scaled for the requirements of transfer learning architecture.
• In my work I have tried to estimate whole body fat percentage, not an absolute value, so I avoided the estimation of exact body and evaluated body shape features mapping the images to a reference scale.

Steps for future:

Measurements are to be put in a vector and regression algorithms are to be trained on sets of feature vectors with associated ground truth fat percentages.

On full body images, only the depth variation in abdominal area was considered, the area of the silouhette, the apparent diameter of the neck, are planned to to be included later.

Challenges ahead:

However, a mandatory step to take will be the creation of larger datasets including non-fit subjects with wider fat variation, trying to balance the dataset and reaching an order of magnitude more in terms of number of samples. To develop a generic application for body fat estimation from depth sensor data, it is necessary to train the system with images of subjects belonging to different ethnic groups as the fat distribution patterns varies within the different subsets.

Technologies Used

• Convolutional neural network
• Image processing
• Machine learning
• Deep learning
• Transfer learning
• Data augmentation
• ResNet-50 regressor