STATE MACHINE DESIGN AND IMPLEMENTATION FOR COMBINING OBJECT TRACKING METHOD: HUMAN CASE STUDY

• Main Author: Maria Teresa R Kinasih, Fabiola
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/36571
• Institution: Institut Teknologi Bandung

Since the Viola and Jones' method on real-time face detection was proposed in 2001, numerous works for object detection, person recognition, and object tracking have been published by papers and journals. Each method has its own strong points and drawbacks. For this thesis will focus to build a system with a goal to track the specific object of interest, in this case, person, it is beneficial to combine those methods using state machine in order to harness the tracker promptness while maintaining the ability to distinguish the object of interest with the other object and backgrounds. Several methods that are considered to be combined are: Face Recognition which is a derivation works from Dlib machine learning toolkit, Face Detection which employed convolutional neural network with Mobilenet-SSD architecture, kernelized correlation filter based object tracker, and a primitive object tracker based on color filter. The FSM implemented in this paper is able to meet the goal with a considerable performance for indoor settings. System performance is between 8-30 frame per seconds, depends on which method that is currently being run, while being able to recognize the object. Coordinate point accuration is about 93-100% for horizontal coordinate and 97-100% for vertical coordinat. No false positive yielded when testing, while false negative could happen for under exposure frame captured. The false negative rate for indoor settings is recorded at 5-10%.