[Paper] First International StepUP Competition for Biometric Footstep Recognition: Methods, Results and Remaining Challenges

Source: arXiv - 2602.11086v1

Overview

The First International StepUP Competition put the newly released UNB StepUP‑P150 dataset to the test, challenging teams to build biometric footstep‑recognition systems that can verify a person’s identity from the pressure pattern of their steps. By gathering 23 teams from academia and industry, the competition demonstrated that deep‑learning models can now achieve double‑digit accuracy on a task that was previously hampered by tiny, homogeneous datasets.

Key Contributions

• Largest public footstep‑pressure dataset (UNB StepUP‑P150) made available for research and commercial prototyping.
• Benchmark competition that defines a realistic verification protocol: limited enrollment data, heterogeneous test conditions (different shoes, speeds, surfaces).
• State‑of‑the‑art baseline: the winning solution (Saeid_UCC) reached an Equal Error Rate (EER) of 10.77 % using a Generative Reward Machine (GRM) optimization framework.
• Comprehensive analysis of failure modes, especially the impact of unseen footwear on recognition performance.
• Open‑source reference implementations and a public leaderboard that will continue to drive progress after the event.

Methodology

1. Dataset preparation – StepUP‑P150 contains >150 k high‑resolution pressure maps from 150 participants, captured across multiple sessions, shoe types, and walking speeds.
2. Enrollment vs. verification split – Teams received a small, homogeneous enrollment set (same shoes, similar speed) and were asked to verify against a much larger, deliberately varied test set.
3. Modeling approaches – Most submissions leveraged convolutional neural networks (CNNs) or hybrid CNN‑RNN architectures to learn spatial‑temporal foot‑pressure signatures. The winning team introduced a Generative Reward Machine that treats verification as a sequential decision problem, optimizing a reward that balances false‑accept and false‑reject penalties.
4. Evaluation metric – Equal Error Rate (EER) was used, where the false‑accept rate equals the false‑reject rate; lower EER indicates better biometric performance.

Results & Findings

• Overall trend: Deep models consistently outperformed classical handcrafted‑feature baselines (which hovered around 20 % EER).
• Footwear sensitivity: All teams saw a sharp EER increase (≈ 5–7 % absolute) when the test shoes differed from enrollment shoes, confirming footwear as the biggest open challenge.
• Speed variation: Moderate walking‑speed changes had a smaller impact (≈ 1–2 % EER rise), suggesting that temporal normalization techniques are already effective.
• Data efficiency: Even with only a few enrollment samples per user, the top models achieved sub‑15 % EER, indicating that the dataset’s richness enables strong generalization.

Practical Implications

• Secure access control: Footstep biometrics could complement badge or facial systems in high‑security facilities where hands‑free, covert verification is valuable (e.g., data centers, labs).
• Smart building automation: Personalized lighting, HVAC, or elevator routing could be triggered by recognizing occupants as they walk, improving energy efficiency and user experience.
• Healthcare & rehabilitation: Continuous monitoring of gait patterns for patient identification and fall‑risk assessment becomes feasible with low‑cost pressure mats.
• Retail analytics: Anonymous footstep signatures can help differentiate repeat customers without invasive cameras, enabling privacy‑preserving personalization.
• Edge deployment: The competition’s focus on limited enrollment data mirrors real‑world constraints on edge devices (e.g., IoT floor sensors), encouraging lightweight model designs that can run on microcontrollers.

Limitations & Future Work

• Footwear generalization remains the dominant bottleneck; future research must explore domain‑adaptation, synthetic shoe‑style augmentation, or multimodal fusion (e.g., combining pressure with inertial data).
• Dataset diversity: While StepUP‑P150 is large, it still reflects a single geographic region and a narrow age range; broader demographic coverage would improve fairness and robustness.
• Real‑time latency: The winning GRM pipeline, though accurate, is computationally heavy; optimizing inference speed for on‑device deployment is an open engineering challenge.
• Security analysis: The study did not evaluate spoofing attacks (e.g., fabricated pressure maps); adversarial robustness will be crucial before commercial rollout.

The StepUP competition has set a solid benchmark for biometric footstep recognition and highlighted a clear roadmap for turning this emerging sensor modality into practical, secure, and privacy‑respectful applications.

Authors

• Robyn Larracy
• Eve MacDonald
• Angkoon Phinyomark
• Saeid Rezaei
• Mahdi Laghaei
• Ali Hajighasem
• Aaron Tabor
• Erik Scheme

Paper Information

• arXiv ID: 2602.11086v1
• Categories: cs.CV, cs.LG
• Published: February 11, 2026
• PDF: Download PDF