With the advent of artificial intelligence era, brain-computer interface (BCI) technology plays an important role in various fields, among which steady-state visual evoked potential (SSVEP) is the most widely used paradigm in BCI. For the SSVEP-BCI with few target stimuli, this paper proposes a feature extraction algorithm based on multi-channel wavelet packet coefficients (MCWPC), and uses a naive Bayes classifier to classify the target stimulus frequencies. It solves the problem of low accuracy of traditional Fourier transform (FFT) algorithm in a short time. By studying 280 groups of experimental data of 7 subjects, it is found that the average accuracy of the algorithm is 95.16% and information transfer rate (ITR) is 49.77bit/min in 2 seconds. It is significantly better than the traditional FFT algorithm and the popular filter bank canonical correlation analysis (FBCCA) algorithm (with a maximal accuracy of 70.67% and 89.51%). The experimental results show that the feature extraction algorithm based on MCWPC has a strong advantage in SSVEP-BCIs with few target stimuli.