CEM-YOLO: multi-branch residual feature fusion and convolutional maxpooling downsampling for real-time vehicle detection in night scenarios

• Published in: Signal, image and video processing Vol. 19; no. 9
• Main Authors: Liu, Li-Juan, Jia, Rushi, Karimi, Hamid Reza
• Format: Journal Article
• Language: English
• Published: London Springer London 01.09.2025; Springer Nature B.V
• Subjects: Ablation; Computer Imaging; Computer Science; Feature extraction; Image Processing and Computer Vision; Machine learning; Modules; Multimedia Information Systems; Night; Object recognition; Original Paper; Pattern Recognition and Graphics; Real time; Redundancy; Signal,Image and Speech Processing; Vision; Deep learning; Convolutional Maxpooling Downsampling; Real-time vehicle detection; MobileViT; Multi-branch Residual Feature Fusion
• ISSN: 1863-1703; 1863-1711
• DOI: 10.1007/s11760-025-04313-2

Real-time vehicle detection at night faces significant challenges: ineffective low-light feature extraction, low computational efficiency hampering real-time performance, subpar detection of small and occluded vehicles, and limited cross-scenario generalization. To address these issues, this paper presents CEM-YOLO, an enhanced YOLO algorithm leveraging deep learning for nighttime vehicle detection. Two novel modules, Convolutional Maxpooling Downsampling and Multi-branch Residual Feature Fusion, are introduced to mitigate model complexity, reduce feature redundancy, and safeguard input features. Additionally, the Efficient Multi-Scale Attention Module is integrated into the Neck network’s detection layers. Extensive experiments and ablation studies on benchmark datasets demonstrate that CEM-YOLO excels in nighttime scenarios, achieving an optimal speed-accuracy balance for real-time applications.