Fig 1.
Illustration of the motivation and the objectives.
Our study is to recognize a set of users from their gait patterns using a encoder network and to provide an interpretable analysis of the network using the XAI method.
Table 1.
List of the related work on gait recognition.
Fig 2.
The insole used to collect the gait data.
Fig 3.
Illustration of computing the prototype of each sensing modality for a subject.
The prototypes (bold solid curves in the rightmost figures) of a subject are computed by averaging over all unit steps. For brevity, the L2 norms of are depicted.
Fig 4.
Illustration of the encoder–decoder architecture.
The encoder and decoder include three sub-encoders and sub-decoders for multimodal sensing.
Fig 5.
Illustration of the prototyping encoder–decoder with triplet loss.
The overall loss function is a linear combination of the multimodal triplet loss function and the prototyping loss function.
Fig 6.
Illustration of gait recognition using the trained encoder.
Here, unit step s*,u is recognized as that of the “green” subject, whereas unit step s*,w is rejected.
Fig 7.
Illustration of splitting the data into the training, known test, and unknown test sets.
Fig 8.
ACC as a function of γ and ν for a value of τ = −0.1.
A similar rate is represented as the same color with the maximum 1% difference, with the highest rates as yellow.
Fig 9.
Performance as a function of τ for fixed values of γ = 2.2, ν = 0.06, and λ = 1.0.
Fig 10.
Averaged relevance score heat maps and the their occluding positions for O1 ⋯ O5 of SA and LRP-ϵ.
In each heatmap, the x-axis and y-axis indicate features of each sensor and time-stamps of each unit step, respectively. (a) Common attribution maps. (Left: SA, Right: LRP-ϵ). (b) Occluding positions (O1, ⋯, O5) for all modal inputs. (Top: SA, Bottom: LRP-ϵ).
Fig 11.
Performance as a function of occluding position O1, ⋯, O5 by SA and LRP-ϵ for fixed γ = 2.2, ν = 0.06, τ = −0.1, and λ = 1.0.
(a) SA. (b) LRP-ϵ.