Fabrication progress detection for concrete T-girders based on improved Yolov4

Abstract

Large precast concrete girder plants have many processes, long cycles and a large amount of data. This study proposes an improved Yolov4 object detection algorithm with a spatio-temporal relationship to detect each fabrication process of precast concrete girders. It realises the digitisation of the fabrication information of traditional precast concrete girder plants. Initially, adding upsampling and convolution layers to the Yolov4 base model enhances the feature extraction ability of the algorithm at different fabrication stages of precast concrete girders. The spatio-temporal relationship is adopted to determine the fabrication progress of precast concrete girders with identical features but are at various fabrication stages. Finally, this research conducts an application analysis of an actual precast concrete girder plant. The analysis result indicated that the improved Yolov4 algorithm significantly raises the mean average precision and average intersection over union in recognition. Besides, the spatio-temporal relationship effectively solves error detection problems caused by the similar appearance at different fabrication stages. This method provides practical support for digitising the fabrication data of traditional precast girder plants.