FCD-Net: Feature Decorrelation and Confidence-Driven Dynamic Fusion for Robust Pedestrian Recognition in Autonomous Driving

Accurate pedestrian recognition is crucial for autonomous driving systems, but existing methods often fail in challenging scenarios due to feature redundancy and unreliable confidence estimation. While rich features are vital for capturing diverse pedestrian appearances, redundancy reduces representation diversity and generalization ability. Moreover, unreliable confidence estimation often results in a mismatch between predicted confidence scores and actual classification accuracy, which can lead to critical failures, such as missed detections or false positives, in dynamic environments. To address these issues, we propose FCD-Net, a novel framework combining feature decorrelation, confidence-driven dynamic fusion, and calibrated confidence estimation. Specifically, a Feature Decorrelation Loss (FDLoss) reduces redundancy to enhance feature diversity, while a Dynamic Fusion Module (DFM) adaptively integrates features based on classification confidence. Additionally, a Score-Accuracy Mapping (SAM) and Distribution Calibration Module (DCM) ensure reliable confidence estimation. Experiments on benchmark datasets (MIT, INRIA, Caltech) show that FCD-Net achieves up to a 5.39% improvement in F1-score on the Caltech dataset, maintaining real-time performance at 43.26 FPS. Deployment in real-world autonomous driving scenarios further demonstrates its effectiveness, particularly in handling occluded or distant pedestrians, where it significantly reduces missed detections and false positives.