Detection of Pedestrians and Helmets in Large Construction Site

Abstract

It is necessary for workers to wear helmets when working in large construction sites. The traditional way to supervise the workers whether wearing helmets or not for safety is artificial, which have brought out many problems such as many blind spots, labor and time costing. Since a large number of surveillance cameras are currently installed in these construction sites, the surveillance video can be developed in taking the place of human supervision in an intelligent way. This paper designs a pedestrian and helmet detection network based on Faster R-CNN. In feature extraction, we have chosen the Residual Network (Resnet) combined with the Feature Pyramid Network (FPN) because the objects have small size, low resolution and less semantic information in whole scenes. We have also designed a parallel residual Block (PRB) combined with the Receptive Field Block (RFB) to strength feature extraction. The feature maps obtained from different convolution layers have been fused twice. And we have studied two fusion methods. Experiment results from our own dataset show that our proposed detection network improves the mAP by 8.74% and 2.3% respectively compared with Yolov3 and Faster R-CNN, at the cost of 0.3 FPS slower than Faster R-CNN.