Non-contact, non-visual, multi-person hallway gait monitoring

Abstract

This paper presents a multi-person gait monitoring system designed for efficient operation in cluttered environments. The system demonstrates robust capabilities in tracking multiple closely spaced individuals and accurately extracting the walking speed, even in the presence of others. We address two significant challenges, including enhancing radar resolution and mitigating multipath effects in cluttered settings. Our method shows remarkable accuracy, with a maximum error of 0.33 m/s and a minimum of 0.005 m/s, as validated through 25 walking tests in a bedrest study. Its adaptability makes it a valuable clinical tool, offering insights for predicting underlying health issues in older adults.

Keywords: Activity recognition; Autonomous systems; Multi-person gait monitoring; Sequential Deep learning; mm-wave radar.

Fig. 1

Participants’ walking speed in the BR study obtained by radar empowered by the proposed STA algorithm (brown) and a stopwatch (blue).

Fig. 2

Repeated measures Bland-Altman plot comparing the walking speed calculated from the stopwatch and radar data. A given colour represents a unique individual. The solid and dashed horizontal lines represent the bias and 95% limits of agreement.

Fig. 3

Comparison of walking speed obtained by stopwatch before bedrest (PRE), the first day (R + 1), third day (R + 3) and 4 weeks (R + 4wk) after getting out of bed. A given colour represents a unique participant. *Different () from PRE.

Fig. 4

Schematic of the multi-person hallway walking tests setup. The minimum and the maximum relative angle between the radar position and the hallway walls are 1 and 2, respectively.

Fig. 5

Multi-person hallway walking tests (a) a snapshot of the video in the middle of walking and (b) corresponding detected subjects’ clusters and ghosts.

Fig. 6

Proposed algorithm for multi-person gait monitoring.

Fig. 7

Details of the proposed STA algorithm.

Fig. 8

Representation of two walking subjects in a walking environment (a) range-azimuth heatmap of frame #1 obtained by the Capon beamformer, and (b) a snapshot of the video at the beginning of walking tests at frame #1.

Fig. 9

Representation of two walking subjects in the hallway environment (a) detected points by CFAR in frame #1 (b) clusters obtained by DBSCAN in frame #1.

Fig. 10

Representation of two walking subjects in a walking environment (a) seven detected clusters after performing the proposed 1D-Azimuth clustering in frame #1 (b) associated seven parents to the seven detected clusters in frame #1.

Fig. 11

Details of the proposed STA algorithm (a) the procedure of allocating a new cluster that meets the criteria to previous classes and (b) the procedure of creating a new class for a new cluster that does not meet the condition to be allocated to the previous classes.

Fig. 12

All detected clusters throughout the entire walking lap.

Fig. 13

Subjects’ clusters detected by the proposed STA algorithm.