MSc Thesis Proposal by: Subah Ibnat Karim

Obstacle and change detection using a single video of change and obstacle free paths is explored to detect any changes that occur while traveling the same paths in the future in this study. This approach starts from learning the background model of the given path as preprocessing, detecting changes starting from the first frame till last without redundancy and determining the current location in the path by frame number. The main purpose of this study is to improve the efficiency of obstacle and change detection for autonomous navigation systems while maintaining high matching score, particularly in familiar environments with reduced size of dataset. Two different approaches: geometry-based and machine learning, are implemented and compared to enhance obstacle and change detection techniques using RGB cameras. In the geometry-based approach, techniques such as SIFT, ORB, and BRISK are utilized for feature extraction from the frames, generated from the single video. Moreover, Brute Force and FLANN matchers are used for feature matching to remove the similar featured frames for reducing the size of dataset and to detect changes in test image by comparing its features with filtered frames. Feature matches undergo geometric verification using homography estimation with RANSAC homography to filter out spurious matches. In the machine learning approach, a pre-trained model such as the ResNet50 model employs transfer learning for feature extraction from the video frames. K-Means clustering, an unsupervised learning algorithm, is utilized to remove similar frames for dataset reduction, optimizing memory usage. Test image features are then matched with filtered frame features using cosine similarity. Performance evaluation includes best feature matching score between test image and best matched frame by comparing matching score between test image feature with each frame features. Moreover, time complexity and memory usage are evaluated by measuring feature extraction time, feature matching time, CPU usage, and memory usage. Comparing the results for both approaches indicate that geometry-based approaches perform competitively with machine learning methods, with the geometry-based approach particularly standing out with higher matching scores. This underscores the significance of geometry-based techniques, especially in scenarios where reducing the dataset is crucial for obstacle and change detection using RGB Cameras.

 

Keywords: Feature extraction, Feature matching, Geometric verification, Transfer learning, Unsupervised learning

Thesis Committee:

Internal Reader: Dr. Imran Ahmad          

External Reader: Dr. Mohamed Belalia  

Advisor: Dr. Boubakeur Boufama