OBJECT DETECTION SYSTEM BASED ON YOLO AND PALLET POSE ESTIMATION WITH ARUCO MARKER FOR AUTONOMOUS FORKLIFT

• Main Author: Sean Kesuma, Eric
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/73660
• Institution: Institut Teknologi Bandung

Utilizing technology such as artificial intelligence and robotics potentially improves E-Commerce trends on the production and consumption side. Autonomous forklift is the innovation that can be implemented for picking and distributing goods in warehousing with pallets. Pallet is a horizontal rectangular structure that has a cavity for placing objects on it. The picking system in the warehouse using autonomous forklift requires a perception system which consists of detecting and estimating poses on the pallet when in the picking mode and detecting the dynamic object while in the navigation mode. This research will cover how to estimate pose pallets with some different approaches as well as human and another forklift detection. ArUco marker will be used on the custom pallet model to estimate the position and orientation in real time using the coordinates transformation system as well as information from roll pitch yaw angles representation. The object detection system will be implemented using the You Only Look Once (YOLO) method in estimating dynamic objects with various lightning condition in a warehouse. Camera calibration, distance, position, orientation, and the data communication protocol will be discussed in this study. Position estimation system has the value of error 1.621 cm, 2.99 cm, and 2.7966 cm in the x, y, and z planes respectively. Orientation estimation also has the value of error 0.3512 degrees in roll, 2.094 degrees in pitch, and 1.4992 degrees in yaw. The speed of object detection can reach 24 frames per second (FPS) with the converted YOLOv5 to ONNX (Open Neural Network Exchange) model and the distance to marker with regression method in calibration has an error rate of 3.99 cm on the first camera and 2.26 cm on the second camera. The detection of the marker also reached 30 FPS on camera. Based on the experimental results, the object detection model and pose estimation on pallet have provided the expected estimation results locally.