Data, language and graph-based reasoning methods for identification of human malicious behaviors in nuclear security

Abstract

Identification of human malicious behaviors is a demanding problem in nuclear security due to its potential to cause huge personnel and economic casualties. With the visibly rapid development of deep learning technology especially computer vision, the automatic identification of human malicious behaviors become feasible, offering significant savings in human resources. However, there is few research focusing on the explainable reasoning process based on the calculation results of computer vision. In this paper, an innovative framework is proposed, which contains two modules called computer vision module and reasoning module. Among them, the computer vision module could extract and detect human actions by some computer vision technologies, while these actions would be comprehensively utilized in reasoning module to derive the reasoning results. Furthermore, within the reasoning module, three novel reasoning methods are developed which are data-based reasoning method, language-based reasoning method and graph-based reasoning method. For evaluating the practicality and effectiveness of these methods, experiments are conducted on four typical scenarios in nuclear security, which are normal status, fence climbing, wire cutting and weapon holding. The results show that the proposed language-based reasoning method turns out to be the best one, which obtain a higher precision value of 0.7917 and a perfect recall value of 1.0000.