Complex highway vehicle detection algorithm based on improved YOLOv11

Abstract

To address missed detections and truncations caused by extreme scale variations, target fragmentation, and background interference in highway monitoring, a high-precision vehicle detection algorithm based on improved You Only Look Once (YOLO)v11n is proposed. First, Spatial and Channel Reconstruction Convolution (SCConv) is introduced to dynamically suppress background noise. Second, a 4× downsampled P2 high-resolution branch is cascaded to accurately anchor distant tinytargets. Finally, Bi-level Routing Attention (BiFormer) is embedded for global feature stitching of heavy long vehicles. Evaluated on a custom highway dataset, the model achieves 94.9% mean Average Precision (mAP)@0.5 and 77.1% mAP@0.5:0.95 with only 2.69M parameters, outperforming the baseline by 1.7 and 2.2 percentage points, respectively. Compared with mainstream models like YOLOv12n, the proposed algorithm significantly enhances detection robustness and accuracy with low computational overhead, offering strong practical value for intelligent transportation.