Research on Real-time Monitoring Method of Small Box Girder Erection Progress Based on Improved Yolov4

This paper proposed an object detection algorithm based on improved Yolov4 to monitor the erecting progress of small box girders in real-time. Firstly, this paper creates a dataset from the collected surveillance videos and annotates them. Secondly, the Yolov4 network structure is improved in this paper to enhance the network characteristics and detection capability. Then, the k-means++ algorithm labels are re-clustered to enhance the fit between anchor boxes and feature layers. Finally, cosine annealing and transfer learning methods are introduced to enhance the model training effect. A new evaluation metric-AMDR (average miss detection rate) is also proposed in this paper. Compared with Yolov4, the improved Yolov4 not only improves the mAP (average mean accuracy) of the test set by 3.3%, but also reduces the AMDR by 10.64%, which significantly improves the recognition accuracy of small box beams. Therefore, the algorithmic model has strong engineering application value in civil bridge construction.