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

arXiv:2502.17244 · cs.CV, cs.LG · Submitted 2026-08-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "A dataset of high-resolution plantar pressures for gait analysis across varying footwear and walking speeds".

Jane: The paper was written by Robyn Larracy, Angkoon Phinyomark, Ala Salehi, Eve MacDonald, Saeed Kazemi et al. from University of New Brunswick.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title and First Impressions: Tom: Welcome back to the show, everybody. Today we’re cracking open a brand new paper from arXiv, and it’s called “A dataset of high-resolution plantar pressures for gait analysis across varying footwear and walking speeds.” Jane, I have to say, just reading that title got me excited, because we finally have a big, public dataset for foot pressure.

Jane: Tom, I’m right there with you. Plantar pressure just means the pressure under your foot, and this paper is all about capturing that in incredible detail. They used a walkway that’s covered in sensors, and it’s not just a tiny mat either. This thing is over three meters long, so people can take multiple natural steps in a row.

Tom: Right, and that’s the part that got me. Most of the older datasets, they had people stepping on a single plate or a very short mat, so you’d get maybe one or two footsteps. Here, they’ve got a one point two by three point six meter grid, which means they’re catching four to six steps per pass. That’s a huge leap forward for studying how people actually walk.

Jane: And the resolution is wild. We’re talking about four sensors per square centimeter. That’s like having a tiny grid of pressure detectors every five millimeters. For comparison, some of the older public datasets had sensors that were maybe five times bigger, so you’d lose all the fine detail of how your heel strikes or how your toes push off.

Tom: The team behind this is from the University of New Brunswick, and they went all in. They recruited one hundred fifty people, and they had them walk under four different footwear conditions—barefoot, a standard shoe they provided, and then two pairs of the participant’s own shoes. Plus, they varied the walking speed: preferred, slow, fast, and even a slow-to-stop condition.

Jane: That slow-to-stop one is clever, because that mimics real life, like when you’re approaching a door or a security checkpoint. You don’t just walk at a constant speed all day. So having that in the dataset makes it much more realistic for building systems that work in the real world.

Tom: And the sheer volume of data is staggering. Over two hundred thousand footsteps. The previous biggest dataset of this kind had about twenty thousand so this is a tenfold increase. That’s the kind of scale you need if you want to train modern deep learning models, which are hungry for data.

Jane: Exactly. And they didn’t just dump raw sensor readings on us. They also provide preprocessed data, where each footstep is cut out, aligned, and normalized. That saves researchers weeks of work just cleaning the data, so they can jump straight into the science.

Tom: I love that they thought about the user. They give you both raw trial files and these segmented footsteps, plus a spreadsheet with all the participant demographics. So you can study how age, sex, body size, or shoe type changes the pressure patterns.

Jane: And that’s the hook for our next segment, because it’s not just about having a big pile of data. It’s about what you can actually do with it. We’re going to dig into the methods and how they made sure all those footsteps were labeled correctly.

Tom: Stay with us, because this dataset could change how we think about gait recognition and biomechanics.

Summary and Methodology: Tom: Welcome back. We’re still on “A dataset of high-resolution plantar pressures for gait analysis across varying footwear and walking speeds,” and Jane, I want to get into the nitty-gritty of how they actually built this thing, because the summary in the paper is dense.

Jane: It is, but the core idea is simple. They set up this long walkway with pressure tiles, and they had people walk back and forth for ninety seconds at a time. But the clever part is how they turned that continuous stream of pressure frames into individual, labeled footsteps.

Tom: Right, because you can’t just look at one frame and say, "that’s a foot." They used a tracking algorithm called SORT, which is usually used for tracking objects in video. Here, they applied it to track blobs of pressure across time, so they could follow a foot from heel strike to toe-off.

Jane: And that gives you a three dee bounding box for each step, with time, height, and width. That’s the raw extraction. But then they had to figure out which foot it was, left or right, and which direction the person was walking. They used the center of pressure trajectory for that, which is basically the weighted average of where the pressure is at each moment.

Tom: I read that they got the left/right classification right ninety-nine point seven percent of the time with an automated algorithm. But they didn’t stop there. They had two human reviewers visually inspect every single footstep. That’s a massive quality control effort, and it’s exactly what you need for a benchmark dataset.

