Gesture recognition based on sEMG signals has been studied for many years, but its practical applications are still limited due to the gap between laboratory study and practical applications. Among the many factors that influence the sEMG signals, the time factor and environmental noise was focused in this study. The data acquisition experiments were conducted for three consecutive days when seven electrodes were placed on the subject's forearm and hand. For each session, the subject was seated in a non-vibration environment and in a chair that vibrates in the vertical direction with vibration frequencies ranging from 0 Hz to 200 Hz randomly. Results of signal recognition with k-Nearest Neighbors (kNN) classifier show that the vibration will reduce the recognition accuracy, but the influence is limited when the test samples are from the same environment as the training samples. And the results of inter-session tests show that history sEMG signals is unreliable for new session sample recognition. When only one sample per movement was used as the training sample, the recognition accuracy was much higher than that with historical information. To solve such a problem, a fast online adapting algorithm based on training sample management and kd-tree was proposed. The results show that a much higher accuracy can be obtained when classifier is updated by using the samples that incorrectly recognized during calibration stage. The accuracies can be enhanced by 9.9% ± 8.2%, and 21.8% ± 10.7% respectively for the samples from non-vibration environment and vibration environment.