An Adaptive Physics-Guided Ensemble Learning Framework for Dynamic Object Detection

Changqiang Wang

doi:10.1109/jstars.2026.3712395

An Adaptive Physics-Guided Ensemble Learning Framework for Dynamic Object Detection

2026-07-10

0

OA

AI

DOI：10.1109/jstars.2026.3712395

Original

Original request for help

Save

Abstract

En 中文

Traditional machine-learning models in dynamic object detection often overlook the temporal continuity of moving objects, leading to unstable and inaccurate results. To address these limitations, this article develops an adaptive physics-guided ensemble learning framework for dynamic object detection. The proposed method first employs simple linear iterative clustering superpixel segmentation to preserve spatial structure and improve feature consistency. Based on the segmented regions, multidimensional spatiotemporal features are constructed by integrating optical-flow statistics, motion-consistency residuals, texture descriptors, geometric characteristics, and temporal statistical information. To enhance the physical interpretability of feature representations, an adaptive physics-guided calibrator is introduced. By evaluating the consistency between observed motion patterns and locally estimated rigid-body motion behavior, the calibrator dynamically adjusts feature weights and emphasizes physically meaningful motion cues while suppressing unstable observations and noise interference. The calibrated feature representations are subsequently utilized to train light gradient boosting machine, extreme gradient boosting, and support vector machine classifiers. To further improve robustness and generalization capability, an uncertainty-aware adaptive ensemble strategy is employed, where classifier predictive entropy is utilized to dynamically determine sample-level fusion weights and adaptively aggregate multimodel predictions. Experiments conducted on multiple sequences of the KT3DMoSeg dataset demonstrate that the proposed framework achieves dynamic object recall rates exceeding 80% while maintaining stable performance under illumination variations, background disturbances, and complex motion scenarios. The results indicate that adaptive physics-guided feature calibration effectively improves detection completeness, temporal consistency, and model interpretability, providing a practical and computationally efficient solution for dynamic object detection.

Keywords:

Dynamic object detection

machine learning

physics-guided ensemble learning

uncertainty-aware

An Adaptive Physics-Guided Ensemble Learning Framework for Dynamic Object Detection

Abstract

Journal

Organization

Cited Papers