DEVELOPMENT OF A FACE RECOGNITION SYSTEM AS A PERSON RECOGNITION SYSTEM ON A ROBOT IN THE FORM OF ROGA MASCOT

• Main Author: Ahmad, Farrel
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85059
• Institution: Institut Teknologi Bandung

Roga is a mascot of Bandung Institute of Technology with the abbreviation of Robot Gajah (Elephant Robot). The capstone project was initialized to create Roga as a real robot in the form of interactive application on a static robot with the functionality as information service center. Roga interactive application had the capability to see, hear, and talk thus there were vision module, interaction module, and location module. A system was needed to know who Roga was talking to during interaction with humans. Vision module as the robot’s computer vision was able to see human with their face as the most apparent physical features. According to this, a face recognition system was used to recognize a person based on their face. Face recognition system consisted of some program components in a unity. The problem to be solved was the development of a face recognition system that is fast, capable of training during runtime, and capable of prediction during runtime. Development was done using ROS (Robot Operating System) that consisted of three main programs, which were camera node, face trainer node, and face recognizer node. There were three face recognition models that were tested, which were LBPH, SFace, and VGG-Face. The experiment was done and measured using standard metric of IEEE 2945-2023. There were 13,724 face images that were spread into 5,754 unique identities as testing data. According to the test, SFace had the best performance between the other two models with the average score of FAR = 0.00360, FRR = 0.07480, face verification response time = 0.03857 s, FPIR = 0.003, FNIR = 0.293, and face identification response time = 1.036 s. The output of face recognition was an ID number that must be sent to the interaction module. This must be done as fast as possible using MQTT (Message Queuing Telemetry Transport) protocol based on the lowest end-to-end latency. The lowest latency was done by own MQTT broker nearest to the robot’s location in Bandung, that was AWS (Amazon Web Services) in Jakarta with average latency of 35.7 ms.