Adaptive HOG-LBP Based Learning for Palm Tracking

Abstract

Hand tracking is a challenging research direction in computer vision field. Although gesture tracking algorithm has been widely applied to the system of human-computer interaction, it is difficult to meet the robustness and the real-time requirements because of the hand lacks sufficiently rich texture information for discrimination. In this paper, we propose a new method named Adaptive HOG-LBP features to track the palm in the unfettered color images by fusing the HOG features and LBP features. The fused features could give more texture information of the edge features and sub-structure of the palm, and are less sensitive to light variations and background clusters. Macro structural features of the palm contain the maximum amount of information that can be used for discrimination, so we can use the fusion of HOG features and LBP features to process palm detection. The fusion features are inputted into linear SVM classifier learning. Experimental results show that in our own established palm dataset, performance has been improved significantly. Particularly worthy of note is that in our own established database, most of gesture pictures are of low quality, and the palm library contains a wealth of inter-plane and out-of-plane rotated pictures.