Scholarlink

<p>With the rapid advancement of intelligent driving technology, accurately predicting the motion trajectories of traffic participants in dynamic scenes has become a key challenge in achieving safe and efficient autonomous driving. However, existing methods often struggle with modeling spatio-temporal dependencies and capturing fine-grained dynamic features, especially for high-speed objects. To address these issues, this paper proposes FSTINet, a BEV-based framework that jointly integrates frequency-domain and spatio-temporal information. Two core modules are introduced: the Hierarchical Temporal-Spatial Integrated Module (HTSIM), which combines motion-consistency guided temporal pooling with hierarchical interaction learning to capture multi-scale spatio-temporal dependencies while suppressing noise; and the Frequency-Spatial Fusion Module (FSFM), which leverages a 2D discrete wavelet transform (2D-DWT) to decompose features into high-frequency and low-frequency components, enabling complementary fusion that preserves global semantics and enhances local motion details. Extensive experiments on the challenging nuScenes dataset demonstrate the effectiveness of the proposed method: in low-speed scenarios (speed <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\le \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>≤</mo> </math></EquationSource> </InlineEquation> 5 m/s), the mean error decreases from 0.2535m to 0.2273m, while in high-speed scenarios (speed &gt; 5 m/s), the error drops from 1.0345m to 0.8314m. These results confirm the improved accuracy and robustness of FSTINet in handling diverse and complex traffic dynamics, offering a more reliable solution for class-agnostic motion prediction in autonomous driving. The code will be available at <a href="https://github.com/ZehuaChenLab/FSTINet">https://github.com/ZehuaChenLab/FSTINet</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FSTINet: Frequency-Enhanced Spatial-Temporal Integrated Network for Class-Agnostic Motion Prediction

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Yuzhuo Feng,
  • Andong Xue,
  • Zehua Chen

摘要

With the rapid advancement of intelligent driving technology, accurately predicting the motion trajectories of traffic participants in dynamic scenes has become a key challenge in achieving safe and efficient autonomous driving. However, existing methods often struggle with modeling spatio-temporal dependencies and capturing fine-grained dynamic features, especially for high-speed objects. To address these issues, this paper proposes FSTINet, a BEV-based framework that jointly integrates frequency-domain and spatio-temporal information. Two core modules are introduced: the Hierarchical Temporal-Spatial Integrated Module (HTSIM), which combines motion-consistency guided temporal pooling with hierarchical interaction learning to capture multi-scale spatio-temporal dependencies while suppressing noise; and the Frequency-Spatial Fusion Module (FSFM), which leverages a 2D discrete wavelet transform (2D-DWT) to decompose features into high-frequency and low-frequency components, enabling complementary fusion that preserves global semantics and enhances local motion details. Extensive experiments on the challenging nuScenes dataset demonstrate the effectiveness of the proposed method: in low-speed scenarios (speed \(\le \) ≤ 5 m/s), the mean error decreases from 0.2535m to 0.2273m, while in high-speed scenarios (speed > 5 m/s), the error drops from 1.0345m to 0.8314m. These results confirm the improved accuracy and robustness of FSTINet in handling diverse and complex traffic dynamics, offering a more reliable solution for class-agnostic motion prediction in autonomous driving. The code will be available at https://github.com/ZehuaChenLab/FSTINet.