A method for pattern recognition using bag-of-words model and neural network

Abstract

The purpose of the project is to create an algorithm for real-time hand gesture recognition in video frames captured directly from the camera. The proposed algorithm is based on the bag-of-features (or bag-of-words) model, SURF-descriptor, k-means clustering, and neural network classification method. The bag-of-words model combined with SURF and k-means is used to create feature vectors, which then are fed as input data for the neural network. The algorithm is trained and tested with a self-made image data set. Experiments with various testing data sets demonstrate that the proposed algorithm ensures a high processing speed (less than 40 ms for each frame) to be able to perform in real time with data captured directly from a camera, keeps being invariant to transformations of the object in the video frame (including rotation, scaling and affine transition), and provides high recognition accuracy (~ 90%).