Scholarlink

<p>Uncertainty prediction remains a core challenge in pedestrian trajectory forecasting, particularly in complex and dynamic environments such as intersections, where human motion exhibits significant variability and strong contextual interdependencies. Conventional deterministic predictors often lack adaptability and fail to capture individual-specific motion patterns, whereas stochastic sampling approaches are prone to producing long-tailed distributions, resulting in unrealistic or inconsistent trajectory outcomes. To rigorously address the inherent uncertainty and multimodality in trajectory forecasting, this study presents an augmented GRU-based probabilistic predictor equipped with feature-gated units and Bayes By Backprop (BBB) integration to improve prediction accuracy and diversity. Specifically, we first develop a Bayes By Backprop-Fully Connected (BBB-FC) feature learning framework to extract individual motion features. On this foundation, an adaptive Bayesian Dynamic Graph (BDG) extractor is introduced to model spatial interactions and capture critical inter-pedestrian dependencies. Subsequently, we propose an augmented Bayes By Backprop–Gated Recurrent Unit (BBB-GRU) that integrates acceleration-gated, velocity-gated, and fusion-feature-gated units to characterize spatiotemporal uncertainties while preserving individual features. Finally, a composite loss and a posterior perturbation sampling (PPS) scheme are formulated to enforce kinematic realism, promote forecast diversity, and holistically capture predictive uncertainty. Comprehensive experiments on the ETH, UCY, and SDD benchmarks demonstrate the effectiveness of the proposed predictor. Notably, our predictor outperforms the DSTIGCN baseline, yielding relative reductions in ADE/FDE of 48.00%/60.00% on ETH-UCY and 8.11%/8.20% on SDD.</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

Uncertainty-aware pedestrian trajectory forecasting with a feature-augmented gated recurrent unit

  • 期刊论文

  • 发表日期：

    2026年7月9日

查看全文
  • Yanbo Zhang,
  • Liying Zheng

摘要

Uncertainty prediction remains a core challenge in pedestrian trajectory forecasting, particularly in complex and dynamic environments such as intersections, where human motion exhibits significant variability and strong contextual interdependencies. Conventional deterministic predictors often lack adaptability and fail to capture individual-specific motion patterns, whereas stochastic sampling approaches are prone to producing long-tailed distributions, resulting in unrealistic or inconsistent trajectory outcomes. To rigorously address the inherent uncertainty and multimodality in trajectory forecasting, this study presents an augmented GRU-based probabilistic predictor equipped with feature-gated units and Bayes By Backprop (BBB) integration to improve prediction accuracy and diversity. Specifically, we first develop a Bayes By Backprop-Fully Connected (BBB-FC) feature learning framework to extract individual motion features. On this foundation, an adaptive Bayesian Dynamic Graph (BDG) extractor is introduced to model spatial interactions and capture critical inter-pedestrian dependencies. Subsequently, we propose an augmented Bayes By Backprop–Gated Recurrent Unit (BBB-GRU) that integrates acceleration-gated, velocity-gated, and fusion-feature-gated units to characterize spatiotemporal uncertainties while preserving individual features. Finally, a composite loss and a posterior perturbation sampling (PPS) scheme are formulated to enforce kinematic realism, promote forecast diversity, and holistically capture predictive uncertainty. Comprehensive experiments on the ETH, UCY, and SDD benchmarks demonstrate the effectiveness of the proposed predictor. Notably, our predictor outperforms the DSTIGCN baseline, yielding relative reductions in ADE/FDE of 48.00%/60.00% on ETH-UCY and 8.11%/8.20% on SDD.