Detection Of Key Positions In Production Line Based On Machine Vision

Human body target detection is a complex and challenging task.With the continuous development of computer vision and machine learning technology,human target detection has been extensively studied.However,the current researchers also mainly in the theoretical study of human target detection,due to the complexity of the actual scene,the human body target detection in the actual environment has not been large-scale applied.In this paper,we focus on the specific application environment of air conditioning assembly production line,and use the method of combination of regional and convolution neural network to solve the problem of detecting key personnel in the production line.In the stage of human target region extraction,this paper extracts the human target area based on the selection search and RPN algorithm respectively.The selection search mainly adopts the color space,shape,texture and level of the image itself to carry on the hierarchical grouping,and combines the exhaustive search and the segmentation intensity and so on,simultaneously uses the image structure to guide the characteristic sampling process.With the selection search,the human target location in the image will be captured as much as possible.RPN-based human target area extraction is mainly achieved by convolution neural network,that is,the use of human target samples to train a convolution neural network model,which does not need to consider the image of its own texture and other characteristics of factors.In order to produce a possible target frame area,a small grid is used to slide on the convolution feature plot of the last layer of convolution layer,which is slid on the input convolution feature graph using 3x3 window,and each sliding window is mapped the vector of lower dimension,which is sent to two independent all-connected layers: one is the regression layer of the human body target and the other is the target classification layer of the human body,and the non-maximum suppression algorithm is combined with the output a range of possible human target areas.According to the specific application scene of the key positions of the production line,this paper collects and designs a special sample set of human target and uses it to train the key position detection model.In the design of human convolution neural net-work,this paper uses ZF and VGG-16 convolution neural network to train.In the stage of human target network training,we use the pre-trained ImageNet network model to initialize the weight of the model network,and to fine tune the neural network weight parameters,which greatly reduces the calculation of the model and accelerates the training speed of the model.In addition,in determining the posture of key positions,this article makes full use of the characteristics of the target detection box,through the human body target detection box width and high ratio to determine the post is a posture or sitting position.The actual test results show that the algorithm is feasible and efficient,and the testing requirements of the key positions are reached in the production line.