HUMAN RE-IDENTIFICATION SYSTEM USING FACE BIOMETRIC FEATURE

• Main Author: Gunawan, Agus
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39317
• Institution: Institut Teknologi Bandung

The human re-identification system is a system that can re-identify a human who has appeared on one of the cameras connected to the system. This system has three main components, namely human detection, tracking, and re-identification. This system has previously been made but has several disadvantages, such as failing to detect people with poses that are not ideal and obstructed, using tracking components that are not optimal, and recognizing different people as the same person. The results of this final project are intended to overcome those weaknesses. Therefore, the objectives of this thesis research are: 1) building a tracking component that can be used to generate keyframes, 2) looking for the most optimal facial biometric features for the re-identification component, 3) building a human re-identification system using those components and detection component with a deep learning approach, then 4) measure the quantitative performance of the human re-identification system being built. The development of the system begins with building a detection component with a YOLOv3 deep learning model. After that, the development continued with a tracking component consisting of two processes, namely mapping the frame with a specific identity and extracting the keyframe from the resulting frames. Mapping frames with certain identities is done by the DeepSort method, while keyframe extraction is done by using the face tilt angle as a determinant of a frame to be categorized as a keyframe. The keyframe extraction process begins with face detection using the MTCNN technique which produces the best accuracy of 99.41% from the experimental results and can be used to determine the face tilt angle based on the facial landmark extraction. The next step is to build the re-identification component using the VGG-Face ResNet50 technique which produces the best accuracy of 87.5% from the results of face recognition experiments based on feature extraction. The system that was successfully built with each component is evaluated using the evaluation dataset resulting an accuracy of 97.14% and succeeded in resolving any weaknesses that existed in the previous system.