Industrial Part Counting using Computer Vision

Abstract

Reliable quantity verification is important in logistics and manufacturing, where incorrect deliveries can lead to delays, customer claims, and increased costs. At the complete knock down (CKD) operations at Volvo Trucks Tuve, quantity verification is currently performed manually, making the process time-consuming and vulnerable to human error. Therefore, this thesis investigates the potential of using computer vision for industrial part counting. The aim of the thesis was to evaluate computer vision-based part counting at CKD packing stations by reviewing the current state of the art and assessing the performance of the Volvo Vision System (VVS). The study followed a design science research methodology approach and included a current state analysis, stakeholder analysis, interviews, observations, a literature review, and practical tests using VVS and YOLOv5 object detection models. Different hardware settings, lighting conditions, and dataset sizes were tested to evaluate the feasibility of the system in an industrial environment. The results show that computer vision-based object counting has potential for industrial applications, primary in structured environments. The experiments demonstrated that VVS could successfully detect and count several industrial parts under controlled conditions. Increasing the number of training images per class slightly improved the model performance, and the system achieved high accuracy for multiple object classes. However, the study also identified several challenges related to lighting variations, object overlap, reflections, scalability, and long-term maintainability. The study concludes that VVS has high potential and performance when it comes to object detection. However, in cases such as this one where there is a large amount of highly varied products and manual work, there are a lot of challenges that hinders a smooth implementation of a vision system.