Aiming at the deficiency of using empirical threshold to detect the starting and ending points of active segment of surface electromyography (sEMG) signal in traditional methods, an adaptive detection method of sEMG starting and ending points based on time-frequency point density is proposed. Time-frequency point density is innovatively proposed as the characteristic parameter of surface EMG signal in this method. Firstly, butterworth bandpass filtering and wavelet threshold denoising are used to preprocess the sEMG signals in Ninapro DB8 dataset. Short-time Fourier transform is used for time-frequency analysis of signals. Secondly, the sEMG signal is divided into several continuous unit time-frequency windows, the number of frequency points in the windows is counted, and the time-frequency point density (TFPD) characteristic parameters are extracted. Finally, the TFPD results are adaptively normalized in the interval ［-1,1］, and the start and end of EMG signals are detected by using the binary judgment method based on sliding window. The experimental results show that this method can detect the start and end of sEMG signal activity segment in quasi-real time within 0.5 s, and the accuracy is nearly 100%. Compared with other common algorithms, the proposed method has better accuracy. The influence of individual differences can be eliminated through normalized positive and non-positive values, and the adaptability is strong. In addition, the proposed method has strong practicability in gesture recognition system.