Context-based Indoor Object Detection

鲁棒高效的室内物体检测可以帮助重度视力障碍患者独立进入陌生的室内环境。然而，大多数现有的物体检测方法都是针对特定类型的物体（例如人脸）或一般的自然物体（例如建筑物、天空等）开发的，由于以下挑战，这些物体不能直接应用于室内物体：1)不同室内环境中物体模型的类间变化大，2)不同物体模型的类内变化小，3)与自然场景或室外环境中的物体相比，纹理较少，4)由于遮挡或用户失明，只能捕获部分物体，5)由于用户与物体之间的位置和距离变化导致物体的视图和比例变化，6)缺乏合适的数据库。这个EAGER项目是通过结合标志（文本和图标）和其他视觉线索（如浴室和电梯楼层号码）的上下文信息来探索检测室内物体的新方法。该研究通过结合上下文信息丰富了目标检测的研究，并对现有方法进行了重大改进。本项目开发的方法为盲人和视障人士进入陌生的室内环境提供了新的策略和技术。该研究还将为视频监控、智能会议室、视频索引和人机交互等许多重要研究领域带来益处。

Robust and efficient indoor object detection can help people with severe vision impairment to independently access unfamiliar indoor environments. However, most existing object detection methods are developed either for a specific type of object (e.g. face) or for general nature objects (e.g., building, sky, etc) which cannot be directly applied to indoor objects due to following challenges: 1) big inter-class variations of the object model among different indoor environments, 2) small intra-class variations of different object models, 3) less texture compared to objects in natural scene or outdoor environments, 4) only part of the object is captured due to occlusions or blind user, 5) view and scale variations of the objects caused by the position and distance change between the user and the object, and 6) lack of suitable databases.This EAGER project is to explore new methods to detect indoor objects by incorporating the context information from signs (both text and iconic) and other visual clues such as signage of bathrooms and elevator floor numbers. The research enriches the study of object detection by incorporating context information, and leads to significant improvements over existing methods. The methods developed in this project provide new strategies and technologies for the blind and visually impaired to access unfamiliar indoor environments. The research also benefits many important research areas including video surveillance, intelligent conference rooms, video indexing, and human-computer interactions.