Action Recognition without a Camera Using Low Dimensional Micro-Doppler Signatures

Unmet Need
Despite the desire for high quality action recognition and gesture tracking in a myriad of fields, recognition tools are limited in several aspects. Reliance on cameras and visual cues limit the ability of action recognition software to process motion in anything other than a 2 dimensional plane. Utilizing proximity sensors or multi-camera optical arrays some systems have found workarounds to this forced 2 dimensionality, but these approaches have been mediocre at best with low accuracy and onerous requirements placed upon the subject. Furthermore, these systems universally rely on the presence of cameras, meaning their functionality is limited in low light or low visibility situations.

Technology Overview
Utilizing low dimensional micro-Doppler frequencies in congruence with modern machine learning techniques, Johns Hopkins Researchers have discovered a new method for highly accurate, 3 dimensional action recognition that does not require a camera. Modeled off sonar based visualization tools utilized by horseshoe bats, this technology trains deep learning networks attached to low-dimensional micro-Doppler emitters to recognize how human movement impacts those frequencies. Because it is not reliant on visual cues, this system can function in low light and poor visibility. Additionally, this system is capable of identifying 3 dimensional gestures without the need for complex optical arrays, and with much more freedom of movement on the part of the subject. This system represents the first example of non-optical based gesture recognition and opens the door to diversifying the technology utilized in gesture tracking.

Stage of Development
Proof of concept. Preliminary testing

Publications
Bio-Inspired Human Action Recognition With A Micro Doppler Sonar System

A Generative Model for Doppler-Modulations of Action Sequences
 

Category(s):