Dynamic hand gesture recognition from Bag-of-Features and local part model

This paper discusses the use of Bag-of-Features and a local part model approach for bare hand dynamic hand gesture recognition from video. We used dense sampling to extract local 3D multiscale whole-part features. We adopted three dimensional histograms of a gradient orientation (3D HOG) descriptor to represent features. K-means++ method has applied to cluster the visual words. Dynamic hand gesture classification was completed by using a Bag-of-features (BOF) and non-linear support vector machine (SVM) method. A BOF do not track the order of events. To counter the unordered events of BOF approach, we used a multiscale local part model to preserve temporal context. Initial experimental results on newly collected complex dataset show a higher level of recognition.