Low-Light Gait Recognition

A method to identify individuals by their gait, at a distance and in low-light conditions, combines infrared imagery and neural network–based pose estimation.
An illustation of a U.S. servicemember looking through a camera, highlighting two figures in the distance. A zoom in view of the figures highlights points on their body that can be used to recognize their gait.
The Low-Light Gait Recognition program seeks to provide a method to identify individuals at a distance and in low light conditions. Image: Tom Sheehan, Shawn Dufour

Many defense and homeland security missions would benefit from the ability to surveil and identify individuals from far distances and in poor visibility. In these conditions, customary identification technology — such as fingerprints, iris imagery, or facial recognition — may be impractical or unreliable. One promising method to identify a person in these scenarios is by observing their gait, or their unique pattern of walking.

Gait has long been studied as a mechanism to identify individuals with good results when using clear, visible images and consistent viewpoints. Limited research has also been conducted with non-visible imaging modes, such as infrared (IR) imagery, which is particularly promising for identifying subjects in low-light conditions. But challenges remain in reliably identifying subjects when faced with inconsistent angles or partial obstructions.

We seek to extend gait recognition techniques to operate on IR images. Importantly, the program is also examining the use of recent developments in pose estimation techniques to improve performance when subjects are viewed from differing angles and when some level of obstruction may be present. These pose estimation techniques use advanced neural networks to locate body positions (e.g., joints) in images of subjects. An initial study we conducted showed encouraging results with respect to reliable detection of key points on subjects, provided that the images are of sufficient resolution, which is achievable with practical optics.

Our team is now focusing on enhancing the performance of key point detection in IR imagery and in constructing a recurrent neural network–based gait identification system. This identification system operates directly on key points from a pose estimation algorithm. By applying 3D pose estimation techniques to the imagery, we believe we can improve our systems' robustness to view-angle and obstructions.