Technology
Traditional technologies are being used to perform human gait analysis to identify targets in circumstances where the environment is not supportive, for example where there is a lack of facial images when using facial recognition, or losing “target behaviour” after crossing checkpoints, etc. Radio frequency based human gait can be used in such situations as radio frequency is always operational. For example, the use of traditional video surveillance devices is limited in highly classified work environments or sensitive private compounds such as changing rooms and toilets. Radio frequency imaging tends to be less invasive and obtrusive as compared to video imaging. Such motion detection methods do not involve the use of cameras, contact sensors, or other intrusive equipment which means privacy is respected and time is being saved.
We create specialised radio frequency access points supporting Doppler effect which uses radio disturbances (frequency amplitude, phase over time) to track human gait and location without additional devices. Beyond capturing the raw radio frequency signals, we apply advanced digital signal processing to create the most robust deep machine learning algorithms for recognising human activities, their locations and creating their unique radio frequency biometric signatures for tracking purposes. The real-time analysis of “target behaviour” is also much more efficient when using radio to monitor “targets”.
The Science Behind What We Do
Channel state information (CSI) measurements are first collected from radio frequency devices. A collection of data pre-processing schemes is then applied to sanitise and calibrate the noisy and erroneous CSI data samples. Next, a multiple-layer deep convolutional neural network, is developed to automatically learn the salient features from the pre-processed CSI data samples. The extracted features constitute a latent representation for each person’s gait identity such that one person can be distinguished from another. Using the latent biometric representation, a multi-class classifier is adopted to achieve accurate user identification. Extensive experiments in typical indoor and outdoor environments are conducted to demonstrate the effectiveness of our system.