Surface electromyography (sEMG) has been applied extensively in gestures recognition. In order to improve the recognition accuracy, multi-channel sEMG is conventionally sampled, which also increases the complexity of applications. To solve the problem, a novel gesture recognition method based on sEMG decomposition is proposed. Sampling sEMG signals from the muscle of human upper limb by a single-channel electrode; then decomposing the sEMG into six motor unit action potential trains (MUAPTs) and the decomposition process includes 2-order differential filtering, threshold calculation, spike detection and hierarchical clustering. Afterwards, the features, including integral of absolute value, maximum value, median of non-zero value and semi-window energy, are extracted to form a feature matrix, whose dimension is then reduced by the principal component analysis. Finally, support vector machine is employed to recognize five different hand gestures, and 80.4% of accuracy can be obtained, while only about 70% of recognition accuracy can be achieved by traditional methods without sEMG decomposition.