Appearance-based material classification after occlusion removal for operation-level construction progress monitoring

Progress monitoring;Material Classification;Construction site

Muthukumar, Banu ; Golparvar-Fard ; Mani

关键词: Progress monitoring;    Material Classification;    Construction site;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/88097/MUTHUKUMAR-THESIS-2015.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Today, the availability of a large number of smart devices on construction sites, has significantly interest popularity of appearance-based methods for automated construction progress using site photographs monitoring. These methods, however, face a number of technical challenges that limit their applicability including low spatial resolution of images, and static anddynamic occlusions due to the construction progress and moving resources (equipment, workers,scaffolding, etc). To address these limitations, this paper extends on an existing model-driven appearance-based material classification method for appearance-based construction progressmonitoring using 4D BIM and site photologs. Specifically, it introduces a robust occlusion removal algorithm that can lower false positives in material recognition. The method leveragesthe depth information from the 4D BIM as well as the 3D point cloud created through Structure from Motion procedures. Once the occluded regions are removed, square-shape patches can be extracted from the back-projection of the BIM elements on site images. These improved image patches are then used in the material recognition pipeline to create a vector quantized histogramof all the material classes. The material class with the highest frequency is chosen as the material type for the element and this appearance information is used to infer the most updated state ofprogress for the elements. To validate, four existing incomplete and noisy point cloud models from real world construction site images and their corresponding BIMS were used. An extended version of the Construction Material Library (CML) developed at the University of Illinois at Urbana-Champaign’s Real-time and Automated Monitoring and Control (RAAMAC) lab was used to train the material classifiers and the experimental results shows an average accuracy of 90.9%. The occlusion removal and subsequent classification for the four datasets resulted in an accuracy of 92.2% compared to 89.9% in the existing method, demonstrating a definite improvement. By predicting the material present in an element, the status of that element can be identified as “in progress” or “completed’ and compared with the schedule. Since staticocclusions are detected, analyzed, and removed, this method has potential to be effective for appearance-based progress monitoring methods and can results in higher accuracy material classification.