Research On Detection Of Hardhats Worn By Construction Based On Deep Learning

The construction industry has become one of the important material production departments and pillar industries of China’s national economy,At the same time,it is also one of the most dangerous work areas.Careers in the construction field are always involved with risks and engender dangers to which workers and professionals are exposed.Hardhats play an essential role in protecting construction individuals from accidents.It is also one of the basic safety regulations on the construction site,which all workers and visitors should always abide by.However,for various reasons,people do not strictly enforce the rules of wearing helmets.To enhance construction sites safety and to facilitate the safety monitoring work of safety inspectors,The traditional method is to deploy part-time safety officers to inspect the construction site regularly,which has high cost and poor effect.With the mature development of artificial intelligence technology,it is possible to use machine learning algorithm to realize the intelligent monitoring of safety helmet.Most of existing projects are used to monitor the wearing of safety helmets through multi-stage data processing,but these methods rely on the manual production to detect the personnel on the construction site heavily.There are also limitations in adaptability and universality.Based on this,in this thesis,a one-stage system based on convolutional neural network is proposed to automatically monitor whether construction personnel are wearing hardhats and identify the corresponding colors.It is motivated by the development of the Single Shot Multi-box Detector,which detects objects by directly regressing bounding boxes via a single CNN.However,the SSD commonly fails to detect small-scale objects because of its inherent properties of weak features at the bottom layers of high resolution,although it has built a feature pyramid networks.In order to detect small-scale objects effectively,such as hardhats.we develop a novel aggregation framework combined with the presented reverse progressive attention,which propagates the semantically strong features back to the bottom layers progressively,and the different features extracted from different layers can be fused to generate a new feature pyramid,then the SSD is used to predict the final detection results.In this way,it can learn the feature information from the top level effectively.To facilitate the study,this work constructs a new hardhat wearing detection benchmark dataset,which consists of 2116 images covering various on-site conditions.Each object is annotated with a class label and its bounding box.The experimental results demonstrate that the proposed system is effective under all kinds of on-site conditions,which can achieve83.89% m AP with the input size512 ×512.