Object Recognition and Tracking of Bolts: A Comparative Analysis of CNN Models and Computer Vision Techniques : A Comparison of CNN Models and Tracking Algorithms

Abstract: The newer generation industry 4.0 focuses on development of both flexibility and autonomy for power tools used by companies in different mechanical areas and assembly lines. One area for automation is the application of computer vision in power tools to detect, identify and track bolts. This technology can be utilized for indoor localization of the tool or for automatically adjusting the torque to the appropriate level. This study investigates the performance of different object detection models, namely YOLOv8, SSD and Faster R-CNN, using a custom dataset for bolt detection. The study aims to evaluate the effectiveness of these models in terms of inference speed and mean Average Precision (mAP) score, while also addressing challenges related to dataset quality and data association methods for object tracking. The results indicate that the YOLOv8 model, particularly the YOLOv8n variant without the P5 detection layer in the head, outperforms the other models in terms of both inference speed and mAP score. As for the two implemented tracking algorithms, namely ORB feature matching and Kalman filter, the ORB feature matching method exhibited commendable performance and less ID swaps when detections were consistent and occlusions were minimal compared to the Kalman filter. However, it exhibited limitations in handling occlusion and struggled with undetected or out-of-frame objects. On the other hand, the implemented Kalman filter partially mitigated occlusion-related challenges but encountered difficulties in maintaining accurate associations over multiple frames.