The School of Computer Science is pleased to present…
Date: Monday, May 12, 2025
Time: 1:00 PM
Location: Essex Hall, Room 122
Obstacle and change detection using RGB cameras is implemented with multiple training videos of obstacle-free paths, considering surroundings to detect obstacles and changes on test paths. This study presents a dual approach combining geometric and machine learning approaches with a parallel processing architecture to efficiently reduce dataset size by removing similar frames. The approach is robust to scalability, lighting, and orientation variations through augmented datasets. In the geometric approach, features extracted by SIFT, BRISK, and ORB are matched using BF and FLANN matches and filtered via RANSAC Homography estimation. The machine learning approach employs CNN models, such as MobileNetV3-large, EfficientNetB3, ResNet50, DenseNet121, and their feature-level ensembled model to extract features of the dataset, which are matched using cosine similarity. Performance metrics include accuracy, precision, recall, f1-score, specificity, speed (training/detection time), and resource usage (CPU/memory). Geometric ORB with BF matcher achieves the highest performance with the fastest detection time, while MobileNetV3-Large achieves the highest performance with a longer detection time. In this study, the geometric approach outperforms the machine learning approach in performance and speed, demonstrating practical applicability for autonomous navigation and future enhancement potential.
Internal Reader: Dr. Imran Ahmad
External Reader: Dr. Mohamed Belalia
Advisor: Dr. Boubakeur Boufama
Chair: Dr. Pooya Moradian Zadeh
