Image-based sitting posture detection

Abstract

This project investigates image-based sitting posture detection, where a camera mounted on a computer is used to take a picture while a person is using the computer. Image enhancement techniques such as hitogram equalization, Gaussian filter and median filter are applied to the datasets. We found that converting RGB space to HSI space before historgram equalization improves the effect of contrast enhancement much. Bottleneck feature extraction based on CNN is used to extract features. Fine-tuning backward to last 3 convolutional layers helps improve the feature extraction. Then we compared the performances of various kinds of classification algorithms such as MLP, RVFL and RBF. The architecture and hyper-parameters of MLP are determined by 10 fold cross validation. Random search method is applied to RVFL for tuning the random weights. The center vector of RBF is determined by SOM.