Research On The Detection Technology Of Normativity Of Worker's Wearing Based On Image Recognition

Safety is a constant topic in industrial production and construction,the wear of working clothes and safety helmet is necessary for safe production which plays an important role in the prevention of accidents.In this thesis,a set of algorithms are proposed based on image recognition method for detecting normativity of workers' wearing and have been successfully applied in the monitoring system of Karamay oilfield in Xinjiang.Firstly,classical HOG feature based pedestrian detection algorithm is implemented.Due to the shortcoming of HOG feature's poor illumination robustness,HOG-LBP fusion feature based pedestrian detection algorithm is applied.Focusing on the shortcomings of HOG feature's higher dimension,the PHOG-LBP feature based pedestrian detection algorithm with PCA dimension reduction algorithm is applied.Based on comparing three different algorithms on feature extraction time,detection time and recognition rate,an appropriate algorithm is chosen which detection time(43ms)and recognition rate(95.8%)meet the requirements of both recognition rate and real-time performance.Secondly,a new worker-wearing normative detection algorithm based on HSV model is proposed.In this algorithm,human body area which is obtained from a pedestrian detection algorithm is chosen as ROI area and original RGB model image is transformed into HSV model image.Then,based on HSV color threshold,the image is transformed into a binary image.After a series of morphological operations on the binary image,situations like whether the human body wearing working clothes and pants,and the clothes is opened which are determined through the distribution and arrange of black and white pixels in binary image.A patent of this algorithm has been applied.Finally,two algorithms for detection of helmet wearing are proposed.The first is the helmet detection algorithm based on the random Hough transform head location.The random Hough transform circle detection is used to locate the head area of the human body,and HSV color model is used to determine whether the helmet is worn at the head position.The second is the helmet detection algorithm based on convolutional neural network and transfer learning.The method of transfer learning is adapted to learn the feature extraction method of the convolutional neural network model of Inception v3,the corresponding Softmax function is trained for image classification to determine whether there is a helmet exist in the image.The experimental results show that the latter has higher recognition rate and stronger generalization ability which can solve the problem of low recognition rate due to low pixels.Practical engineering applications show that overall recognition rate can reach more than 90% indicates proposed algorithms can effectively detect the normality of worker's wearing in camera surveillance,and alarm the irregular wearing condition.