Ball detection for KRSBI soccer robot using PeleeNet on omnidirectional camera

Abstract

Kontes Robot Sepak Bola Indonesia (KRSBI) is an annual event for contestants to compete their design and robot engineering in the field of robot soccer. Each contestant tries to win the match by scoring a goal toward the opponent's goal. In order to score a goal, the robot needs to locate the ball. We employ an omnidirectional vision camera as a visual sensor for a robot to perceive ball. We calibrate streaming images from the camera, in order to remove the mirror distortion. We deploy PeleeNet as our deep learning model for object detection. We fine-tune PeleeNet on modified PASCALVOC 2007-2012 dataset with the additional ball object. Our experiment results show PeleeNet has the potential to be deployed as a deep learning mobile platform for KRSBI as the ball detection architecture. It has a perfect combination of memory efficiency, speed and accuracy.