A recent study in Sustainability presents a new deep neural network model aimed at improving construction safety by detecting proximity and relationships simultaneously to address risks associated with the use of robots. The study highlights the high accuracy achieved in relationship detection evaluations and suggests areas for further refinement to enhance hazard identification efficacy.
A recent study published in Sustainability has introduced a deep neural network (DNN) model to enhance construction safety by performing proximity and relationship detections simultaneously. The study, conducted by Kim, Goyal, Lee, Kamat, and Liu (2024), addresses the risks of forcible contact associated with the increasing use of robots in the construction industry.
The research utilized a single-shot DNN model, Pixel2Graph, which specializes in multi-scale feature abstraction and single-shot relationship detection. Training involved a baseline model pre-trained with Visual Genome, using construction-specific data for fine-tuning. The models aimed to perform relationship detection, object classification, and localization.
Three models with varying difficulty levels were tested: Only-Rel (relationship detection), Cla-Rel (object classification and relationship detection), and Loc-Cla-Rel (localization, classification, and relationship detection). Results showed high accuracy in Recall@5 evaluations: 90.63% for Only-Rel, 72.02% for Cla-Rel, and 66.28% for Loc-Cla-Rel on test datasets.
While the study demonstrates the potential for enhancing construction safety through DNN-based hazard identification, the models’ performance on test datasets indicates a need for further refinement. The research proposes that improving training data and advancing DNN architectures could enhance single-shot visual relationship detection efficacy.
Journal Reference: Kim, D., Goyal, A., Lee, S., Kamat, V.R., & Liu, M. (2024). Single-Shot Visual Relationship Detection for the Accurate Identification of Contact-Driven Hazards in Sustainable Digitized Construction. Sustainability, 16(12), 5058.
