An obstacle detection system for automated guided vehicles

Abstract

The objective of this master's thesis is to investigate the utilization of computer vision and object detection as an integral part of an automated guided vehicle's navigation system, which operates within the facilities of the target company. The rationale for conducting this research and developing an application for this purpose arises from the inability of automated guided vehicles to detect smaller or partially obstructed objects, and the lack of differentiation between stationary and moving objects. These limitations pose a safety hazard and negatively impact the overall performance of the system. The anticipated outcome of this thesis is a proof-of-concept computer vision application that would enhance the automated guided vehicle's obstacle detection capacity. The primary aim is to offer practical insights to the target company regarding the practical implementation of computer vision by developing and training a YOLOv7 object detection model, as a proposed resolution to the research problem.

A thorough theoretical part of the required technologies and tools for training an object detection model is followed by a plan for the application to define requirements for the object detection model. The training and development are conducted using open-source and standard software tools and libraries. Python is the primary programming language employed throughout the development process and the object detector itself constitutes a YOLOv7 (You Only Look Once) object detection algorithm. The model is trained to identify and classify a predetermined number of objects or obstacles that impede the present automated guided vehicle system. Model optimization follows a fundamental trial-and-error methodology and simulated testing of the best-performing model. The data required for training the object detection model is obtained by attaching a camera to an automated guided vehicle and capturing its movements within the target company's facilities. The gathered data is annotated using Label studio, and all necessary data preparation and processing are carried out using plain Python.

The result of this master’s thesis was a proof of concept for a computer vision application that would improve and benefit the target company’s day-to-day operations in their production and storage facilities in Vaasa. The trained model was substantiated to perform up to expectations in terms of both speed and accuracy. This project not only demonstrated the application's benefits but also laid grounds for the business to further utilize machine learning and computer vision in other areas of their business regarding the operational improvement competency of the target company’s employees. The results of this master’s thesis showed that finding an optimal object detection model for a specific dataset within a reasonable timeframe requires both appropriate tools and sufficient research data premises in terms of model configuration. The trained model could be utilized as a foundation for similar projects and thereby reduce the time and costs involved in preliminary research efforts.