The automatic recognition of pedestrian intentions is a difficult issue in location-based services, which is crucial for establishing intelligent navigation services and new human-computer interaction method. Currently, using behavior patterns to estimate pedestrian intentions has become a mainstream solution, but this approach relies on multiple sensors and has time delays. This article proposes a pedestrian intention detection method based on brain imaging technology, which interprets pedestrian turning intentions through multi-channel, high-resolution EEG signals. Firstly, according to the standard motor imagery paradigm, EEG samples corresponding to four types of intentions within road intersection scenes were collected, including straight ahead, stop, left turn, and right turn. Then, by fusing the features of EEG in time-frequency domain, spatial domain, and functional connectivity domain, the spatiotemporal functional connectivity networks (STFCNs) of EEG are constructed to express the process of EEG activity, facilitating the capture of EEG features highly related to the intent. Finally, a graph convolutional neural network was used to encode the STFCNs, completing the mapping from EEG to four types of navigation intentions. The experimental results show that the average accuracy (F1 score) of detecting four types of intentions using a short time window (1 s) is 0.443±0.062, and the highest accuracy can reach 0.571. The average accuracy with a long time window (6 s) is 0.525±0.084, and the highest accuracy is 0.665. The detection accuracy of this method is slightly better than other classification algorithms, and its detection ability for forward and stop intentions is excellent, up to 0.740 and 0.700.
