Local Self-attention Temporal Encoding Network for Trajectory Prediction
In response to the challenge of traditional encoders failing to capture local trajectory changes,such as brief stops or sharp turns,at the local temporal scale,leading to decreased prediction accuracy,we propose a Local-SelfAttention Temporal Encoding Architecture(Loc-SelfAttention).This algorithm fully utilizes the superior local perception capabilities of small-scale convolution kernel to sharply capture and extract local trajectory changes.Leveraging self-attention mechanism,it dynamically assigns attention weights to extracted local features based on their impact on future trajectory distribution,effectively filtering out noise and outliers,selecting effective local features,and thereby enhancing trajectory prediction accuracy.The experimental results show that the Loc-SelfAttention algorithm reduces minimum average displacement error by 3.7%,minimum final displacement error by 3.1%,and miss rate by 4.8%compared to the HOME trajectory prediction algorithm on the Argoverse Motion Forecasting Benchmark.It has certain application prospects in fields such as autonomous driving and intelligent traffic management.
local temporal scalesmall-scale convolution kernelself-attention mechanismtemporal encoding
万方数据
维普
