Design Of Helmet Detection System Based On Deep Learning

Construction production is an important part of economic construction,and production safety is the lifeblood of construction production.As the last guarantee of workers’ lives,safety helmet plays an important role in construction safety.In order to ensure production safety,we should not only make workers voluntarily abide by safety regulations and wear safety helmets,but also establish an efficient and accurate monitoring mechanism.With the rise of smart city construction,image recognition and target classification methods based on deep learning provide methods for automatic supervision and monitoring.This thesis focuses on the monitoring video analysis of smart construction site,and studies from four aspects: feature extraction,regression positioning,semantic analysis and algorithm application,taking helmet monitoring as an example.First,this thesis starts with the feature extraction of target detection,analyzes the problems of current deep learning model in the application field through the migration of existing models,and proposes a feature extraction method based on irregular sampling for this problem,which can fully extract the foreground features of the target and prevent the misjudgment caused by background and foreground fusion.Furthermore,this thesis designs a new neural network framework based on this sampling method,and proves the feasibility of the transformation through comparison.Secondly,this thesis starts from the regression of target detection,analyzes the characteristics of high confidence box selection when the target regression,and designs a more flexible box selection strategy for the current situation of insufficient regression accuracy of the existing model.In the multi-scale detector,multiple boxes in the confidence interval can be fully used to make the detection results more accurate.Thirdly,starting from the classification of target detection,this thesis predicts the problem of similar semantic target detection by model comparison and verifies the problem by targeted samples.In this thesis,a new classification strategy is designed from the perspective of fuzzy semantic classification,and a classifier suitable for helmet detection is selected through experiments,and the detection framework and data set are further modified to adapt.Finally,this thesis designs a complete set of helmet detection system,which successfully deploys our algorithm on the cloud platform to provide external services.In addition,this thesis also successfully adapts the algorithm to the mobile platform,which provides diversity for the algorithm deployment.