A deep learning approach for real-time rebar counting on the construction site based on YOLOv3 detector

Counting steel bars is a routine daily task for steel manufacturers and most building construction sites. Currently, counting is usually performed manually, which is laborious, time-consuming, and error-prone. This study proposes a deep learning approach based on YOLOv3 detector for automatic steel bars detection and counting through images. Three new measures, including an additional feature pyramid, complete intersection over union (IoU) loss and focal loss, and bag of freebies, were introduced to improve rebar detection and counting performance. A dataset containing 74,824 rebar sections from real construction sites was constructed and utilized to evaluate the proposed approach. The application results demonstrate that the suggested measures can significantly improve the performance of the YOLOv3 detector to high average precision of 99.7% at IoU of 0.5. Comparisons with other detectors show that the proposed approach is fast, accurate, and robust at rebar counting under different construction site conditions. It can be facilitated as the basis of a real-time, cost-effective rebar counting scheme.

Automated summary

This study proposes a deep learning approach based on the YOLOv3 detector for automatic steel bar detection and counting, which significantly improves performance through the introduction of new measures. The results from the dataset containing real construction site data demonstrate that the method performs robustly and accurately under different conditions for rebar counting.