<p>Crowd flow prediction has become an important issue in urban management, especially in regulating crowd flow during congested periods. Accurately predicting future congestion on a road section requires in-depth analysis of crowd flow data and influencing factors. However, existing prediction methods fail to fully integrate spatiotemporal features and effectively utilize environmental and historical information. This paper proposes a spatiotemporal multi-head attention graph convolutional network for pedestrian flow prediction, enhanced with knowledge graphs for improved accuracy(STMHAGCN-KG). First, we build online and offline knowledge graphs based on external scene factors, and integrate historical pedestrian traffic with knowledge through a dedicated module to obtain a pedestrian traffic matrix that integrates knowledge. Secondly, we capture spatial features through multiple feature graphs, use enhanced LSTM and multi-head attention mechanisms to model spatiotemporal dependencies, fuse spatiotemporal features with the pedestrian traffic matrix, and finally generate traffic prediction values in the fully connected layer. Experiments on real-world pedestrian data show that the proposed method achieves superior performance compared to traditional and state-of-the-art models.</p>