DeepSafety: a deep neural network-based edge computing framework for detecting unsafe behaviors of construction workers

Recently, the development and application of artificial intelligence have received widespread research attentions, and one of important applications is accident prevention. Since most accidents on construction sites are caused by construction works’ unsafe behaviors, unsafe behavior detection is desired. Unlike traditional detection model which focuses only on accuracy of detection and ignore efficiency of detection, a deep neural network-based edge computing framework is proposed for detecting unsafe behavior not only efficiently but also precisely. To address efficiency issue, an object detection model and a posture estimation model are established on edge device for extracting feature from surveillance camera streaming, a time series classification model is developed on server for detecting unsafe behavior according to the features extracted from edge devices. Finally, a comprehensive experimental study based on three datasets collected from three real construction sites is conducted. The results showed that the models proposed in this study can achieve 87% in terms of accuracy and 1.5 s in terms of latency. The improve rate of the proposed DeepSafety is higher than 25% in terms of accuracy and 80% in terms of latency. Accordingly, the proposed edge-computing based framework is shown to deliver excellent performance.