Jane: They even used the video recordings they took from seven cameras around the room to double-check uncertain cases. So if the algorithm flagged something weird, they could look at the video and see what actually happened. That’s how you catch the edge cases, like when someone shuffles their feet during the slow-to-stop trials.

Tom: And they thought about what happens when people step partially off the mat. Those are called incomplete footsteps, and they’re flagged in the metadata. They also flagged standing footsteps, which happen during that slow-to-stop condition, and they used an outlier detection method to flag any steps that just look abnormal.

Jane: That outlier detection is interesting. They used something called an R-score, which basically compares each footstep to the median footstep in that trial. If a step is too different in terms of size, duration, or force profile, it gets flagged as an outlier. That way, researchers can easily filter out the noisy steps if they want to train a clean model.

Tom: And here’s the thing I really appreciate. They provide two different preprocessing pipelines. Pipeline one keeps the foot size and rotation information, so you can still use that for recognition. Pipeline two resizes everything to a common foot size and normalizes the amplitude, which is better for comparing pressure patterns across people.

Jane: Because there’s no single right way to preprocess pressure data. If you resize everyone’s foot to the same size, you lose information about foot size, which might be useful for identifying someone. But if you keep the original size, you can’t directly compare pressure at the same anatomical location across people. So they give you both options and a script to generate your own.

Tom: That flexibility is huge. It means researchers can test different preprocessing choices and see which one works best for their specific task, rather than being locked into one format. And that brings us to the next segment, because I want to talk about what improvements this enables and what new research doors it opens.

Jane: Absolutely, because a dataset like this isn’t just a bigger version of what we had. It’s a different beast entirely.

Improvements and Implications: Tom: So we’ve established that this dataset is big and well-annotated. But Jane, the real question is, what can we do with it that we couldn’t do before? What improvements does “A dataset of high-resolution plantar pressures for gait analysis across varying footwear and walking speeds” actually enable?

Jane: The biggest one is that we can finally train deep learning models properly. Before, with only twenty thousand footsteps, you’d overfit quickly. Now, with two hundred thousand you can train convolutional neural networks or transformers to recognize individuals just from their foot pressure patterns, and actually expect them to generalize.

Tom: And the diversity of the data matters just as much as the volume. They have people from nineteen to ninety-one years old, a balanced mix of men and women, and a wide range of body sizes and foot shapes. That’s critical for building biometric systems that don’t fail on people who aren’t in the training set.

Jane: But it’s not just about recognition. This dataset lets us study how footwear changes your gait. They have the same person walking barefoot, in a standard shoe, and in two of their own shoes. So you can isolate the effect of the shoe itself versus the person’s natural gait.

Tom: That’s a point I want to push on, because the paper shows some really interesting visualizations. They have peak pressure images, and you can see how a steel-toe work boot creates a completely different pressure pattern than a stiletto heel. But the person underneath is the same, so you can start to separate the identity component from the footwear component.

Jane: And that’s the hard problem in gait recognition. When someone changes shoes, their pressure pattern changes, but there’s still something consistent about how they walk. With this dataset, you can actually train models to be invariant to footwear, which is what you need for a real security system.

Tom: There’s also the walking speed variation. They have slow, preferred, fast, and that slow-to-stop condition. The paper shows that walking speed changes the ground reaction force profile significantly, especially the peak force at heel strike. So if you want a system that works in the real world, it has to handle people walking at different speeds.

Jane: And the implications go beyond security. This could be used in rehabilitation. If you have a patient with a foot injury, you can track their pressure patterns over time and see if they’re improving. Or you could use it to design better shoes, because you can see exactly where the pressure is concentrated.

Lu: I’d like to jump in here, Tom. The scale of this dataset also opens up the possibility of generative models. You could train a model to synthesize realistic foot pressure patterns for people who aren’t in the dataset. That could be used to augment training data even further, or to simulate how a new shoe design would affect gait before you even build a prototype.

Meng: And from an engineering standpoint, the fact that they provide both raw and preprocessed data is a godsend. We can test our own preprocessing pipeline against theirs and see if it makes a difference. Plus, the metadata includes spatiotemporal parameters like step length and step width, which are useful for feature engineering in traditional machine learning.

Tom: That’s a great point, Meng. The paper even compares step length and step width across datasets, and they show that the CASIA-D dataset had participants that were almost perfectly separable using just those two features. But in this new dataset, there’s much more overlap, which means the recognition problem is harder and more realistic.

