Detecting safety helmet wearing on construction sites with bounding-box regression and deep transfer learning

摘要

Detecting safety helmet wearing in surveillance videos is an essential task for safety management, compliance with regulations, and reducing the death rate from construction industry accidents. However, it is much challenged by some factors like interocclusion, scale variances, perspective distortion, small object detection, and the carrier recognition of safety helmet. Traditional image-based methods suffer from them. This article proposes a new methodology for detecting safety helmet wearing, which makes use of convolutional neural network-based face detection and bounding-box regression for safety helmet detection. On the one hand, the method can help to recognize the carrier of the safety helmet and detect a multiscale and small safety helmet. On the other hand, deep transfer learning based on DenseNet is introduced and applied using two different strategies, namely, object feature extractor and fine-tuning for safety helmet recognition. To further improve the recognition accuracy, the network model with two peer DenseNet networks is trained by mutual distillation. Extensive analysis and experiments show that the novel methodology has considerable advantages in detecting safety helmet wearing compared to other state-of-the-art models. The proposed model has achieved 96.2% recall, 96.2% precision, and 94.47% average detection accuracy. These results, precision-recall (PR) curve, and receiver operating characteristic (ROC) curve demonstrate the feasibility of the new model.