Hand Gesture Recognition using Gaussian Threshold and different SVM kernels

Abstract

Hands play an important part in expressing one’s actions and ideas thus Hand Gesture Recognition (HGR) is very significant in computer vision based gesture recognition for Human Computer Interaction (HCI). In our work, the dataset has been generated for five static hand gestures (Close Hand, Open Hand, Victory Hand, Thumb Down and Thumb Up), by making videos of 10 different users doing the gestures with all possible variations resulting in total 16,240 images extracted from videos. The objective of this thesis is to develop an efficient hand gesture method using image processing and machine learning algorithms and methods. Background Subtraction is important for extracting hands from a static background. After extraction of the hand, the image processing algorithms like Bilateral Filter and Median Blur have been used for smoothing the images. Gaussian Threshold removes the most of the noise in the images and give the clear outline of the hand gesture. Convex hull points are calculated for each hand gesture and number of points in convex hull are taken as feature. After pre-processing the images, the extracted features are given as input to the Support Vector Machine classifier. Comparison of the performance of different SVM kernels i.e. rbfdot, vanilladot, polydot, tanhdot, laplacedot and besseldot, for HGR is done. The accuracy achieved with different SVM kernels varied from 24.17% to 85.07% with training-testing ratio of 70-30% for 16,240 entries in the dataset. The 10-fold cross validation is performed to prove the robustness of the kernel with SVM.