Robust gesture recognition in video requires segmentation of the meaningful gestures from a whole body gesture sequence. This is a challenging problem because it is not straightforward to describe and model meaningless gesture patterns. This paper presents a new method for simultaneous spotting and recognition of whole body key gestures. A human subject is first described by a set of features encoding the angular relations between a dozen body parts in 3D. A feature vector is then mapped to a codeword of gesture HMMs. In order to spot key gestures accurately, a sophisticated method of designing a garbage gesture model is proposed; a model reduction which merges similar states based on data-dependent statistics and relative entropy. This model provides an effective mechanism for qualifying or disqualifying gestural motions. The proposed method has been tested with 20 persons' samples and 80 synthetic data. The proposed method achieved a reliability rate of 94.8% in spotting task and a recognition rate of 97.4% from an isolated gesture.