China is pioneering efforts to track and surveil their citizens using AI systems with data from facial recognition systems paired with other tracking technologies. Body movement or gait analysis technology has been deployed already in Shanghai and Beijing, and perfectly complements facial recognition technology for this purpose.
One company, Watrix, has software that can identify people based on physical characteristics from up to 50 meters away, whereas facial recognition technologies require a relatively close view of a person’s face. However once scanned at close range, an identified person can be tracked at a distance using gait recognition software from pretty much any system of cameras.
This works regardless of viewing angle and does not require any identifiable body parts to be showing, like faces. Huang Yongzhen, the CEO of Watrix, said in a recent interview, “You don’t need people’s cooperation for us to be able to recognize their identity. Gait analysis can’t be fooled by simply limping, walking with splayed feet or hunching over, because we’re analyzing all the features of an entire body.”
Huang says the system has a 94% accuracy rate, which is good enough for commercial use, although less accurate than facial recognition software. The other main obstacle is that it does not yet work in real time due to the complexity of the processing the data. This is balanced by the fact that high resolution and close proximity are not needed.
These systems typically measure the movement of neck, shoulders, elbows, spine, hips, knees and ankles. Angles between body parts are measured while in motion as well. Indicators usually measured for these techniques are step length, stride length, cadence, speed, progression line, foot angle, hip angle. This identification method is not fooled by limping or changes in walking pattern resulting from self-induced factors like putting a rock in one shoe. However, a person’s gait can be influenced by emotional state.
There were other attempts earlier that did not pan out, but now machine learning systems seem to provide the needed capability to make this work at scale. Groups of people walking around in public places can be identified now, as we’re seeing in China. This is commonly used in the medical community to assess rehabilitation progress, including recovery from osteopathic or chiropractic procedures.
At the opposite extreme we have brain fingerprinting technologies like that being developed by Damien Fair and colleagues. This idea maps connections in the brain as an identifier that is unique to an individual, like a fingerprint on a finger. His lab and others are exploring possibilities like identifying relatives by patterns found in these maps that they’re calling “functional connectomes.”
This method of identifying individuals is not portable or reliable enough at this stage to be ready for broad use, but it suggests that people can be distinguished by the way they process information, in a way that is hard to fake or spoof. The current work uses MRI imagery, which is not portable and requires direct proximity, but one can imagine a few years from now when the situation has advanced.
Another technique that can be used to identify individuals at distance is by profiling their cardiac signature. The US Dept of Defense has a prototype that can acquire a sufficient pattern of signals from 200 meters away. It is thought that this method can be used from even greater distances as prototypes are refined. This can be used for authentication as well, as a smart watch from Nymi shows.
This method works by using lasers to read the heartbeat data, even through clothes. The newest method uses a technology called laser vibrometry, which measures tiny surface movements caused by a heartbeat. Previous attempts have used techniques like using infrared sensors to detect blood flow, but this is accurate at a distance and reliable.
The reason this technique is considered so compelling is that a person’s heartbeat characteristics cannot be faked or easily spoofed. Every person’s pattern is unique, unlike results you might get from facial recognition systems. The accuracy rate is thought to be capable of reaching almost 100%.
Finally let’s take a look at promising new areas of research into extreme facial recognition with limited data. This involves using machine learning to extrapolate from sparse data to create a reasonable guess at what a higher resolution image would show. A low resolution image of a face can be processed to create a higher resolution approximation that has surprisingly good results from as little as 256 pixels.
Researchers in Korea were able to construct facial images that were reasonably good approximations, even if almost no data was available. They fed 16×16 pixel images into this system and it produced accurate, if a bit weird images. This implies that facial recognition techniques might be used even when images are acquired from many kilometers away, perhaps even from satellite imagery. The enhanced imagery might be enough to get good results from facial recognition algorithms.
We covered plenty of current technologies in previous posts like vein mapping, fingerprint identification, and even behavioral biometrics but these were some promising new technologies not yet widely deployed. What will they come up with next? Stay tuned to find out!