Effectively predicting the future risk indicators of multiple traffic participants under the driver's field of vision is the key to providing risk warnings to human drivers and avoiding potential collision risk. Most existing research on risk only considers the pairwise interaction between a single individual and the vehicle in the scene， and conducts research from the perspective of evaluation rather than prediction， while ignoring the different interaction between heterogeneous traffic participants and future risk status. This paper proposes a heterogeneous multi-objective risk prediction method Risk-STGCN based on spatiotemporal graph convolutional neural network， using graph convolution and temporal convolution to learn single-frame scene graph information and timing information respectively， combined with multi-layer timing prediction network to predict the multi-objective risk indicator TTC. Training and verification are conducted on the open source data set BLVD and the real vehicle self-collected data set， which is then compared with commonly used sequence prediction models. The experimental results show that the average TTC error of the proposed model on different data sets is less than 0.95 s， with multiple experimental indicators better than other models mentioned in this paper. The proposed model has good robustness and improves the interpretability of risk prediction in complex traffic scenarios.
