A Multimodal States Based Vehicle Descriptor and Dilated Convolutional Social Pooling for Vehicle Trajectory Prediction

Precise trajectory prediction of surrounding vehicles is critical for decision-making autonomous and learning-based approaches are well recognized the robustness. However, state-of-the-art methods ignore 1) feasibility vehicle's multi-modal state information 2) mutual exclusive relationship between global traffic scene receptive fields local position resolution when modeling vehicles' interactions, which may influence accuracy. Therefore, we propose a vehicle-descriptor based LSTM model with dilated convolutional social pooling (VD+DCS-LSTM) to cope above issues. First, each employed as our model's input new vehicle descriptor encoded by stacked sparse auto-encoders proposed reflect deep interactive relationships various states, achieving optimal feature extraction effective use inputs. Secondly, encoder used encode historical sequences composed novel improve spatial interactions. Thirdly, decoder predict probability distribution future trajectories on maneuvers. The validity overall was verified over NGSIM US-101 I-80 datasets method outperforms latest benchmark.