Performance Comparison of YOLOv4 and YOLOv4-Tiny Algorithm for Object Detection on Wheeled Soccer Robot

This paper aims to describe performance comparison of YOLOv4 with YOLOv4-tiny for object detection on wheeled soccer robot using omnidirectional camera. Wheeled soccer robot mainly detects the ball, the goal, opponents, and its surrounding. Omnidirectional mirror is installed perpendicular to camera position such that the surrounding area which reflected on mirror captured by camera. Captured images are then process following YOLOv4 and YOLOv4-tiny algorithm scheme in order to detect founded object in the field. There are 4 objects to be detected by the robot such as ball, opponents, goal and keeper. There are 1000 images that has been used in training data, validation data, and tested data. Training was built first on the computer, then implemented it to the robot so that data can be collected. The result shows that the YOLOv4 model is better than the YOLOv4-tiny model in terms of mAP parameter based on number of datasets. However, based on frames per second, YOLOv4-tiny produced 7.7 fps compares to YOLOv4 with 2.01fPs. It is said that the YOLOv4-tiny is better 3 to 4 times than YOLOv4 in capturing images implemented on Rapsberry Pi on Wheeled Soccer Robot.