ABSTRACT

Shadows have long been a challenging topic for computer vision. This challenge is made even harder when we assume that the camera is moving, as many existing shadow detection techniques require the creation and maintenance of a background model. This article explores the problem of shadow modelling from a moving viewpoint (assumed to be a robotic platform) through comparing shadow-variant and shadow-invariant image features — primarily color, texture and edge-based features. These features are then embedded in a segmentation pipeline that provides predictions on shadow status, using minimal temporal context. We also release a public dataset of shadow-related image sequences, to help other researchers further develop shadow detection methods and to enable benchmarking of techniques.

摘要

长期以来，阴影一直是计算机视觉的一个具有挑战性的话题。当我们假设相机正在移动时，这一挑战就变得更加艰巨，因为许多现有的阴影检测技术都需要创建和维护背景模型。本文通过比较阴影变化和阴影不变的图像特征（主要是颜色，纹理和基于边缘的特征），从移动的角度（假定是机器人平台）探讨了阴影建模的问题。然后将这些功能嵌入到分段流水线中，该流水线使用最小的时间上下文提供有关阴影状态的预测。我们还发布了与阴影相关的图像序列的公共数据集，以帮助其他研究人员进一步开发阴影检测方法并实现技术基准测试。