The invention discloses a random projection multi-kernel learning-based hand gesture identification method comprising the following steps: hand gesture images are collected and preprocessed, preprocessing operation comprises hand gesture positioning operation and hand gesture segmenting operation, sift characteristics are extracted from preprocessed and segmented hand gestures, a K-means algorithm is adopted for training a learning dictionary, an iteration dictionary is used for updating the algorithm and the dictionary, the gesture images are subjected to space pyramid dividing operation, the trained dictionary is used for encoding the sift characteristics of the hand gesture images in each space pyramid layer, and therefore characteristic vectors can be obtained and subjected to cascading operation; random projection is adopted for subjecting the characteristic vectors to dimensional reducing operation; as for a characteristic vector learning kernel matrix after dimensional reducing of each pyramid layer, a multi-kernel model learning algorithm is adopted for classified learning, and an optimal kernel matrix combination coefficient is obtained. Via the method disclosed in the invention, problems of background interference, high complexity, long time consumption, low identification rate and the like in a conventional hand gesture identification method can be solved.