On the Learnability of Footwear Robustness: Footstep Recognition Under Unseen Conditions
Organizations: Movement and Biosignals Intelligence Lab (MoBIL), University of New Brunswick, Fredericton, NB, Canada
Abstract
Footstep patterns, captured using pressure or force-sensitive walkways during natural walking, offer a convenient and passive approach to biometric recognition. However, footwear variability poses a major challenge to practical deployment, as it impacts both the foot-sensor interface and the user's underlying gait mechanics. This study considers the particularly difficult scenario where users enroll under one fixed footwear condition but must later be recognized in arbitrary, unseen footwear. Using the StepUP-P150 dataset, we show that footwear-induced shift, while substantial and heterogeneous across individuals, contains shared structure that can be learned from a training population and transferred, zero-shot, to previously unseen users and footwear styles. We do this using BiStepNet, a novel architecture that decomposes spatial and temporal identity cues into independently trainable streams and outperforms compact 3D spatiotemporal baselines for cross-footwear recognition. We demonstrate that cross-footwear recognition performance scales systematically with the volume and diversity of footwear representation in the training set, with gains transferring to footwear styles not directly observed during training. Moreover, we show that auxiliary gait datasets with only partial modality overlap (incorporated through the architecture's decomposed design) can be leveraged to further improve robustness to unseen footwear conditions.
Figures & tables
| Dataset | IDs | Strides | Footwear | Other Covariates |
|---|---|---|---|---|
| Force plate (temporal representations only) | ||||
| GaitRec [ 24 ] | 211 | 7,755 | BF (0–1), PF (0–1) | Speed (3) |
| Gutenberg [ 25 ] | 350 | 8,819 | BF | — |
| Bialystok [ 26 ] | 324 | 13,702 | PF (1–7) | — |
| ForceID-A [ 7 ] | 184 | 5,327 | PF (1–2) | Speed (3), Days (2) |
| Pressure platform (spatial representations only) | ||||
| BF → BF | BF → SF | BF → PF | ||||
|---|---|---|---|---|---|---|
| Model | Rank-1 | EER | Rank-1 | EER | Rank-1 | EER |
| R(2+1)D CNN | 96.1 | 1.9 | 79.2 | 5.3 | 42.5 | 15.4 |
| 3D CNN | 95.8 | 2.3 | 79.2 | 5.9 | 44.6 | 15.5 |
| 3D CNN-GRU | 93.3 | 3.0 | 74.3 | 7.0 | 46.3 | 14.6 |
| BiStepNet a | 97.2 | 2.0 | 83.4 | 4.9 | 52.3 | 12.7 |
| BiStepNet b | 95.9 | 2.4 | 83.4 | 5.0 | 65.2 | 9.1 |
| BF → BF | BF → SF | BF → PF | |||||
|---|---|---|---|---|---|---|---|
| Encoder | Datasets | Rank-1 | EER | Rank-1 | EER | Rank-1 | EER |
| Temporal | Single | 74.9 | 8.1 | 44.3 | 15.4 | 45.2 | 15.5 |
| Multi | 77.0 | 7.8 | 46.5 | 14.6 | 48.6 | 14.7 | |
| Spatial | Single | 95.2 | 2.7 | 75.9 | 6.3 | 42.8 | 15.2 |
| Multi | 96.1 | 2.4 | 77.0 | 5.7 | 44.1 | 14.7 | |
| Fused | Single | 95.9 | 2.4 | 83.4 | 5.0 | 65.2 | 9.1 |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Layer | Channels | Hyperparameters | Output |
| 3D CNN (352K params) | |||
| Conv3D † | 1 32 | , | |
| Conv3D † | 32 64 | , | |
| Conv3D † | 64 128 | , | |
| AvgPool3D+FC | 128 256 | pool(1,1,1) | 256 |
| R(2+1)D CNN (364K params) | |||
Gait Recognition
Gait recognition identifies people from their walking patterns, using visual appearance and motion cues across cameras and other sensors.
See all 32 papers