近年來，自動駕駛技術取得了顯著進展，特別是在路徑預測研究領域。軌跡預測在自動駕駛系統中是一個關鍵的技術挑戰，涉及基於車輛感知到的當前環境數據，精確預測未來的行駛路徑。本研究介紹了一個名為Multi-modal Transformer Path Prediction Version 2 (MTPPV2) 的軌跡預測系統。MTPPV2 在前一版本 MTPP 的基礎上，引入了多頭自注意力機制到Transformer 架構中，以更好地捕捉複雜的時空依賴關係。此外，為了提高預測準確性，我們利用線性插值來填補 NuScenes 數據集中的缺失幀，確保處理過的數據保持足夠的連續性，以進行準確的軌跡預測。最後，為了加快模型訓練速度，系統重新設計以支持多 GPU 訓練模式。實驗結果顯示，在 NuScenes 數據集上，與 MTPP 相比，MTPPV2 在長期軌跡預測的 ADE和 FDE 方面分別提升了 50.5% 和 52.6%。;In recent years, automated driving technology has seen remarkable advancements, particularly in the field of path prediction research. Trajectory prediction presents a crucial technical challenge in automated driving systems, involving the accurate forecasting of future driving paths based on current environmental data sensed by vehicles. This study introduces a trajectory prediction system named Multi-modal Transformer Path Prediction Version 2 (MTPPV2). Building on top of the previous version of MTPP, MTPPV2 incorporates a multi-head self-attention mechanism into the transformer architecture to better capture complex spatial and temporal dependencies. Moreover, to enhance prediction accuracy, we utilize linear interpolation to fill in missing frames in the NuScenes dataset, ensuring the processed data maintain the continuity necessary for accurate trajectory prediction. Finally, to expedite model training, the system is redesigned to support multi-GPU training mode. The experimental results on the nuScenes dataset demonstrate that, compared to MTPP, MTPPV2 improves the ADE and FDE of the prediction of long-term trajectory by 50. 5% and 52. 6%, respectively.
