SKDMap-Net System Identifies People by How They Walk, Not by Their Face

A team of researchers led by Binge Quan has described in the International Journal of Reasoning-based Intelligent Systems an artificial intelligence system called SKDMap-Net that identifies people by their gait, the individual pattern of body movement while walking. Unlike facial recognition, the method works even when the face is covered, turned away from the camera, or only partially visible.

SKDMap-Net does not analyze raw video footage but rather skeletal keypoints extracted from it, meaning the positions of joints such as shoulders, elbows, hips, knees and ankles, tracked frame by frame. The system calculates joint flexion angles, movement speed and acceleration, then processes body-position data and movement-dynamics data separately before merging them into a single description of a person's gait.

How the attention mechanism works

A key element of the architecture is an attention mechanism that assigns greater weight to the body parts best visible in a given scene. If the legs are partially obscured by an obstacle or another person, the system automatically relies more heavily on torso and arm movement for recognition. This lets the method outperform older approaches based purely on silhouette analysis, which were easily thrown off by an unfavorable camera angle or a change of clothing.

The authors emphasize that separating static body-position information from movement-dynamics information, then recombining them through attention, produced results that surpassed existing methods on all three public datasets tested.

Why gait is harder to hide than a face

Gait belongs to behavioral biometrics, meaning traits that describe how a person acts rather than how they look. Unlike a face, which can be covered with a mask or a hood, the way a person moves can practically not be consciously altered over the long term without significant physical effort. That makes the method attractive for surveillance systems that need to identify people from a distance or in poor lighting, conditions where facial recognition fails.

At the same time, the very trait that makes gait useful biometrically also makes it hard to protect. Passwords and ID documents can be replaced after a breach, but gait cannot be changed. A movement pattern, once collected, stays with a person for life and can be used to re-identify them even after facial data has been removed.

The authors' case for less invasiveness

The paper's authors present a non-obvious argument in favor of their approach: since the system operates on skeletal point coordinates rather than full video footage, it could in theory require storing less identifying visual data about a person than classic surveillance systems that record high-resolution video.

This approach could make gait recognition more reliable while reducing the amount of visual data about a person that needs to be processed - from the description of the SKDMap-Net methodology in the International Journal of Reasoning-based Intelligent Systems

That argument does not, however, remove the core risk. Reducing the amount of raw footage does not reduce identification capability, since skeletal data is still enough to recognize a specific person across multiple cameras scattered throughout a public space and build a detailed map of their daily routes and habits.

What this means for Poland

Poland is currently working on national implementation of the EU's AI Act, which classifies remote biometric identification in public spaces as high-risk, with some applications subject to strict restrictions. Systems like SKDMap-Net show that the technical capability to identify people is outpacing the speed at which regulations governing their use by public bodies and private operators of urban surveillance are being written.

For operators of surveillance systems in Polish cities and shopping centers, this means that simply forgoing facial recognition no longer guarantees anonymity for people caught on camera. Identity can now be established through methods that are not formally called facial recognition yet still lead to identifying a specific individual.

The paper's authors explicitly argue that before gait-recognition technology is deployed more widely, strict rules are needed governing data storage, access, and how such systems are implemented, so that no loophole emerges that bypasses existing regulations on facial biometrics.

Sources: How you walk could identify you: New AI boosts long-range security checks (techxplore.com), Adeus reconhecimento facial: a sua forma de andar é a sua nova impressão digital (pplware.sapo.pt), Ciemne okulary i kaptur już nie pomogą (spidersweb.pl)