A Motion Deviation Image-based Phase Feature for Recognition of Thermal Infrared Human Activities

• Published in: Engineering letters Vol. 28; no. 1; p. 48
• Main Authors: Tan, Yong, Yan, Wenjuan, Huang, Shijian, Du, Derong, Xia, Liangping
• Format: Journal Article
• Language: English
• Published: Hong Kong International Association of Engineers 22.02.2020
• Subjects: Deviation; Feature extraction; Feature recognition; Human activity recognition; Human subjects; Image quality; Infrared imagery; Infrared imaging
• ISSN: 1816-093X; 1816-0948

A novel activity feature for thermal infrared human activity recognition, which suffers from poor quality infrared imaging and a great variation in human subjects, is proposed in this paper. Extraction of this feature consists of three major stages. First, a coarse-to-fine localization procedure extracts the regions of interest (ROIs) that may contain human subjects from a raw infrared sequence. Second, it generates a motion deviation image (MDI), which is a novel spatio-temporal template, from the available ROIs to represent the infrared activity sequence efficiently. Third, from such MDI, it generates a directional phase congruency-based feature that codes the significant activity information via the steps including calculating phase maps, estimating intrinsic dimensions and reducing vector dimensionality. The proposed feature is validated on the IADB infrared human activity database, and the experimental results show its advantages in recognition accuracy when used for multi-class activity recognition due to its robustness to poor quality thermal infrared imagery, good representation of activity information and effective removal of noise.