The electromyography (EMG) signals detected when muscle activates can reflect the muscle activation level and has a capability of representing human motions. In this paper, three different upper limb motion recognition methods using features extracted from EMG signals were compared to study their properties under our special circumstance. The three recognition methods are wavelet transform packet (WTP) method, weighted peaks (WP) method and detrended fluctuation analysis (DFA) method. The motions to be classified are elbow flexion and extension, forearm pronation and supination and palmar flexion and dorsiflexion. EMG signals are recorded from biceps brachii, brachioradialis, pronator teres, flexor carpi radialis and extensor carpi radialis longus. Three volunteers participate in the experiments. The experimental results indicate that the WP method has the highest recognition accuracy rate while the WTP method is the most suitable one for real-time implementation.