CDS-YOLOv11:vehicle detection in low-light scenarios via a lightweight YOLOv11

Abstract In low-light nighttime conditions, drivers often struggle to accurately identify nearby vehicles, which can lead to traffic accidents. Concurrently, traditional detection methods face significant challenges in such dimly lit environments. To address the aforementioned issues, a nighttime vehicle detection model based on YOLOv11 is proposed. The framework employs the convolutional block attention module to enhance feature extraction and devises a depthwise separable dilated convolution module to capture richer contextual information for fusion, thereby reducing computational parameters and improving efficiency. Furthermore, a squeeze and excitation channel attention module is inserted to deliver higher quality features for neck fusion. The experimental evaluation demonstrates significant improvements of the proposed model over the original YOLOv11s baseline. The refined model demonstrates boosts of 4.0% in precision, 0.8% in recall, 1.8% in mAP@0.5, and 1.0% in mAP@[0.5:0.95]; crucially, its model size is merely 14.5MB, and it also achieves a notable 37.1% cut in computational load floating-point operations per second. The improved model boasts higher detection accuracy alongside significantly lower computational complexity and reduced model size, achieving a lightweight architectural design. It is also more suitable for deployment on edge devices.