A robust deep belief network-based approach for recognizing dynamic hand gestures

The task of dynamic hand gesture recognition includes several challenges arising due to variations in hand appearances, scale of the hand and the spatio-temporal variability of the gestures. To overcome these difficulties, in this work, we present a robust approach to detect and track a hand, and to recognize the hand trajectory based on DBNs (Deep Belief Networks) in conjunction with the SVM+HOG (Support Vector Machines + Histograms of Oriented Gradients) framework and the Mean Shift approach. Our method relies on merging the advantages of accurate feature-based hand detection, fast local searching-based hand tracking, and smart deep learning-based hand trajectory recognition. Our approach entails three different phases, i.e. hand detection, hand tracking and hand trajectory recognition. The first two phases work in collaboration, with the SVM+HOG framework utilized for hand detection and the Mean Shift algorithm used for tracking the hand. For the third phase, we design a method based on DBNs. Through ample experiments, we prove that our proposed hand tracker outperforms several other state-of-the-art trackers and our hand gesture recognition approach shows excellent results on numerous challenging video sequences.