The invention discloses a
gesture recognition method. The method comprises the steps that common
human body basic actions are captured and are stored as sample actions; according to the sample actions, final training actions are obtained;
graphics rendering is conducted on the training actions and a preliminary
depth map and a corresponding position identification graph are generated; according to the generated
depth map, samples similar to the
depth map collected in real time are synthesized; the synthesized samples are used for calculating corresponding depth characteristic vectors, and a
random forest model is obtained through training; accurate depth figure outline is extracted through area growth based on smoothness constraining; depth characteristic vectors of each pixel of the depth figure outline are calculated based on the
random forest model, and position identification probability of each pixel is determined through the
random forest model; noisy points are filtered and identified based on the
human body part corresponding to each pixel and the probability of each pixel, and skeleton nodes are generated in a polymerized mode; a time-sequential sequence of the skeleton nodes is recorded and a skeleton motion track is generated; the motion tracks of the nodes of the hands are extracted and are matched with a predefined template, and the gesture action type is recognized. The invention further discloses a
gesture recognition device.