The Gaitprint: Identifying Individuals by Their Running Style

Abstract

Recognizing the characteristics of a well-developed running style is a central issue in athletic sub-disciplines. The development of portable micro-electro-mechanical-system (MEMS) sensors within the last decades has made it possible to accurately quantify movements. This paper introduces an analysis method, based on limit-cycle attractors, to identify subjects by their specific running style. The movement data of 30 athletes were collected over 20 min. in three running sessions to create an individual gaitprint. A recognition algorithm was applied to identify each single individual as compared to other participants. The analyses resulted in a detection rate of 99% with a false identification probability of 0.28%, which demonstrates a very sensitive method for the recognition of athletes based solely on their running style. Further, it can be seen that these differentiations can be described as individual modifications of a general running pattern inherent in all participants. These findings open new perspectives for the assessment of running style, motion in general, and a person's identification, in, for example, the growing e-sports movement.

Keywords: attractor method; human cyclic motion; individual locomotion; recognition; running quality.

Figure 1

Two-dimensional depiction of the three-dimensional recognition horizon (red) compared to an attractor of the same (subject A, blue) and a different subject (B, green).

Figure 2

Example for two normally distributed probability curves (orange = between subject comparison; blue = within subject comparison; thin and thick dashed lines in orange and blue symbolize mean $(\mu )$ and standard deviation (δ); T is the border for rating different (left of T) and same (right of T); α-area = probability of not identifying a subject; P-area = probability of a false positive identification).

Figure 3

Recognition percentages presenting same-subject values (green dots in (a) & black crosses in (b)) and different-subject (worst-case) values (grey dashes in (a,b)).

Figure 3

Recognition percentages presenting same-subject values (green dots in (a) & black crosses in (b)) and different-subject (worst-case) values (grey dashes in (a,b)).

Figure 4

False positive detection probability chart displaying the individual percentages of all subjects. Squares = Super attractor versus the first run of the same person and filled triangles = Super attractor versus the second run of the same person. When probability is zero, squares are partly hidden by filled triangles.

Figure 5

δM mean of the morphing analysis with according standard deviations sorted by subjects.

Figure 6

Four examples (subject 10 = orange, 9 = green, 2 = blue, 5 = yellow) showing different progressions of transient oscillations which have subsided after minute 10.

Figure 7

Display of the recognition horizon (blue and orange) with a data set from a differing subject over all three axes (a = x-axis, b = y-axis, c = z-axis). The recognition rate was 44%.