Deep Learning on Monitoring Building Construction

Now a days building constructions have led to low efficiency of construction schedule management and caused many construction projects to have cost overruns and legal disputes with site managers due to schedule delays on time. The manual extraction of construction procedural constraints is costly and also time-consuming. Construction industry suffers from the highest number of problems among all industries, i.e., one in five worker deaths in private industry were in construction. Tremendous loss has occurred to the workers families, the industry. Considering the increasing in number of these accidents there is a growing necessity of developing innovative methods to automatically monitor the safety for the workers at construction sites. Since the head is the most critical area of a human body and is the most vulnerable to an impact that could cause serious injury or death, the use of a protective helmet in construction work is needed. In this paper, we aim to automatically detect the uses of construction helmets e.g., whether the construction worker wears the helmet or not by analysing the construction surveillance images. Based on the collected images, we first detect the object of interest i.e., construction worker and further analyze whether the worker wears the helmet or not, by using computer vision and machine learning techniques like using of convolutional neural network process of resnet34 model which is pretrained on image-net dataset. In the first step, we incorporate frequency domain information of the image with a popular human detection algorithm Histogram of py-troch for construction worker detection; in the second step, the combination of colour-based and open cv feature extraction techniques is applied to detect helmet uses for the construction worker. With regard to actual video surveillance of power station, the content change of frames is not so intense, thus, we can combine the detecting results of several adjacent frames to improve the recognition accuracy of whole system. For example, if a safety helmet with similar position coordinates has been detected in four of five real time video frames, then the confidence of safety helmet detecting results will be largely increased.