Abstract:In video-based automatic gesture recognition

interfering factors such as hand occlusion and color closeness between background and skin of the signer often result in errors in hand segmentation and hand tracking. In light of this

this article introduces a method to differentiate left and right hand in videos of signing. Through the differentiation of each hand

hand segmentation in videos can be assisted

and left/right hand tracking error due to hand occlusion can be diminished. In this method

feature of histograms of oriented gradients (HOG) in the training samples using single hand

either left or right hand

was firstly extracted. Next

principal component analysis (PCA) was carried out on hand feature

and the three different patterns of left hand

right hand

and hand occlusion were then generated through eliminating redundant information. Lastly

left-right hand differentiation was realized through pattern matching. The results of the experiment showed that the method adopted in this study helped realize relatively sound recognition rate.