Frequency-spatial decoupled state-space modeling with transformer fusion: a robust YOLOv11 framework for autonomous vehicle perception

Abstract

Visual environmental perception stands as the paramount safety cornerstone for autonomous driving. However, robust vehicle detection remains a formidable challenge, particularly under adverse conditions such as intense glare, severe occlusion, and inclement weather. While YOLOv11 offers an efficient baseline, its inherent limitations, including coupled frequency-spatial feature extraction and inadequate long-range dependency modeling, frequently lead to misdetections of occluded or distant targets. To surmount these critical bottlenecks, we propose a novel and collaboratively enhanced YOLOv11 framework. This framework seamlessly integrates three innovative modules: a Frequency-spatial decoupled feature pyramid network (FSD-FPN), a visual recursive state-space module (VRSSM), and a multi-scale attention convolution-transformer block (MACTB). Specifically, the FSD-FPN is meticulously designed to effectively disentangle high-frequency structural details from low-frequency semantic components, thereby significantly attenuating spectral noise induced by extreme illumination. Concurrently, the Mamba-based VRSSM establishes global contextual dependencies with remarkable linear complexity, crucially restoring feature integrity for heavily occluded vehicles. Furthermore, the MACTB synergistically combines the inductive biases of local convolutions with the global refinement capabilities of Transformers, substantially boosting multi-scale discriminative power. Comprehensive empirical evaluations were conducted on two distinct datasets: the widely recognized public BDD100K dataset and a meticulously self-curated Chengdu urban road dataset. Our framework consistently demonstrates superior performance, achieving a mean average precision (mAP)@0.50.95 of 70.9% and an mAP@0.5 of 93.0% on the combined validation set. These results unequivocally outperform YOLOv7, YOLOv10, the vanilla YOLOv11, and other state-of-the-art detectors, validating the efficacy and robustness of our proposed approach in complex autonomous driving scenarios.