Hand detection is a challenging task in hand gesture recognition system and the detection results can be easily affected by changes in hand shapes,viewpoints,lightings or complex backgrounds.In order to detect and localize the human hands in static images against complex backgrounds,a hand detection method based on a mixture of multi-scale deformable part models is proposed in this paper,which is trained discriminatively using latent SVM and consists of three components each defined by a root filter and three part filters.The hands are detected in a feature pyramid in which the features are variants of HOG descriptors.The experimental results show that the proposed method is invariant to small deformations of hand gestures and the mixture model has a good performance on NUS hand gesture dataset-Ⅱ.