Date of Award
5-28-2025
Publication Type
Thesis
Degree Name
M.Sc.
Department
Computer Science
Keywords
RGB camera; Obstacle-free paths; Environmental factor; Threadpooling architecture
Supervisor
Boubakeur Boufama
Rights
info:eu-repo/semantics/embargoedAccess
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.
Abstract
Obstacle and change detection using RGB camera is implemented with 48 training videos of obstacle-free paths, considering the surroundings as part of the paths to detect any ob- stacles and changes on the test image paths in this study. This study introduces a dual- approach framework combining geometric and machine learning based approaches using par- allel Threadpooling architecture to detect obstacles and changes while reducing the training dataset by removing similar frames. This method is robust and invariant to scalability, lighting, and orientation changes through the augmented and reduced datasets via parallel threadpooling. The main purpose of this study is to implement this dual approach using RGB cameras and threadpooling for higher efficiency, speed, and resource usage, ensuring robustness against environmental factor variations to detect test image paths for obstacles and changes. In geometric based approach, features are extracted using SIFT, BRISK, and ORB, matched using BF matcher and FLANN based matcher, and filtered through RANSAC Homography estimation in threadpooling architecture for the dataset reduction and detec- tion. In machine learning based approach, features are extracted using four pre-trained CNN models such as MobileNetV3-large, EfficientNetB3, ResNet50, and DenseNet121, and their feature-level ensembled model, matched using cosine similarity in threadpooling architecture for dataset reduction and detection. The performance is evaluated by accuracy, precision, recall, F1-score, specificity, speed by training time and detection time, and the resource us- age by cpu usage and memory usage in this study. The geometric based ORB, along with the BF matcher, achieves 100% performance in only 0.32 seconds with an overall accuracy for the geometric based approach of 95.8%. Machine learning based MobileNetV3-Large achieves 100% performance in 2.25 seconds with an overall accuracy for machine learning based approach of 92.5%. According to this, geometric based approach works better than machine learning approach with higher performance and speed. Due to invariance to image size, scalability, lighting, orientation, and environmental factor changes, this study method is suitable for practical use in autonomous systems, along with providing a foundation for future enhancements in obstacle detection and classification and feature-based analysis using RGB cameras.
Recommended Citation
Karim, Subah Ibnat, "Obstacle and Change Detection Using RGB Cameras" (2025). Electronic Theses and Dissertations. 9739.
https://scholar.uwindsor.ca/etd/9739