Research On Helmet Detection Based On Improved CenterNet Algorith

In recent years,the real estate industry has risen,and the number of construction workers has greatly increased.The lack of standardized management at the construction site has led to an endless stream of safety accidents.In safety accidents,the common cause of accidents is that workers do not wear safety helmets.At the construction site,the common detection method is manual detection,but manual detection has limitations such as not being detected in a timely manner,personnel fatigue,and nepotism.Using computer detection is a more effective method.In recent years,the rapid development of computer vision has provided efficient and low-cost detection for helmet wearing.Considering the limitations of traditional machine learning,such as being heavily influenced by environmental factors such as lighting and extracting shallow features,adopting a deep learning model is more effective.However,the detection area of safety helmets is small and belongs to small targets in computer vision.Deep learning models have low accuracy in detecting small targets.Center Net is a 21 year old SOTA(States of the Arts)algorithm that performs well in detecting small targets with high accuracy.This paper analyzes the structure of Center Net,and finds that Center Net has the disadvantages of single feature without reuse,poor performance of loss function,and insufficient information extracted from convolution kernel.In response to these shortcomings,this article proposes corresponding improvement methods aimed at improving Center Net’s performance on the helmet dataset.The main work of this article is as follows:(1)Using crawler technology and image capture technology to obtain a total of7581 related image data,and using Label Img tool to annotate the obtained images.Expanding the dataset using data augmentation techniques such as target displacement,Gaussian noise,and image flipping to enhance the robustness of the network enables the model to perform well in complex environments.(2)In order to solve the problem of single Center Net structure and non reuse of feature maps during the calculation process,FPN Center Net structure,ASFF structure,and DASFF structure are proposed to reuse and fuse features,and experiments are conducted on a helmet dataset.Compared with the original Center Net,the m AP values increased by 4.06%,2.52%,and 3.63%,respectively.Therefore,the FPN Center Net structure with the best performance among the three is selected for feature fusion and subsequent improvements.(3)Center Net does not perform well in the two tasks of category prediction and location prediction,so this paper uses the loss function to improve the fitting effect of the model on the two tasks.The improvement of the loss function is divided into two parts:(1)In order to solve the problem of low category confidence in Center Net’s category calculation process,Focal Mse One Loss and Focal Mse Guss Loss loss function are proposed,Compared with the original Focal Ross loss function on the helmet data set,the MAE value decreased by 3.7591 and 3.9542 respectively(the lower the MAE value,the better),and the m AP value increased by 1.85% and 2.39%respectively.The loss function with the best effect of both was selected as the category loss function;(2)In order to solve the inaccurate problem of Center Net in the process of category positioning,CIo U Loss loss function was used to replace the original positioning loss function to solve the problem of insensitive Io U(Intersection over Union).Through the ablation experiment,compared with the original positioning loss function on the helmet dataset,the m AP value increased by 1.05%;(4)The Center Net structure uses a simple convolution kernel for calculation,which is not sensitive to flipped objects.At the same time,the simple convolution kernel can not extract the information of multiple Receptive field,so the convolution kernel improvement is divided into two parts:(1)In order to solve the problem that Center Net is not sensitive to flipped objects,the ACB convolution core is studied and the ordinary convolution core in the backbone network is modified to ACB convolution core,which increases the m AP value by 0.53%;(2)In order to solve the problem of a single Receptive field in Center Net,Py Conv convolution kernel was studied and the ordinary convolution kernel sampled up was modified to Py Conv convolution kernel to fuse multiple Receptive field,which increased the m AP value by 0.63%.Finally,the improved algorithm was compared with other algorithms on the helmet dataset and MS COCO dataset.Compared with the original Center Net,the m AP values of the improved algorithm on the two datasets were increased by 8.69% and 1.36%,respectively.The FPS on the independent graphics card GPU of the Ge Force GTX 1050 reached 24.81.The improved algorithm proposed in this article not only ensures detection speed,but also significantly improves detection accuracy.