Gait-Based Person Identification and Intruder Detection Using mm-Wave Sensing in Multi-Person Scenario

Each person’s walking pattern or ‘Gait’ is a specific and unique feature. Using millimeter wave or mm-wave sensors to identify unique individuals by their gait is an emerging biometric solution with applications in diverse fields like security surveillance, health monitoring, and automated access control.

However, the existing devices face restrictions due to body contact and privacy invasive devices. Hence, this study proposes MGait, a powerful tool for indoor based, multi-person identification and intruder detection methods based on how they walk, or ‘Gait Identification.’

The gait is captured as m-D (micro-doppler) signatures over low-cost mm-wave radar sensors. The mm-wave radar for recognizing human movement facilitates long-distance detection, light insensitivity, and privacy preservation.

The untrained gait is recorded in an indoor corridor environment at 25fps by a TI 77 GHz frequency-modulated continuous-wave (FMCW) radar. Its high range resolution and velocity provide possibilities for separating multiple subjects and assist in acquiring vast data for extracting subtle and discriminative gait features.

In scenarios with multiple subjects walking simultaneously at different speeds and gaits, the range-Doppler (R-D) space is employed to detect and track each person’s walk frame by frame, extracting their respective m-D gait signatures.

This raw data is received and stored in data cubes and then sent into signal preprocessing pipelines, which eliminates signal noises and reflections from static targets. The backscattered signals from multiple concurrent subjects are superimposed together in a cluster.

These multiple target clusters are then associated frame by frame to form coherent and continuous gait tracks for each individual by implementing a Karman filter, which applies a ‘predict-associate-update’ loop using nearest neighbor data association (NNDA).

The radar echo generates a unique m-D signature, as the gait is unique for each person. However, recognizing the different gaits more than the manual method is required as the differences between these gaits are very subtle. Hence, the system is trained with deep learning methods.

An open-set identification Network is built and trained by a deep learning loss model. The network is processed by a Large margin Gaussian Mixture (L-GM) Loss to train the model to learn highly discriminative feature representations. A density-based spatial clustering of application with noise (DBSCAN) algorithm classifies highly clustered data points as known users and scattered points as outlier points, i.e., intruders.

Using the Class-posterior probability, users can be identified as known persons or intruders, who can then be rejected by setting appropriate threshold parameters in the loss function. Extensive experimental results verify that the MGait system with 77GHz FMCW Radar achieves an accuracy of 88.59% in identifying up to 5 subjects, including 1 known user and 4 intruders, who are freely and simultaneously moving in a corridor space.

Great accuracy with multiple functionalities of this mm-wave sensing system makes it fit for usage in varied fields like advanced spatial access systems, body-language-based behavior studies, and health monitoring. The researchers plan to foray into an in-depth analysis to improve the MGait performance further by using 2-dimensional antenna arrays and/or multi-static radar systems.