Scholarlink

<p>State estimation in multi-object sensing environments is commonly performed using per-trajectory Bayesian filters, while upstream components such as detection, data association, and track management determine identity continuity and system-level performance. In practice, Unscented Kalman Filters (UKFs) are widely used for nonlinear estimation but rely on globally fixed uncertainty parameters, including process noise, measurement noise, and sigma-point scaling, despite strong variation across object categories and sensing conditions. In this work, we introduce Category-Aware Adaptive Unscented Kalman Filtering (CAT-GAN-UKF), a trajectory-level adaptive filtering framework in which uncertainty parameters are updated online using a category-conditioned learning module driven by innovation-based diagnostics and contextual cues. The proposed method preserves the recursive Bayesian structure of the UKF while enabling dynamic, context-dependent adjustment of uncertainty assumptions. Importantly, the evaluation in this study is conducted strictly at the trajectory level using pre-associated object-centric sequences derived from a large-scale autonomous-driving dataset. A leakage-free experimental protocol is employed, in which models are trained on past scenes and evaluated on future scenes, ensuring that adaptation policies generalize across environments. The proposed approach is evaluated against a broad set of classical and learning-based baselines, including EKF, PF, UKF-IAE, KalmanNet, and LSTM-enhanced filters, under identical trajectory-centric conditions. Additional analyses include Q/R sensitivity experiments, Monte Carlo robustness evaluation under measurement perturbations and outliers, and empirical uncertainty consistency metrics. Across diverse object categories and sensing regimes, CAT-GAN-UKF demonstrates consistent improvements in trajectory-level estimation accuracy and improved consistency between predicted uncertainty and realized estimation error. These results establish that category-conditioned online adaptation can significantly enhance estimator-level performance in heterogeneous environments, while full integration with tracking pipelines is left for future work. Importantly, the present study evaluates the estimator strictly at the trajectory level using pre-associated tracks, and does not assess end-to-end multi-object tracking performance, which depends on data association and track management components. Furthermore, all improvements are validated against both static and adaptive baselines under identical conditions, confirming that the gains arise from learned context-dependent adaptation rather than improved parameter tuning alone.</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

CAT-GAN-UKF: category-aware online adaptive unscented kalman filtering for trajectory-level multi-object state estimation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月11日

查看全文
  • Lior Tobaly,
  • Eyal Yaniv,
  • Zeev Zalevsky

摘要

State estimation in multi-object sensing environments is commonly performed using per-trajectory Bayesian filters, while upstream components such as detection, data association, and track management determine identity continuity and system-level performance. In practice, Unscented Kalman Filters (UKFs) are widely used for nonlinear estimation but rely on globally fixed uncertainty parameters, including process noise, measurement noise, and sigma-point scaling, despite strong variation across object categories and sensing conditions. In this work, we introduce Category-Aware Adaptive Unscented Kalman Filtering (CAT-GAN-UKF), a trajectory-level adaptive filtering framework in which uncertainty parameters are updated online using a category-conditioned learning module driven by innovation-based diagnostics and contextual cues. The proposed method preserves the recursive Bayesian structure of the UKF while enabling dynamic, context-dependent adjustment of uncertainty assumptions. Importantly, the evaluation in this study is conducted strictly at the trajectory level using pre-associated object-centric sequences derived from a large-scale autonomous-driving dataset. A leakage-free experimental protocol is employed, in which models are trained on past scenes and evaluated on future scenes, ensuring that adaptation policies generalize across environments. The proposed approach is evaluated against a broad set of classical and learning-based baselines, including EKF, PF, UKF-IAE, KalmanNet, and LSTM-enhanced filters, under identical trajectory-centric conditions. Additional analyses include Q/R sensitivity experiments, Monte Carlo robustness evaluation under measurement perturbations and outliers, and empirical uncertainty consistency metrics. Across diverse object categories and sensing regimes, CAT-GAN-UKF demonstrates consistent improvements in trajectory-level estimation accuracy and improved consistency between predicted uncertainty and realized estimation error. These results establish that category-conditioned online adaptation can significantly enhance estimator-level performance in heterogeneous environments, while full integration with tracking pipelines is left for future work. Importantly, the present study evaluates the estimator strictly at the trajectory level using pre-associated tracks, and does not assess end-to-end multi-object tracking performance, which depends on data association and track management components. Furthermore, all improvements are validated against both static and adaptive baselines under identical conditions, confirming that the gains arise from learned context-dependent adaptation rather than improved parameter tuning alone.