BEGIN:VCALENDAR METHOD:PUBLISH PRODID:-//Apple Computer\, Inc//iCal 1.0//EN X-WR-CALNAME;VALUE=TEXT:USC VERSION:2.0 BEGIN:VEVENT DESCRIPTION:Speaker: Preetham Aghalaya Manjunatha , Astani Department of Civil and Environmental Engineering Talk Title: Vision-Based and Data-driven Analytical and Experimental Studies into Condition Assessment and Change Detection of Evolving Civil, Mechanical and Aerospace Infrastructures Abstract: Civil, mechanical, and aerospace infrastructures are subjected to applied loads and environmental forces like earthquakes, wind, and water waves in their operating lifespan. These factors will slowly deteriorate the structures during their service period, and often subtle observations of substantial damages are challenging. Due to the cost-effectiveness of high-resolution color, depth cameras, location sensors, and Micro Aerial Vehicles (MAVs), image processing, computer vision, and robotics techniques are gaining interest in Non-Destructive Testing (NDT) and condition assessment of infrastructures. In this study, several promising vision-based and data-driven, automated, and semi-automated condition assessment techniques are proposed and evaluated to detect and quantify a class of problems under the umbrella of infrastructure condition assessment. \n \n A synthetic crack generation methodology is introduced to generate zero-labeled samples for training the classical classifiers. This classifier was tested on a real-world dataset using the gradient-based hierarchical hybrid Multi-scale Fractional Anisotropy Tensor (MFAT) filter to segment the cracks. The results demonstrate the promising capabilities of the proposed synthetic crack generation method. Furthermore, the textural noise suppression and refinement are carried out by using an anisotropic diffusion filter. Guidelines are provided to select the parameters for the anisotropic diffusion filter. Further, this study presents the semantic segmentation of the cracks on concrete surface images using a deep Convolutional Neural Network (CNN) that has fewer parameters to learn. Several illustrative examples are presented to demonstrate the capabilities of the CNN-based crack segmentation procedure. The CNN was tested on the four real-world datasets, and the results show the proposed CNN's superiority against four state-of-the-art methods.\n \n As a part of this study, an efficient and autonomous crack change detection, tracking, and evolution methodology is introduced. Among the image registration methods, feature-based registration is robust to the noise, intensity change, and partial affine motion model. This study uses an efficient $k$-d tree-based nearest neighbor search which is faster than the quadratic computational complexity of the current pairwise search. Furthermore, unlike other methods, the fixed camera assumption is relaxed in this study. Another significant contribution is a probabilistic measure of the reliability of the analysis results that can aid the prognosis damage detection models.\n \n After the nearest neighbor search, the SURF-based keypoints are extracted from the images in the previous database and the current one. This is followed by the Random Sample Consensus (RANSAC)-based outliers rejection, bundle adjustment to refine the homographies, gain exposure compensation and multi-band blending for the seamless registration images. Lastly, the registered image is compared to the current images for the change detection in crack physical properties. To demonstrate the capabilities of the proposed method, two datasets were utilized; a real-world dataset, and a synthetic dataset. The experimental results show that the performance of the proposed methodology is suitable for detecting the crack changes in two datasets.\n \n This work also studies the condition assessment of public sewer pipelines. The visual-bags-of-words model was evaluated for classifying the defective and non-defective sewer pipeline images using two feature descriptors. Three classical classifiers are trained and tested on a moderate-sized dataset of 14,404 images. The experimental results demonstrate that the classification accuracy of the visual-bags-of-words model is satisfactory and comparable to deep learning methods given the moderate dataset size.\n \n Lastly, defect detection of the three-dimensional surface of mechanical parts is studied. A preliminary study on a vision-based semi-autonomous spatio-temporal method to detect, locate and quantify the defects such as loose bolts, displacements, pipe chafing, or deformation is proposed. In addition, a probabilistic reliability quantification method based on the ensemble averaging of the Cloud-to-Cloud (C2C) distances is introduced for mechanical systems. Several quantitative and qualitative examples are presented to illustrate the capabilities of the proposed method. The results show that the proposed method is promising and robust to register the complex shapes, and detect and locate the changes in the mechanical systems.\n \n \n \n \n \n \n \n \n Webcast: https://usc.zoom.us/j/98823087239? Meeting ID: 988 2308 7239 Passcode: 729479 SEQUENCE:5 DTSTART:20211213T160000 LOCATION: DTSTAMP:20211213T160000 SUMMARY:Astani Department of Civil and Environmental Engineering Ph.D. Dissertation UID:EC9439B1-FF65-11D6-9973-003065F99D04 DTEND:20211213T180000 END:VEVENT END:VCALENDAR