雨雾环境下的行人检测基准

A Benchmark for Pedestrian Detection in Rain and Fog

摘要: 随着针对雨雾天气行人目标检测研究的深入，现有数据集的局限性日益凸显：一方面，现有真实数据集存在严重的标注偏差，仅标注清晰可见的行人目标而忽略模糊实例，导致模型评估结果虚高；另一方面，现有合成数据集虽样本数量多，但其模拟雾效与真实场景存在显著缺陷。为此，本研究重新界定检测目标与评估维度，并且构建了新型真实雨雾行人检测数据集，其包含了2400张图像，通过数据清洗与重新标注。重点优化了低分辨率、小目标及复杂场景下的标注质量，解决了现有公开数据集在雨雾场景下样本不足、标注不精确等问题。实验结果表明，该数据集能够有效评估行人检测模型在复杂气象条件下的检测性能，为雨雾天气行人目标检测任务提供了更贴近实际的实验基准。

Abstract: As research into pedestrian target detection in rainy and foggy weather has deepened, the limitations of existing datasets have become increasingly apparent: on the one hand, existing real-world datasets suffer from severe annotation bias, as they only label clearly visible pedestrian targets while ignoring blurry instances, leading to inflated model evaluation results; on the other hand, while existing synthetic datasets have a large number of samples, their simulated fog effects exhibit significant deficiencies compared to real-world scenarios. To address these issues, this study redefines the detection targets and evaluation criteria and constructs a new real-world rain and fog pedestrian detection dataset comprising 2400 images, achieved through data cleaning and re-labeling. The dataset prioritizes improving annotation quality in low-resolution, small-object, and complex-scenario conditions, addressing the limitations of existing public datasets, such as insufficient samples and imprecise annotations in rain and fog scenarios. Experimental results demonstrate that this dataset can effectively evaluate the detection performance of pedestrian detection models under complex weather conditions, providing a more realistic experimental benchmark for pedestrian target detection tasks in rain and fog weather conditions.