Pose-Based Identification Using Deep Learning for Military Surveillance Systems

Surveillance systems research area these days has become more approachable duo to the increment of crimes and terrorism events, those systems ideal goal is to be able to identify any person anywhere at anytime, to prevent crimes before they even happen. This paper proposes a gait recognition scheme for person identification based on spatiotemporal pose by using skeletal data that was collected from RGB-D sensor. Concretely, the human gait over a period of time is described by 3D geometrical features with considering both distances and orientations of joints. The extracted features are then fed to an identification model which is fine-tuned from a pre-trained Convolutional Neural Network (CNN). According to the experiment on the benchmark dataset UPCV Gait, the system achieves the person identification accuracy of 92.88%, which is surprisingly good for using a dataset such as UPCV Gait with a fine-tuned pre-traind CNN.