It's on My Radar

Using a miniaturized radar-on-a-chip and 3D-printed reflectors, it is possible to identify objects and create interactive environments.

Nick Bild
16 days agoMachine Learning & AI
Radar-based object identification (📷: T. Gunasekaran et al.)

Radar has long been used in tracking and ranging applications, but with recent advances in device miniaturization, a whole new batch of use cases has started to appear. These applications range from gesture detection to breathing rate sensing, material identification, and object tracking. These use cases are certain to expand in the future as miniature radar units have become more common features in smartphone designs. One area that has not previously been well explored is radar-based object identification and tracking using custom radar reflectors. This area has a lot of potential to alter and enhance the ways in which we interact with everyday objects.

Recognizing the significance of this gap in knowledge, a group of researchers centered at The University of Auckland have created RaITIn, a radar-based system for identification. RaITIn works with embedded radar reflectors at a tabletop scale to enhance human interactions with real world objects. A machine learning algorithm was employed to decipher the meaning of the raw radar reflections.

The researchers designed their prototype around Google’s Soli miniature radar system that was designed especially for use in small consumer devices like smartphones and tablets. Soli is a five millimeter wavelength radar unit that operates in a frequency range from 57 GHz to 63 GHz, and has a range-bin resolution of 2.5cm. This sensor was positioned 40cm above the surface of a table, where it was able to capture reflections from any part of the table.

The radar reflectors were modeled after the traditional passive octahedral radar reflectors that are commonly found on boats. Various shapes of reflectors, which were designed to exhibit different reflection patterns, were 3D printed, then covered with aluminum tape before being placed into an outer casing. Since radar waves can penetrate many materials, such as plastics, the reflectors can be transparently embedded within other objects. In total, six custom radar reflectors were produced with radii of 2cm, 3cm, 4cm, 5cm, 6cm, and 7cm. To allow for more than six identifiable objects, these reflectors can be stacked in various patterns, which produces a unique radar reflection fingerprint.

The primary signal generated by the Soli sensor is the range-Doppler, which reveals the energy intensity of the reflections, radial distance, and target velocity. This signal was cleaned up by using the Soli C++ SDK to reduce noise coming from the stationary objects on the table. After a bit more preprocessing of the data, it was fed into a machine learning algorithm to help with interpretation. In particular, an AdaBoost classifier, in conjunction with a decision tree was selected to translate radar reflections into predictions of which specific reflector is present. A grid search was used to choose the best hyperparameters for the model, then it was trained. This method was able to clearly distinguish between reflectors with a classification accuracy of 99.17% being found.

With a working setup in place, the researchers got to work building some example applications to demonstrate its capabilities. In one demonstration, an interface with a button, toggle switch and slider was created using the reflectors, which could be embedded within nearly any everyday object. The button, in particular, is interesting in that it uses a pair of reflectors, and when the reflectors come closer together, a button press is assumed to have occurred. In another demonstration, the team developed a 2D Pokemon fighting game in which the characters in the game were made of paper. As the Pokemon evolve and gain experience, it is represented by stacked reflectors.

The present prototype can only recognize a single radar reflector at a time in a given interaction space, which somewhat limits the system’s possibilities. The researchers are, however, working on multi-object identification for a future version. Perhaps the biggest limitation is that RaITIn cannot identify radar reflections if a hand is present in the interaction space. That is a big drawback for a system that was built for human interaction. Limitations aside, RaITIn is a big first step towards using new, low cost radar sensors for a whole new class of applications.

Nick Bild
R&D, creativity, and building the next big thing you never knew you wanted are my specialties.
Latest articles
Sponsored articles
Related articles
Latest articles
Read more
Related articles