GLD-YOLO: an improved YOLO11n-based object detection method for complex traffic scenarios
TL;DR
To address detection degradation caused by illumination variation, complex backgrounds, scale variation, and dense occlusion in traffic scenes, the proposed Global-Local-Dynamic YOLO (GLD-YOLO), an improved YOLO11n-based detector is proposed, demonstrating competitive inference efficiency on the evaluated platform.
Abstract
To address detection degradation caused by illumination variation, complex backgrounds, scale variation, and dense occlusion in traffic scenes, we propose Global-Local-Dynamic YOLO (GLD-YOLO), an improved YOLO11n-based detector. C3k2_LIA and C2PSA_DLKA are integrated into shallow and deep backbone stages for local key-feature selection and contextual modeling, DySample is used in the neck for content-adaptive feature alignment, and Inner-WIoUv3 is adopted for bounding-box regression. Over five independent seeds on the main Traffix Vehicles dataset, GLD-YOLO achieves 83.90±0.85% precision, 69.42±1.54% recall, 78.56±0.40% mAP@0.5, and 54.64±0.73% mAP@0.5:0.95, corresponding to mean gains of 4.70, 0.24, 1.98, and 1.14 percentage points over YOLO11n. Direct five-seed evaluation on Cars Detection and an expanded 3004-image Road Vehicle Images set yields positive paired mAP gains on both external datasets, although the magnitude depends on the distribution shift. GLD-YOLO achieves 88.51 FPS under a unified batch-1 benchmark on the tested RTX 5070 Ti Laptop GPU, demonstrating competitive inference efficiency on the evaluated platform.