高度な注意機構を用いた車両軌跡予測の向上

抄録

This paper employs deep learning techniques and an LSTM network to output driver attention weights in both temporal and spatial dimensions. A three-lane driving scenario was designed on a driving simulator, where vehicle trajectory data and eye movement data were collected from twelve drivers as they completed lane-changing tasks. The collected data were used to obtain the distribution of spatial and temporal weights under the DS dataset, and the results were compared with the conclusions derived from the NGSIM dataset. Finally, the processed gaze behavior data were fed into the LSTM model, improving the accuracy of trajectory prediction.