Suspicious behavior recognition using deep learning

Focusing on keypoint 2D scaling

Abstract

The purpose of this study is to reinforce the defense and security system by recognizing the behaviors of suspicious person both inside and outside the military using deep learning. Surveillance cameras help detect criminals and people who are acting unusual. However, it is inefficient in that the administrator must monitor all the images transmitted from the camera. It incurs a large cost and is vulnerable to human error. Therefore, in this study, we propose a method to find a person who should be watched carefully only with surveillance camera images. For this purpose, the video data of doubtful behaviors were collected. In addition, after applying a algorithm that generalizes different heights and motions for each person in the input images, we trained through a model combining CNN, bidirectional LSTM, and DNN. As a result, the accuracy of the behavior recognition of suspicious behaviors was improved. Therefore, if deep learning is applied to existing surveillance cameras, it is expected that it will be possible to find the dubious person efficiently.