The Research On Multimodal Trajectory Prediction Method Based On Spatiotemporal Attention

Trajectory prediction is a technology that uses data mining,machine learning,and other techniques to predict the movement trajectory of traffic participants.This technology has been widely used in traffic management,driving assistance,and autonomous driving,and has therefore received great attention.With the continuous advancement of technology,trajectory prediction methods have evolved from physical models and traditional machine learning models to deep learning models.Although there has been significant progress in trajectory prediction,there are still many difficulties and problems.Firstly,trajectory prediction is multimodal,meaning that even if the historical trajectories are completely the same,the future trajectories may still differ.In addition,the movement trajectory of a single traffic participant depends not only on its own historical trajectory,but also on the influence of other traffic participants.Finally,due to the high safety requirements of the traffic environment,the predicted trajectory must be real-time and reliable.Therefore,there are still many difficulties and problems in trajectory prediction that require further research.This thesis conducted the following research based on existing studies.(1)Addressing the problem of only using the last moment feature for spatial information interaction.A vehicle trajectory prediction method based on spatiotemporal dynamic attention network was proposed.After encoding the trajectory data using a recurrent neural network,this method uses a hierarchical attention mechanism to model time and space interactions.Then,a specific driving intent feature fusion module was proposed to adaptively integrate the extracted time and space features,thereby generating multimodal trajectory predictions for different driving intents.(2)Addressing the problem of the lack of holistic attention to the Transformer model in existing trajectory prediction methods.A vehicle trajectory prediction method based on intent-aware Transformer was proposed.This method applies Transformer to trajectory prediction,uses an enhanced graph attention mechanism to model spatial interactions between vehicles,and uses an attention-based temporal attention module to capture temporal dependencies.By stacking the spatial attention and temporal attention modules,higher-order interaction features can be obtained.In addition,this method proposes an intent-aware query generation module to incorporate driving intent into the decoding phase,enabling simultaneous prediction of future trajectory at different time steps.(3)Addressing the problem of indirect dependence between spatial interaction features at different time steps when modeling time and space separately.A pedestrian trajectory prediction method based on spatiotemporal joint attention Transformer was proposed.This method improves the original graph attention network by integrating time sequence encoding,pedestrian identity,and sparse attention into the graph attention network to extract time and space interaction features in trajectory data.This method also designs a random query generation module,enabling the parallel generation of multiple possible trajectories.In summary,the trajectory prediction methods proposed in this thesis can capture the time and space features in trajectory data and generate reasonable multimodal trajectories.Experimental results show that compared with other trajectory prediction methods,the proposed method can achieve more accurate predictions.