PERCEPTION SYSTEM DESIGN FOR MULTIPLE OBJECT TRACKING IN AUTONOMOUS VEHICLES

• Main Author: Abel Oktavianus, Joshua
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/49919
• Institution: Institut Teknologi Bandung

Autonomous vehicles, right now, is undergoing a rapid development by reputable automotive industries and notable universities all around the world. This is mainly caused by the potential of autonomous vehicles to provide safer driving experience than the normal vehicles which are driven by a person. Indonesia’s new capital city plan establishes higher opportunity in the development of autonomous vehicles by building the infrastructures specifically for autonomous vehicles and electric vehicles. To further increase the safety of autonomous vehicles, a reliable perception system is neccessary. One of the most important tasks in perception system is object tracking. Object tracking extracts important information such as position, and velocity of multiple objects which surround the autonomous vehicle before sending it to the motion planner for determining the most optimal maneuver. Object tracking gives rise to a challenge, that is, performing measurements without placing any sensors to objects surrounding the autonomous vehicle. This research proposes a perception system design for multiple objects tracking with 3 components as the basis: object detection and classification, object matching, and state estimation. In this design, You Only Look Once (YOLO) is used as the object detector, Feedforward Neural Network, and Oriented FAST and Rotated BRIEF (ORB) serves as the basis for object matching, and last Extended Kalman Filter (EKF) is used for state estimation. This design has been proven to complete object tracking with only 0,15330 meter (x-axis), and 0,48577 meter (zaxis) mean absolute error in an experiment ranging from -2 to 2 meter (x-axis), and 4,55 to 9,55 meter (z-axis).