A Research And Implementation Of Safety-helmet Video Detection System Based On Human Body Recognition

Much of the accidents occured in the production and construction sites are related to the violations of rules by personnels.Not wearing a helmet properly is the most common violation behavior in most production and construction sites.Therefore it is crucial to use modern techniques,which are required to be real-time,accurate and foreseeable,to eradicate these violations in the production and construction sites.This goal is achievable by the fact that based on the techniques of human body recognition,it is feasible to detect if the human body is wearing a helmet or not.The goal of this thesis is to study the relevant algorithms and implement the system,to realize the automatic detection of helmet wearing condition of personnels in the production field.The thesis studied the theories and algorithms applied by the system of helmet detection based on human body recognition,including the moving target detection based on background subtraction division,and the characteristic of the background subtraction division and applicable scenarios,and studied how to locate the head area and the computation of the helmet's color feature.Next,the enhancement of low illumination video images is studied,this problem is mainly about the condition of low illumination,and low contrast.The algorithm of image enhancement with color retaining capability is introduced,including its theory and the concrete steps of the algorithm,the algorithm can enhance image while retaining its hue and saturation information,it provides a solid foundation to compute the color fearture of helmets.The method to locate head area is studied,the method of reducing the computation complexity of HOG feature by applying HOG integral figure is studied and implemented,the mathematical principles of batch processing of SVM is studied and implemented.Based on relevant algorithms,this thesis designed and implemented the system of helmet detection based on human body recognition,at the same time,the support vector machine capable of batch processing and multithread processing is implemented,eliminated the bottleneck of the real-time performance of the system and solved the problem of out of memory while training huge chunk of data.The system utilizes the HOG feature to detect human body,based on this,the system adopts the combination of HOG and color feature to detect helmet.The experimental results indicate that the system implemented in this thesis has a good effect on recognition rate and adaptation.