Garcia-Jaimes Keyla, Ballesteros John R, Branch-Bedoya John W
Departamento de Ciencias de la Computación y de la Decisión, Facultad de Minas, Universidad Nacional de Colombia, Medellín 050041, Colombia.
Data Brief. 2025 Jul 20;62:111912. doi: 10.1016/j.dib.2025.111912. eCollection 2025 Oct.
Street vending is a prominent component of the informal economy, yet its prevalence remains poorly quantified due to the limitations of traditional survey methods, which are costly, invasive, and labor-intensive. To enable scalable, image-based assessments of this activity, we present the StreetVendor-SLI dataset, specifically designed for detecting vendors in urban environments. The dataset comprises 2794 high-resolution images (2416×1359 px), obtained from video footage recorded with a user grade camera mounted on a motorcycle. The original dataset contains 1397 images, with an average size of 5 MB per image, resulting in a total dataset size of 4.63 GB. Privacy compliance with GDPR guidelines was achieved by anonymizing pedestrian faces and vehicle license plates using an open-source YOLO object detection pipeline. Every image is annotated utilizing the YOLO format, with vendors enclosed in bounding boxes and classified into three categories: fixed-stall vendor (1774 labels), semi-fixed vendor (459 labels), and itinerant vendor (124 labels). To address class imbalance and enhance model generalization, data augmentation techniques-including geometric transformations (rotation, flipping, scaling, shearing) and spectral adjustments (brightness, contrast, hue)-were applied. The Steet-level Imagery dataset thus provides an openly available option for the detection of street vendors, offering a valuable resource for researchers studying informal economic activities and urban policies.
街头小贩是非正规经济的一个突出组成部分，但由于传统调查方法存在成本高、侵入性强和劳动密集型等局限性，其普遍程度仍难以量化。为了实现对这一活动的可扩展的基于图像的评估，我们展示了StreetVendor-SLI数据集，该数据集是专门为在城市环境中检测小贩而设计的。该数据集包含2794张高分辨率图像（2416×1359像素），这些图像来自安装在摩托车上的用户级相机录制的视频片段。原始数据集包含1397张图像，每张图像平均大小为5MB，数据集总大小为4.63GB。通过使用开源的YOLO目标检测管道对行人面部和车辆牌照进行匿名化处理，实现了符合GDPR准则的隐私保护。每张图像都使用YOLO格式进行标注，小贩被包围在边界框内，并分为三类：固定摊位小贩（有1774个标注）、半固定小贩（有459个标注）和流动小贩（有124个标注）。为了解决类别不平衡问题并增强模型的泛化能力，应用了数据增强技术，包括几何变换（旋转、翻转、缩放、剪切）和光谱调整（亮度、对比度、色调）。因此，街道级图像数据集为街头小贩的检测提供了一个公开可用的选项，为研究非正规经济活动和城市政策的研究人员提供了宝贵的资源。
