基于图像增强的改进YOLO11低光照车辆行人检测算法

Improved YOLO11-Based Low-Light Vehicle and Pedestrian Detection Algorithm Using Image Enhancement

摘要: 为提高低光照环境下车辆与行人目标的检测精度，本文提出了一种融合图像增强与改进YOLO11的兼顾检测精度与计算效率的低光照车辆行人目标检测算法。首先，引入HVI-CIDNet对原始低光照图像进行增强，以有效恢复图像的光照与结构信息。其次，在YOLO11n基础上进行了两方面改进：一是于骨干网络中引入全局边缘信息传递模块，提取边缘特征并将其有效融入主干特征表示；二是对检测头结构进行重构，并引入PConv，在降低模型参数量的同时保持检测性能。此外，采用EMASlideLoss损失函数，通过对不同难度目标分配差异化权重，缓解低光照交通场景下样本分布不均衡问题。实验结果表明，与未引入图像增强的YOLO11n相比，本文方法在精确率、召回率和mAP50上分别提升2.89%、3.03%和2.83%，且模型参数量仅小幅增加，计算复杂度与原YOLO11n基本保持一致。在所选的多种轻量级目标检测算法对比中，本文方法mAP50最高，充分验证了其在低光照环境下的检测优势，为低光照交通场景下的车辆与行人目标检测提供了一种参考。

Abstract: To improve the detection accuracy of vehicles and pedestrians under low-light conditions, this paper proposes a low-light vehicle and pedestrian detection algorithm that balances detection accuracy and computational efficiency by integrating image enhancement with an improved YOLO11 model. First, HVI-CIDNet is employed to enhance the original low-light images, effectively restoring illumination and structural information. Subsequently, two improvements are introduced based on YOLO11n. A global edge information propagation module is incorporated into the backbone network to extract edge features and effectively integrate them into the backbone feature representation. In addition, the detection head structure is redesigned and PConv is introduced, enabling a reduction in model parameters while maintaining detection performance. Furthermore, the EMASlideLoss loss function is adopted to assign differentiated weights to targets of varying difficulty, alleviating the issue of sample distribution imbalance in low-light traffic scenarios. Experimental results demonstrate that, compared with YOLO11n without image enhancement, the proposed method achieves improvements of 2.89%, 3.03%, and 2.83% in precision, recall, and mAP50, respectively, while only slightly increasing the number of model parameters and maintaining computational complexity comparable to the original YOLO11n. Among several selected lightweight object detection algorithms, the proposed method achieves the highest mAP50, fully validating its superiority in low-light environments and providing a valuable reference for vehicle and pedestrian detection in low-light traffic scenarios.