Hybrid-YOLO: Lightweight Mamba-Transformer Hybrid with multi-scale fusion for real-world traffic detection

Vehicle detection in complex traffic scenes remains challenging due to frequent occlusions, lighting variations, and extreme weather. We present Hybrid-YOLO, a real-time detection framework that unifies Mamba-based state space modeling, Transformer-driven global attention, and multi-scale feature fusion to achieve high accuracy at low computational cost. At its core, Hybrid-YOLO introduces a Dynamic Residual Stem (DR Stem) for adaptive feature calibration, a Hexa-Scan Selective Block (HSSBlock) for six-directional structural perception, and a Selective State Space Model (SSM) for efficient long-range dependency modeling. A Cross-Stage Scales Feature Extraction (CSSFE) module enriches spatial semantics for small-object detection, while a Sparse-Queries Cascade Self-Attention (SCS) module focuses computation on informative regions, enhancing robustness to clutter and background noise. Extensive experiments on KITTI, BDD100K, and IITM-HeTra show that Hybrid-YOLO achieves 90.11 mAP@0.5 at 66.3 FPS, surpassing state-of-the-art methods in both accuracy and efficiency, and offering a promising solution for real-world intelligent transportation systems.