Examining Quality of Hand Segmentation Based on Gaussian Mixture Models

Results of examination of various implementations of Gaussian mixture models are presented in the paper. Two of the implementations belonged to the Intel's OpenCV 2.4.3 library and utilized Background Subtractor MOG and Background Subtractor MOG2 classes. The third implementation presented in the paper was created by the authors and extended Background Subtractor MOG2 with the possibility of operating on the scaled version of the original video frame and additional image post-processing phase. The algorithms have been evaluated for various conditions related to stability of background. The quality of hand segmentation when a whole user's body is visible in the video frame and when only a hand is present has been assessed. Three measures, based on false negative and false positive errors, were calculated for the assessment of segmentation quality, i.e. precision, recall and accuracy factors.