Brain-computer interface is a transformative human-computer interaction. Brain-computer interfaces based on EEG account for most of the research, and functional near-infrared spectroscopy based brain-computer interfaces are beginning to be valued by researchers because of their unique advantages. In the study, fNIRS was used to measure the oxygenated hemoglobin (HbO) concentration of 15 subjects during walking imagery and idle state, and to perform band-pass filtering and baseline drift correction of HbO signals. Then we extracted the mean, peak, root-mean-square and their combinations of HbO as classification features, and finally used SVM, KNN and LDA for classification, and tested the classification performance of different time windows during the task. The experimental results show that the classification accuracy of the three combined features extracted by SVM is significantly higher than other features and classifiers, reaching 90.374.42%; the classification accuracy of the 2~8s time window is higher than that of other time windows. This study is expected to provide a new alternative active rehabilitation training method for patients with walking dysfunction.