With the widespread adoption of Intelligent Transportation Systems, traffic flow forecasting has gradually become a crucial task. Due to the strong spatial-temporal dependencies inherent in traffic flow data, predicting traffic flow has emerged as a challenging task. Existing methods often employ pre-defined static graphs, leveraging prior knowledge of traffic road networks to learn spatial relationships between different traffic segments. However, these methods overlook the spatial-temporal characteristics of traffic flow data, failing to fully capture the temporal dependencies within the data. To address this limitation, we propose a Dynamic Spatial-Temporal Perception Graph Convolutional Networks (DSTPGCN) to capture the complex spatial-temporal dependencies in traffic flow data. Firstly, we utilize the inherent dynamic patterns within historical data across successive time slices, introducing a spatial-temporal perception graph to replace the conventional pre-defined static graph. This graph captures the spatial-temporal dependencies in traffic flow data. Secondly, we design a dynamic spatial-temporal perception graph convolutional module, which aggregate hidden states of neighboring nodes onto the target node, capturing spatial dependencies. Simultaneously, it extracts temporal dependencies from multi-scale temporal convolutions. Extensive experiments on real-world datasets demonstrate the superior predictive performance of our approach.