Jane: So it’s not just a bigger dataset. It’s a harder dataset, which is exactly what we need to push the field forward. And that leads us to our final thoughts.

Conclusion: Tom: We’ve spent the whole show on “A dataset of high-resolution plantar pressures for gait analysis across varying footwear and walking speeds,” and I think we’ve only scratched the surface. Jane, what’s the big takeaway for our listeners?

Jane: The big takeaway is that this dataset is a game-changer for anyone studying how we walk. It’s the largest, most detailed public dataset of underfoot pressure ever released, with one hundred fifty participants and over two hundred thousand footsteps. It covers multiple footwear types and walking speeds, and it’s meticulously annotated.

Tom: And it’s not just for biometrics. It’s for biomechanics, sports science, rehabilitation, shoe design, and even generative AI. The fact that they provide raw data, preprocessed data, and a flexible preprocessing script means that researchers can adapt it to almost any task.

Jane: I also want to highlight the effort they put into quality control. Two human reviewers checked every footstep, and they used video to confirm uncertain labels. That level of care is rare, and it makes the dataset trustworthy as a benchmark.

Tom: And the authors made it easy to use. It’s available on the Federated Research Data Repository, with both Python and MATLAB formats. So whether you’re a deep learning researcher or a biomechanist, you can start using it right away.

Lu: I’ll add that this dataset will likely become the standard benchmark for footstep recognition. The previous datasets were too small to train modern models, so this fills a critical gap. I expect to see a wave of new papers using this data within the next year.

Meng: And from a practical standpoint, the inclusion of the slow-to-stop condition is brilliant. That’s the exact scenario you’d encounter at a secure access point, so it makes the dataset directly relevant for deployment.

Tom: Alright, we’ve said our goodbyes to this paper. It’s been a pleasure, and we’re already looking forward to the next one. Thanks for listening, and we’ll see you on the next episode.

Jane: Goodbye, everyone. Keep walking, and keep measuring.

Robyn Larracy, Angkoon Phinyomark, Ala Salehi, Eve MacDonald, Saeed Kazemi, Shikder Shafiul Bashar, Aaron Tabor, Erik Scheme

University of New Brunswick

cs.CV, cs.LG

Submitted: 2026-08-12

Journal ref: Scientific Data 12 (2025) 1415

DOI: 10.1038/s41597-025-05792-1

Code: https://github.com/UNB-StepUP/StepUP-P150

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 72/100

The gist: Preferred Speed (W1), Slow-to-Stop (W2), Slow (W3), and Fast (W4).

Terminology

Summary

Summary

The paper introduces the UNB StepUP-P150 dataset, described as a footStep database for gait analysis and recognition using Underfoot Pressure, including data from 150 individuals. This dataset comprises high-resolution plantar pressure data (4 sensors/cm2) collected using a 1.2m by 3.6m pressure-sensing walkway. It contains over 200,000 footsteps from participants walking with various speeds (preferred, slow-to-stop, fast, and slow) and footwear conditions (barefoot, standard shoes, and two personal shoes).

The motivation for the dataset is that underfoot pressures during walking remain underexplored due to the lack of large, publicly accessible datasets. While other gait databases exist, the majority use video-based systems, motion capture technologies, and wearable devices. Databases using floor sensors "generally collect data from force plates, as seen in the GaitRec, Gutenberg, and ForceID A datasets, which provide more constrained and less comprehensive information than the complex spatio-temporal data offered by emerging high-resolution pressure sensors. Conversely, databases emphasizing high-resolution underfoot pressure data such as the CASIA-D, SFootBD, CAD WALK, and UoM-Gait-69 datasets, usually address a limited scope, often with little or no consideration of covariates that may confound performance, and generally involve small sample sizes."

The dataset was collected over 18 months at the University of New Brunswick. A total of 180 individuals participated, but 30 were excluded due to experimental deviations (N = 2), missing data (N = 4), hardware malfunction (N = 15), and no accompanying video data to confirm annotations (N = 9). The final dataset includes 150 participants. The demographics include 74 males and 76 females with an average age of 34 years, spanning 19 to 91 years. The dataset maintains an approximately balanced sex/gender distribution and includes both younger and older adults. Most participants identify as White (N = 106, 71%), with additional identifications as Asian (N = 36, 24%), Other/Multiple (N = 6, 4%), and Unknown/Not Specified (N = 2, 1%). The dataset includes participants with height ranging from 151 to 196 cm, weight between 46 and 148 kg, BMI spanning 17 to 39 kg/m2, foot length measuring 20 to 30 cm, foot width between 7 and 11 cm, and UK shoe sizes from 4 to 12.5.

