Developing an algorithm for person re-identification using attribute recognition

This thesis aims to develop a tracking system that integrates person tracking and reidentification processes, which play an important role in security and forensic investigations. Traditional surveillance systems face difficulties in manual reviews due to extensive camera networks, which increases the need for automated tracking methods. However, existing tracking algorithms face challenges such as occlusions, identity changes, and distinguishing between individuals wearing similar clothing. In this context, the proposed system combines the Kalman filter, MOSSE tracker, and Mask R-CNN methods to perform validation through both motion prediction and visual attributes. The Kalman filter and MOSSE Tracker work as two components that validate each other, thereby ensuring that the tracking process is more stable and resilient to errors. Mask R-CNN is used to determine the initial positions of individuals and extract their visual attributes. In addition, people who meet the criteria determined based on the keywords provided by the user are tracked through the dynamic filtering mechanism. The system offers solutions to the problems encountered by traditional tracking algorithms, but it also has some limitations. Due to changes in lighting, differences in posture, or distortions in image quality, clothing detection may occasionally fail, and it can be difficult to distinguish individuals with similar attributes in crowded environments. Future studies should focus on more robust visual representations and adaptive learning mechanisms to overcome these limitations.