ConstructNet: A Deep Learning Object Detector for Construction Site Surveillance

The construction industry is increasingly recognizing the imperative for automation, and one promising application lies in employing deep learning models for monitoring and surveillance of construction sites. Deep learning object detection algorithms have demonstrated their ability to swiftly identify a wide range of objects within images and videos. This paper introduces a novel architecture, ConstructNet, tailored for object detection of construction site machinery. Since the images of construction machines have large bounding boxes, ConstructNet focuses on detecting medium and large-sized objects. Our approach combines an FCOS-style architecture, a slim neck design paradigm, and CIOU loss. On the Alberta Construction Image Dataset (ACID), our model achieves an mAP of 0.71, whereas YOLOv8n has an mAP of only 0.65. ConstructNet is also compact, with a model size of only 13.58MB.