The instrumentation consisted of a 2 × 6 grid of commercial pressure sensing tiles developed by Stepscan Technologies Inc. The runway has a total active recording area of 1.2 m × 3.6 m, allowing for the collection of multiple consecutive steps (typically 4-6 steps). Each tile is 60 cm × 60 cm and contains a 120 × 120 sensor grid, corresponding to a spatial resolution of 4 sensors/cm2. Collectively, the grid comprises 172,800 piezoresistive sensors (i.e., 240 × 720), each with a threshold sensitivity of 10 kPa, sensing range of 1,510kPa (10 kPa resolution), and 100 Hz sampling rate. Seven RGB video cameras were also installed for visual verification, and a flatbed scanner was used for digital scans of personal footwear.

The experimental protocol involved a 90-minute session per participant. After a 30-minute preparation, participants performed three separate 30-second balance tests for each footwear type (both feet, left foot, right foot). The walking experiment involved sixteen 90-second walking trials, featuring four different footwear conditions, each executed at four distinct walking speeds. The four walking speeds were: Preferred Speed (W1), Slow-to-Stop (W2), Slow (W3), and Fast (W4). The four footwear conditions were: Barefoot (BF), Standard Shoe (ST) - a common pair of Adidas Grand Court 2.0 sneakers, and two pairs of Personal Shoes (P1 and P2). The average preferred walking speed was 1.12 m/s, fast was 1.45 m/s (26% faster), and slow was 0.83 m/s (29% slower). The order of footwear and walking speed conditions was randomized for each participant.

Data processing involved several steps. For balance trials, recordings were cropped to a region of interest and rotated to a common orientation, resulting in a 3D tensor of 3000 frames × 180 pixels × 180 pixels. For walking trials, footstep detection used a simple connected pixels object detection technique with a 10 kPa threshold, morphological operations, and SORT (Simple Online and Realtime Tracking) for tracking objects across frames. This produced 3D bounding boxes with dimensions (time, height, width) for each step.

The dataset provides two preprocessing pipelines. Pipeline 1 includes spatial rotation, spatial zero padding, spatial translation, and temporal interpolation, resulting in footsteps standardized to 101 frames with dimensions of 75 × 40 pixels. Pipeline 2 builds upon Pipeline 1 by adding spatial resizing and amplitude normalization using the mean total pressure, resizing soles to 70 pixels in length and 25 pixels in width.

Labels were derived algorithmically and confirmed through manual inspection. The Side label (left/right) was determined using a pixel counting method outlined by MacDonald et al. The Orientation label was determined by analyzing the center of pressure (COP) trajectory. Incomplete footsteps were flagged if they fell partially outside the tile grid or recording. Standing footsteps during slow-to-stop trials were identified by analyzing the slope of the anteroposterior COP. An automated outlier detection method using the R-score was also applied, with footsteps having an R-score of 2.0 or higher flagged as outliers, totaling 29,511 flagged footsteps throughout the entire dataset.

The data is organized with a top-level 'participant metadata.csv' spreadsheet containing demographic and anthropometric information. The pressure data is available in both Python-compatible (NPZ) and MATLAB-compatible (MAT) formats. For each participant, folders are organized by footwear condition (BF, ST, P1, P2) and trial type (S1, S2, S3, W1, W2, W3, W4). Each walking trial folder contains a raw trial file, two preprocessed pipeline files, and a metadata.csv file with per-footstep details.

Technical validation compared the dataset to existing public datasets. The StepUP-P150 offers the highest spatial resolution at 5 × 5 mm compared to other pressure-based gait datasets. The average GRF and COP waveforms were comparable to those from existing datasets. Spatiotemporal gait parameters such as step length, cadence, step width, and toe-out angle were extracted and compared to the CASIA-D dataset, with significant differences found in step length, cadence, and toe-out angle. The dataset also showed higher variability across the four gait parameters compared to CASIA-D due to the larger participant pool and varied conditions.

