IMPLEMENTATION OF DETECTOR-BASED AND DETECTOR­ FREE FEATURE MATCHING ALGORITHMS IN MULTI­ CAMERA OBJECT TRACKING APPLICATION DEVELOPMENT

• Main Author: Kinaya Maureen Sibayak, Nissha
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/82366
• Institution: Institut Teknologi Bandung

Smart surveillance is a crucial need for security across various sectors, including public environments, offices, and industrial facilities. Traditional surveillance systems often face limitations in area coverage and detection accuracy. Additionally, single-camera systems cannot provide a comprehensive view and often have blind spots. In this final project, a multi-camera object tracking application was developed for smart surveillance. The focus of this project is the development of a multi-camera object tracking application using both detector­ based and detector-free feature matching algorithms. The aim of this project is to develop a multi-camera object tracking application that can effectively and accurately integrate data from multiple cameras using detector-based feature matching algorithms, such as SIFT and Brute-Force matcher, and detector-free algorithms, such as LoFTR. The methodology used in this research involves several stages,from data collection, feature matching, to data integration from multiple cameras. The research results indicate that detector-free feature matching algorithms provide better performance compared to detector-based feature matching algorithms in terms of matching accuracy and homography stability. The LoFTR algorithm demonstrates higher homography stability with lower mean absolute error and standard deviation values compared to SIFT Further analysis reveals that the developed system experiences performance degradation at extreme angles and distances, particularly noticeable at angles of 75° and 90°, and distances over 3 meters, where feature matching accuracy and homography stability significantly decrease . Recommendations for future research include the development of more advanced feature matching algorithms, testing in more dynamic environments, and optimizing system performance to improve scalability and processing speed. The use of deep learning algorithms like LoFTR shows great potential in enhancing the performance of surveillance systems, opening opportunities for broader applications in the future.