3.8 Article

Machine learning methods for the industrial robotic systems security

出版社

SPRINGER FRANCE
DOI: 10.1007/s11416-023-00499-6

关键词

Information security; Security management; Security systems; Information processes; Mobile robots; Neural networks; Convolutions; Big data; Car parking

向作者/读者索取更多资源

This article discusses the trends in the introduction of industrial and logistics robots to ensure the safety of civilian facilities, the problems of increasing crime in the Russian Federation, and the decline in the identification of criminals. It also explores the use of BigData methods, specifically an ensemble of computer vision algorithms and convolutional neural networks, for timely detection of emergency situations in parking lots using mobile robots. The implementation of training on convolutional neural networks and the addition of a Squeeze-and-Excitation block (SE) improved the accuracy by 2-3%, reaching 88%, 91%, and 92% respectively. Comparison with the HOG-BoVW-BPNN method showed that DenseNet121 + SE achieved the same accuracy of 86% but with a 40% faster speed, making it a more attractive option for a car park computer vision system.
The trends in the introduction of industrial and logistics robots into the social sphere of activity in order to ensure the safety of civilian facilities, the current problems of the growth of crime in the Russian Federation, as well as the downward trend in the identification of persons who have committed crimes (taking into account the criminogenic situation in the Russian Federation) are discussed. The features of the application of BigData methods are considered, in particular, the use of an ensemble of computer vision algorithms and the mathematical apparatus of convolutional neural networks for the timely detection of emergency situations in parking lots by means of mobile robots. Implemented training of convolutional neural networks on the MobileNetV2, ResNet50 and DenseNet121 architectures with the addition of Squeeze-and-Excitation block (SE) to solve the problem of identifying semantic signs of vehicle damage in parking lots. The methods improved using the SE block made it possible to increase the accuracy by 2-3%, which amounted to 88%, 91% and 92%, respectively. At the second stage of work, when training neural networks, about 20% of the images obtained in difficult video shooting conditions (at dusk, during rain, snow, etc.) were used. The results of testing retrained neural networks on images obtained in difficult video conditions were compared with the method based on the HOG descriptor selected by us for similar conditions (a combined approach based on directional gradient histograms, the bag-of-visual-words method, and a neural network of inverse distribution). Compared with the comparable method, DenseNet121 + SE showed an accuracy of 86%, which is the same as the accuracy of the HOG-BoVW-BPNN method, but the speed of DenseNet121 + SE is 40% faster, which makes it a more attractive method for a car park computer vision system.

作者

我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。

评论

主要评分

3.8
评分不足

次要评分

新颖性
-
重要性
-
科学严谨性
-
评价这篇论文

推荐

暂无数据
暂无数据