A dataset of high-resolution plantar pressures for gait analysis across varying footwear and walking speeds

This paper introduces the UNB StepUP-P150 dataset, a large-scale, high-resolution collection of plantar pressure data from 150 individuals across varying walking speeds and footwear conditions, designed to advance research in biometric gait recognition, biomechanics, and deep learning.

The Gist

Imagine you are trying to teach a computer how to recognize a person just by looking at the way they walk. For a long time, scientists have done this by filming people from the side (like a security camera) or using special suits with sensors. But there's a problem: cameras can be blocked, and suits are expensive and uncomfortable.

What if we could recognize someone just by looking at the "fingerprint" their foot leaves on the ground?

This paper introduces a massive new tool called the UNB StepUP-P150 dataset. Think of it as the "ultimate library" of footprints for scientists to study. Here is a simple breakdown of what they did and why it matters.

1. The "Giant Pressure Mat" (The Instrument)

Imagine a hallway that is 12 feet long and 4 feet wide. Now, imagine the floor of that hallway isn't just wood or tile, but a giant, high-tech carpet made of 172,800 tiny pressure sensors.

• The Analogy: Think of it like a giant, super-sensitive piano keyboard. Every time someone steps on it, the sensors "play" a note, recording exactly how hard and where the foot pressed down.
• The Resolution: This mat is incredibly detailed. It's like having a camera that takes a photo of your foot with 4 pixels for every single square millimeter. This allows researchers to see tiny details, like how the pressure shifts from your heel to your toe, even through thick shoes.

2. The "Cast of Characters" (The Participants)

To make sure this library works for everyone, they didn't just ask a few young athletes to walk across the mat. They invited 150 different people to participate.

• The Diversity: The group was a mix of men and women, ranging from teenagers (19 years old) to seniors (91 years old). They included people of various heights, weights, and ethnic backgrounds.
• The Wardrobe Change: To make the data realistic, participants didn't just walk in one pair of shoes. They walked in:
  • Barefoot (or with socks).
  • Standard Sneakers (a generic pair provided by the researchers).
  • Two Pairs of Their Own Shoes (ranging from high heels and running shoes to heavy work boots and flip-flops).

3. The "Actors' Routine" (The Experiment)

Each person spent about an hour and a half on the mat. They didn't just walk in a straight line; they acted out different scenarios:

• The "Normal" Walk: Walking at their own comfortable pace.
• The "Slow-Down" Walk: Walking slowly and stopping abruptly (like approaching a security gate).
• The "Power Walk": Walking as fast as they could.
• The "Slow" Walk: Walking as slowly as possible.

They did this routine four times, changing their shoes each time. In total, the dataset contains over 200,000 individual footsteps. That's like having a library with 200,000 unique pages of footprints!

4. Why This is a "Game Changer"

Before this, the biggest libraries of footprints were small. Some had only 100 people, and many only had data for people walking in bare feet or one type of shoe.

• The Old Way: It was like trying to learn a language by reading a dictionary with only 50 words.
• The New Way (StepUP-P150): This is like getting the entire dictionary, plus a dictionary of slang, formal speech, and different accents.

Because the data is so huge and varied, it allows scientists to:

• Train Better AI: Computers can learn to recognize a person even if they are wearing heavy boots or walking slowly.
• Study Health: Doctors can use it to see how diseases (like Parkinson's) or injuries change the way we walk.
• Design Better Shoes: Shoe companies can use the data to see exactly how their designs affect pressure on the foot.

5. The "Safety Net" (Quality Control)

The researchers didn't just collect the data and leave it. They acted like strict editors.

• They used video cameras to watch the people walk and double-check the computer's work.
• They manually fixed mistakes (like if the computer thought a left foot was a right foot).
• They even added a "quality score" to every step, flagging any steps that looked weird (like if someone stumbled or tripped) so researchers could ignore them if they wanted.

The Bottom Line

The UNB StepUP-P150 is a massive, open-source treasure chest of footstep data. It's designed to help researchers build smarter security systems, better medical tools, and more comfortable shoes. By making this data public, the researchers are saying, "Here is the most complete map of human walking we have ever made; now, go explore it and discover something new."

Technical Summary

1. Problem Statement

Gait analysis is critical for biometrics, biomechanics, sports science, and rehabilitation. While traditional methods rely on video and motion capture, plantar pressure sensing offers deeper insights into footstep mechanics. However, the field is hindered by a lack of large-scale, publicly accessible datasets that capture high-resolution underfoot pressure data under diverse real-world conditions.

• Limitations of Existing Datasets: Current databases are often small, lack demographic diversity, use low-resolution sensors, or fail to account for confounding variables like walking speed and footwear types. Many rely on force plates (constrained data) or have limited sample sizes that do not represent real-world variability.
• The Gap: There is a need for a comprehensive dataset with high spatial resolution, a large number of participants, varied footwear, and multiple walking speeds to support advanced deep learning models and robust biometric recognition.

