Vision-Based Positioning Estimation on the ERSOW Robot Soccer by Utilizing Unique Landmarks in the Field with a Computational Process using GPU

摘要

Robot soccer is an autonomous robot capable of playing soccer. ERSOW is the name of a soccer robot that was researched by the Surabaya State Electronics Polytechnic which took part in the Indonesian Wheeled Soccer Robot Contest (KRSBI Wheeled). All the rules in the KRSBI Wheeled match follow the RoboCup rules. One of the abilities that a soccer robot must have is determining the position (x, y, θ) of the robot in the field or known as localization. Previously, the localization method on the ERSOW robot only used odometry sensors and IMU sensors, but this method has a high percentage of error due to the slip movement of the rotary encoder when the robot is maneuvering in the field so that the target coordinates that should be targeted are shifted. This shift is known as the odometric error shift. From these problems, we propose a Localization self-positioning estimation on the ERSOW robot while maneuvering in the field. Self-positioning estimation on the ERSOW Robot utilizes the vision camera sensor by estimating its position (x, y) against unique landmarks in the field. The unique landmarks that are used as references are the foot of the goalposts, the white field line in the x-axis direction, the white field line in the y-axis direction, and 2 points where the circular lines meet in the middle of the field. As for the heading position (θ) using the IMU sensor. The goalpost detection uses the tiny Yolov4-Radial Search Lines method and the line detection in the field uses the Radial Search Lines method. Robot Operating System (ROS) is used to process data for each robot’s work process. Vision data processing is handled directly by the GPU to improve the computing process. The results of using this GPU system are able to detect the goal with a processing speed of 58.39 FPS and 62.60 FPS for the detection of circular line meeting points. This system succeeded in estimating the approximate coordinates of the robot (x, y, θ) to the field with a success of 94.93% with an error of 5.17%. Errors in the estimation of coordinate positions are caused by less than optimal landmark detection. However, when the robot is on the line, the estimated position of the robot (x, y, θ) is 100% in accordance with real conditions.