Multi-modality trajectory prediction with the dynamic spatial interaction among vehicles under connected vehicle...

Area of Science:

• Intelligent Transportation Systems
• Computer Vision and Pattern Recognition
• Machine Learning and Artificial Intelligence

Background:

• Connected vehicle environments introduce complex interactions and large data inputs, overwhelming traditional trajectory prediction models.
• Existing models struggle with dynamic, interactive lane-changing scenarios due to their reliance on historical data of only the target vehicle.
• A need exists for stable, targeted lane-changing behavior prediction methods tailored to connected vehicle characteristics.

Purpose of the Study:

• To propose a multi-modality trajectory prediction model (STA-LSTM) for analyzing interactive behaviors in connected vehicle lane-changing scenarios.
• To enhance trajectory prediction by incorporating dynamic spatial interactions and expanding multi-modality feature inputs.
• To improve the accuracy and robustness of lane-changing predictions in complex, interactive traffic.

Main Methods:

• Developed STA-LSTM, a multi-modality trajectory prediction model integrating spatial grid occupancy for vehicle interaction modeling.
• Introduced a space-dimensional attention mechanism to adaptively weigh surrounding vehicle influences on the target vehicle.
• Incorporated an attention module into the LSTM decoder (time dimension) to identify significant historical features and contextual information for robust prediction.

Main Results:

• The STA-LSTM model demonstrated lower Root Mean Square Error (RMSE) and Negative Log-Likelihood (NLL) compared to baseline models.
• Achieved RMSE values of 0.46m, 1.15m, 1.89m, 2.84m, and 4.05m at prediction times of 1s, 2s, 3s, 4s, and 5s, respectively.
• The model accurately predicts vehicle interactions and travel paths of surrounding vehicles in connected environments.

Conclusions:

• STA-LSTM effectively handles complex interactive lane-changing scenarios in connected vehicles.
• The model's attention mechanisms and interaction modeling improve prediction accuracy and robustness.
• This approach offers a more reliable method for trajectory prediction in advanced driver-assistance systems and autonomous driving.