| dc.description.abstract | Currently, research on securing safety by unmanned systems is being actively conducted. Development is underway to reduce costs and secure worker safety by filling safety-related personnel’s blind spots and reducing their burden. For intelligent safety security, we propose artificial intelligence models that can detect, identify and distinguish major objects based on photographic information. In addition, Frequency Channel Attention Network (FcaNet), which supplements the existing Global Average Pooling (GAP) method, is used to improve the existing algorithm, and the accuracy is improved.
For this purpose, 12,000 pieces of photographic data images are collected for 5 major equipment to be encountered in the actual construction environment. The detection and identification performance of the model is maximized by using the FcaNet layer for learning through the existing Faster-RCNN, Libra-RCNN, and Double-Heads model. As a result, the accuracy of the test dataset is improved by 6%, 0.4%, and 0.4%, respectively. And, through using random initialization and improved batch normalization, the shortcomings of limited data are reduced, and the effect of pretraining is obtained without. This results in an improvement of more than 20% in each model, and the revised model shows 0.5% higher than the existing one. It is hoped that these results will be reflected in the work environment intelligence project to further reduce the burden on manpower and improve efficiency. | |
