Radio Killed the Video Stream

Re-identification of people across time and place is typically accomplished by collecting RGB video or images and algorithmically classifying them based on past examples. These methods learn to detect surface features, such as clothing and hairstyle, that can change greatly from day to day and reduce their effectiveness over longer periods of time. In a recent paper, a team from MIT CSAIL has proposed a new approach (RF-ReID) that side steps the limitations of visible light imagery by harnessing radio frequency (RF) signals.

Unlike RGB imaging, RF signals penetrate clothing and reflect directly off of the body. As such, these signals can detect more persistent identifying features, like body size and shape. RF also has the advantage of being effective in the absence of light, and maintaining effectiveness in the presence of certain obstructions that would render an RGB camera useless.

An FMCW radio with a horizontal and vertical antenna array, operating between 5.4 and 7.2 GHz, is employed to give an effective detection range of 12 meters. The radio signal is returned in the form of a pair of heatmaps (horizontal and vertical), at a rate of 30 per second. Using existing methods, an RF tracklet (reflected RF signals, and the bounding box of the person) is extracted from each frame. This information is used to determine the 3D location of 18 major body joints. This data is supplied to a machine learning model that extracts features that identify the individual body shape and walking style of different people.

RF-ReID is an improvement in many ways from a privacy perspective; cameras can be intrusive and collect very sensitive information. The new method, on the other hand, collects less intrusive measurements. It also provides an advantage when compared to other location tracking devices such as smartwatches. These devices are only effective when the people to be recognized remember (or choose) to wear them.

The privacy-preserving features of this technique may allow it to be applied in healthcare settings, where RGB cameras are likely to be avoided over privacy concerns. The tracking features could be used in long-term care settings for the safety of residents. Another interesting application, made possible by recognizing the walking style of individuals, is in the potential to diagnose movement disorders, such as Parkinson’s disease.