International Journal of Science, Engineering and TechnologyAuthors: Shubham Bawari, Paras Chandra
Abstract: The NeuroFit application before was using an CNN in web browser to decide the body type by looking at webcam pictures, and after that, Google Gemini system was used for producing a fitness plan. Even if this way worked, it always needed internet to be on, so it failed sometimes for students having a bad connection. With version two, the main system changed. Now classification for body type used a Random Forest which looks at seven biomechanical measurements that are coming from the Media Pipe Pose points instead of using original image pixels. Dependencies like Media Pipe WASM files and other necessary files come with the app so after the first setting up there are not any internet calls. Instead of three possible body types, the classifier increased decision categories to four BMI groups based on WHO: Underweight, Normal Overweight and Obese. The algorithm learned patterns from 6,000 made-up examples with classes mixed at BMI cutoffs, and gets 82 percent accuracy on the test samples. Height error for 30 sample testers was an average of 4.2 centimetres. Usability for 25 AKTU students got a SUS mark of 81.2 which was more than first version’s 78.5. The system’s countdown was better too, increasing from 60 to 94 percent after updating thresholds in frame. The main discovery is that the Random Forest with landmarks can be explained with feature importance calculation, is accurate and does not need the network.
