YOLO-PBE: An Improved YOLOv11 Vehicle Detection Algorithm for Complex Traffic Scenes

Addressing the two critical challenges of missed detection of distant small targets and difficulty in identifying occluded targets under complex road conditions, this paper proposes YOLO-PBE, an improved high-precision vehicle detection model based on YOLOv11n. First, to tackle the fine-grained feature loss caused by conventional strided convolutions during downsampling, we add a high-resolution P2 detection layer and introduce SPD-Conv, a lossless spatial-to-depth feature transformation technique, for feature extraction. By preserving complete pixel-level information, the model's perception accuracy for distant small vehicles is enhanced. For feature fusion, we design an improved BiFPN incorporating a Ghost module. Specifically, by pruning redundant fusion nodes, replacing computationally expensive operations with inexpensive linear transformations, and employing an adaptive weighted fusion mechanism, the model optimizes feature fusion efficiency while ensuring precise semantic alignment across multi-scale features. Finally, we construct a deeply embedded C2f-EMA module that seamlessly incorporates an efficient multi-scale attention mechanism into the C2f unit. Through cross-dimensional feature interaction and pixel-wise reweighting, the module effectively suppresses background noise interference, strengthens the reconstruction ability of damaged features, and improves robustness in detecting occluded and overlapping vehicles. Experimental results demonstrate that compared with YOLOv11n, the mean Average Precision (mAP@0.5) improves by 5.8 and 4.2 percentage points on the UA-DETRAC and KITTI datasets, respectively. This study provides a new technical approach for vehicle detection in complex traffic scenarios.