- 15 views
- 42 downloads
Spatial Data Analytics for Pattern Recognition of In-Situ and Wide-Area Contexts
-
- Author / Creator
- Lugo Bustillo, Gabriel A
-
Pattern recognition aims to differentiate patterns and regularities across diverse data types. Pattern recognition can identify unfamiliar objects, localize objects from various perspectives or resolutions, and infer patterns even when they are partially occluded. The ubiquity of sensors in various applications has made this topic a vital component of modern technology. Challenges encompass spatial variability, partial occlusion, data heterogeneity, dynamic environments, limited training data, noise and uncertainty, generalization issues, and performance constraints. Addressing these challenges is crucial for advancing the effectiveness of pattern recognition. This thesis focuses on exploring different data modalities and proposing solutions, with experimental validations, to solve pattern recognition problem for visual analysis, positioning, and pose estimation, particularly within the context of spatial and geospatial data. Real-world scenarios, ranging from local-area in-situ sensing to wide area monitoring (WAM) remote sensing, serve as the testing grounds for this research. The proposed novel solutions tackle unique challenges posed by these domains.
In the first scenario, we focus on 3D object recognition and 6D pose estimation of texture-less objects. These are crucial and fundamental endeavors for industrial assembly line automation. While the problem of textured objects has been extensively studied, there is still an open research topic for texture-less industrial parts, which are symmetric, causing ambiguity. In this scenario, we propose a novel object localization and pose estimation technique using RGB images and depth maps of industrial assembly parts. The proposed segmentation model is fully morphological and unsupervised for localizing the region of interest containing the target object extracted from the depth map. For object classification and pose estimation, we combine descriptors with dynamic time warping. We demonstrate how synthetic training images generated from Computer-Aided Design (CAD) models can facilitate pattern recognition.
In the second scenario, this thesis focuses on pattern recognition from 2D videos. The proposed approach is designed for the study of vehicles, with a primary focus on enhancing the assessment of goods and their value. Traditional approaches for vehicle classification often have relied on manual observation and limited sensor data, which have posed challenges in accuracy and scalability. Our proposed closed-loop system integrates deep learning and computer vision to detect, track, count, timestamp, and estimate the direction of vehicles, laying the groundwork for in-depth traffic flow analysis. The framework incorporates a unique data processing mechanism within a crowdsourcing environment, enhancing the scalability of our system. Experimental results demonstrate the effectiveness, efficiency, and robustness of the proposed system on challenging scenes, and adaptability with active learning for vehicular analysis, where the model will improve over time based on new data.
In the third scenario within the scope of WAM, our primary focus is to extract linear features from aerial images acquired by Unmanned Aerial Vehicle (UAV) for land surveying (LS). We observe that aerial photography can provide more precise geospatial and rich semantic information over large areas than conventional on-site surveying methods. However, drone captured data inherits imperfections when it is used to build and create 2D/3D maps of a physical scene. In particular, linear features are often affected in the pixel domain by factors such as complex backgrounds, different levels of occlusion, and lighting changes. In this scenario, we propose a framework for automatic surveying of road markings using drone imagery. We propose a semantic segmentation technique and a refining stage to enhance predicted masks for line connectivity.
In the fourth and fifth scenarios under wide-area remote sensing, the thesis focuses on pattern recognition of LiDAR. We introduce a novel 3D benchmark, "LiSurveying", which is a large-scale point-cloud dataset with over a billion points and uncommon urban object categories in complex outdoor environments. We propose an automatic and effective object detection and key-point feature detection pipeline on dense point-cloud scenes. The proposed method consists of a multiscale voxelization strategy to reduce the computational load and complexity of dense point clouds. Hierarchical features are then extracted to localize the objects of interest. Consequently, we propose an automatic strategy to locate the object's centroid point using a learning-based method with a space-partitioning data structure stage.
The contributions of this thesis encompass algorithm formulation. Whereas conventional pattern recognition techniques exhibit limitations preventing them from adequately addressing the aforementioned challenges, we introduce methodologies applicable in diverse scenarios.
-
- Subjects / Keywords
-
- Graduation date
- Spring 2024
-
- Type of Item
- Thesis
-
- Degree
- Doctor of Philosophy
-
- License
- This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.