Abstract: YOLO series object detection algorithms exist unbalanced accuracy and computational cost, as well as insufficient model generalization. To address these issues, this paper proposes a high accuracy and fast vehicle and pedestrian detection model based on YOLO, called YOLO-Day Night and Fast（YOLO-DNF）, which detects vehicle and pedestrian under different lighting scenarios. This paper analyses the effects of convolutional structure and network depth on the feature extraction capability and calculation cost of the backbone in relation to the convolutional neural networks used in the mainstream detection models today. Then this paper proposes an ACNet network with low computational cost and high feature extraction capability by selecting the convolutional structure arrow-block and CSP-Block for different levels of the network and determining the depth of the network by quantifying the computational cost of stacked units. In addition, the article analyses the difference in luminance between daytime and nighttime images and introduces a data enhancement strategy of HSV domain perturbation and luminance processing to enhance the detection accuracy of the model and improve the problem of insufficient generalization of the model. The experimental results show that the YOLO-DNF model achieves 32.8% detection accuracy of full time mAP at 24.36 frames per second after training in the training set of SODA10M dataset containing only daytime images, which exceeds the current mainstream detection models in terms of detection accuracy and speed. The nighttime accuracy reaches 27.7%, which improves the nighttime detection capability of the model and extends the detection application scenarios of the model.

Key words: object detection, convolutional neural network, data augmentation, YOLO

摘要： 为了解决YOLO系列目标检测算法存在的精度与计算成本不均衡、模型泛化性不足的问题，提出了可满足不同光照场景下目标检测需求的高精度快速的车辆与行人检测模型YOLO-Day Night and Fast（YOLO-DNF）。文中结合当下主流检测模型所使用的卷积神经网络分析卷积结构与网络深度对于主干网络特征提取能力和计算成本的影响，针对网络不同层次选取卷积结构Arrow-Block与CSP-Block搭建网络并通过量化堆叠单元的计算成本确定网络深度，提出低计算成本、高特征提取能力的ACNet网络。此外分析了白天与夜间图像的亮度差异，引入了HSV域扰动并提出亮度处理的数据增强策略，提升了模型的夜间检测精度，改善了模型泛化性不足的问题。实验结果表明：YOLO-DNF模型在SODA10M数据集仅含白天图像的训练集中训练后以每秒24.36帧的检测速率达到32.8%的全时段mAP检测精度，检测精度与速度超过目前主流检测模型。其中夜间精度达到了27.7%，扩展了模型的检测应用场景。

关键词: 目标检测, 卷积神经网络, 数据增强, YOLO