2. Methodology

Data Collection Setup

• Instrumentation: The study utilized a specialized 1.2m × 3.6m pressure-sensing walkway composed of a 2×6 grid of modular tiles.
  • Sensors: 172,800 piezoresistive sensors (240 × 720 grid) providing a spatial resolution of 4 sensors/cm² (5mm × 5mm).
  • Specs: 100 Hz sampling rate, 10 kPa sensitivity threshold, and a range up to 1,510 kPa.
  • Auxiliary: Seven 1080p RGB video cameras (20 fps) positioned at various angles for validation, and a flatbed scanner for digital shoe sole scans.
• Participants:
  • Cohort: 150 healthy individuals (74 male, 76 female) aged 19–91 years.
  • Diversity: Includes varied ethnicities (White, Asian, Others), body types (BMI 17–39), and foot sizes.
  • Exclusions: 30 participants were excluded due to hardware malfunctions, missing video data, or protocol deviations, leaving the final dataset with 150 subjects.
• Experimental Protocol:
  • Conditions: Each participant completed 16 walking trials (4 speeds × 4 footwear types) plus balance tasks.
  • Footwear: Barefoot (BF), Standard Shoes (ST - Adidas Grand Court 2.0), and two pairs of Personal Shoes (P1, P2) covering diverse categories (athletic, casual, boots, heels, etc.).
  • Walking Speeds: Preferred (natural), Slow-to-Stop (simulating security checkpoints), Slow, and Fast.
  • Duration: Approximately 24 minutes of walking data per participant, resulting in ~1,400 steps per person.

Data Processing & Preprocessing

The raw data (3D tensors of pressure over time) underwent rigorous processing:

1. Footstep Detection: Used connected pixel object detection and the SORT (Simple Online and Realtime Tracking) algorithm to identify and track individual steps across frames.
2. Labeling: Automated algorithms assigned labels for foot side (left/right), orientation, and completeness. These were manually verified by researchers using video and pressure visualizations.
3. Normalization Pipelines: To support diverse machine learning tasks, the dataset provides two distinct preprocessing pipelines:
   • Pipeline 1: Spatial rotation, zero-padding, translation, and temporal interpolation (101 frames). Retains foot size/weight info.
   • Pipeline 2: Adds spatial resizing (to a common sole dimension) and amplitude normalization (by mean total pressure). Removes foot size/weight info to focus on pressure patterns.
4. Quality Control: An automated R-score outlier detection method was applied to flag irregular gait patterns (e.g., stumbling, sensor noise), marking ~13.7% of footsteps as outliers.

3. Key Contributions

• UNB StepUP-P150 Dataset: The largest footstep database to date, containing over 200,000 footsteps from 150 participants. This surpasses the previous largest dataset (SFootBD) by more than tenfold.
• High Resolution: Offers the highest spatial resolution (5mm × 5mm) among public pressure-based gait datasets, capturing fine-grained pressure textures and shoe sole details.
• Comprehensive Covariates: Uniquely captures the interaction between gait and multiple variables:
  • Footwear: Barefoot, standard, and two distinct personal shoe types per subject.
  • Speed: Four distinct speeds including a "slow-to-stop" maneuver.
  • Demographics: Balanced sex distribution and a wide age range (19–91).
• Dual Data Formats: Provides both raw trial-by-trial data and preprocessed, step-segmented data in both Python (.npz) and MATLAB (.mat) formats.
• Rich Metadata: Includes detailed anthropometric data, shoe specifications, and step-level metadata (COP trajectories, GRF, bounding boxes, outlier flags).

4. Results & Technical Validation

• Data Quality: Manual inspection confirmed high accuracy in automated labeling (e.g., 99.7% accuracy for left/right foot detection). The R-score analysis successfully identified irregular steps, ensuring a clean dataset for training.
• Comparison with Existing Datasets:
  • Spatial Resolution: StepUP-P150 (5mm) significantly outperforms CASIA-D and CAD WALK (approx. 7.6mm × 5mm) and custom systems like SFootBD (27mm sensors).
  • Spatiotemporal Parameters: The large walkway (3.6m) allowed for the capture of 4–6 consecutive steps, enabling reliable extraction of gait speed, cadence, step length, and toe-out angle.
  • Variability: The dataset exhibits higher variability in gait parameters compared to smaller datasets, reflecting the complexity of real-world populations and diverse footwear.
  • Biomechanical Consistency: Ground Reaction Force (GRF) and Center of Pressure (COP) waveforms align with established biomechanical trends (e.g., increased peak GRF at faster speeds).

5. Significance

• Benchmark for Biometrics: Establishes a new standard for gait recognition research, particularly for testing model robustness against covariates like footwear changes and speed variations.
• Deep Learning Advancement: The high resolution and large volume of data enable the training of complex Convolutional Neural Networks (CNNs) and 3D deep learning models that were previously impossible with smaller, lower-resolution datasets.
• Biomechanical Research: Facilitates the study of how external factors (shoe type, speed) and internal factors (age, gender, body mass) influence plantar pressure distribution.
• Standardization: Provides a unified, open-access resource that allows researchers to compare algorithms fairly, promoting the standardization of gait analysis methodologies.
• Future Applications: Supports research in normative gait modeling, anomaly detection (e.g., for fall risk or neurological conditions), and the development of adaptive footwear or assistive devices.

In summary, the UNB StepUP-P150 dataset fills a critical void in gait research by providing a massive, high-fidelity, and diverse resource that bridges the gap between controlled laboratory studies and real-world gait variability.