The dataset is available on the Federated Research Data Repository (FRDR) at https://doi.org/10.20383/103.01285. The Python-compatible dataset is approximately 50 GB, while the MATLAB-compatible version is approximately 118 GB. Custom scripts for processing and technical validation are provided on the companion GitHub page (https://github.com/UNB-StepUP/StepUP-P150). The dataset supports research in biometric gait recognition, biomechanics, and deep learning, and can be used for statistical models of normative walking gait, differences in pressure-based gait patterns between demographic subgroups, novel machine learning and deep learning models for gait recognition, and development and evaluation of state-of-the-art techniques for gait analysis and gait recognition.

Improvements for AI systems

Based on the paper, here are the specific improvements I can implement in AI systems, along with the resulting capabilities:

  • Improvement: Train a deep learning model on the 200,000+ footsteps with 4 sensors/cm2 resolution, using the two provided preprocessing pipelines (Pipeline 1: rotation+padding+translation+interpolation; Pipeline 2: adds resizing+amplitude normalization).

  • Capability: Recognize individuals with high accuracy from a single footstep, robust to changes in footwear (barefoot, standard shoes, personal shoes) and walking speed (preferred, slow-to-stop, slow, fast). The system can leverage the Outlier, Standing, and Incomplete metadata flags to filter low-quality samples, improving real-world reliability.

  • Improvement: Use the raw trial data (9000×720×240 tensors) to compute spatiotemporal parameters (step length, step width, cadence, toe-out angle) and time-series signals (GRF, COP) via the provided scripts.

  • Capability: Automatically quantify gait abnormalities, track recovery progress, or assess fall risk in older adults (age range 19–91). The system can compare a patient’s gait to the normative distributions from 150 healthy participants, stratified by sex, age, and body size.

  • Improvement: Train a classifier on peak pressure images (Fig. 8) to distinguish footwear types (athletic sneakers, boots, sandals, high heels, etc.) using the Shoe1Category/Shoe2Category metadata.

  • Capability: Identify the type of shoe worn from underfoot pressure data alone, enabling applications in security (e.g., detecting prohibited footwear) or retail (e.g., analyzing customer preferences). The system can also detect unusual footwear (e.g., steel-toe boots, stilettos) that produce distinct pressure signatures.

  • Improvement: Develop a model to predict demographic attributes (age, sex, BMI) from footstep pressure patterns, using the participant metadata.csv for labels.

  • Capability: Provide non-invasive health screening tools, such as estimating BMI or identifying age-related gait changes. The system can flag potential musculoskeletal issues by comparing a user’s pressure distribution to age/sex-matched norms from the dataset.

  • Improvement: Implement the SORT-based tracking algorithm described in the paper (Kalman filter + linear motion model) on raw pressure frames.

  • Capability: In real-time, detect and segment individual footsteps from a continuous pressure stream, even with high-arch shoes or shuffling (as in slow-to-stop trials). The system can output 3D bounding boxes (time, height, width) for each step, enabling live gait monitoring in smart floors or security checkpoints.

  • Improvement: Use the provided preprocessing scripts (Table 3) to create multiple dataset variants, then train a domain-adaptation model to transfer knowledge from StepUP-P150 to smaller datasets (e.g., CASIA-D, SFootBD).

  • Capability: Improve recognition accuracy on legacy or lower-resolution pressure datasets by leveraging the high-resolution, diverse StepUP-P150 data. The system can handle varying sensor densities (Fig. 12) and different walking protocols.

  • Improvement: Train a regression model using the Incomplete and Rscore metadata to estimate missing gait parameters (e.g., step length, COP trajectory) from partial footstep data.

  • Capability: Provide accurate gait analysis even when footsteps are partially off the sensor grid (e.g., in real-world deployments with smaller sensing areas). The system can flag low-confidence predictions using the Outlier field.

  • Improvement: Use the two personal shoe conditions (P1, P2) per participant to train a user-specific model that adapts to footwear changes.

  • Capability: Maintain high recognition accuracy when a user switches shoes, by learning a footwear-invariant representation. This is critical for biometric systems where users may wear different shoes daily.

  • Improvement: Provide a standardized evaluation protocol using the dataset’s metadata (e.g., Side, Orientation, Standing) to split data into train/test sets.

  • Capability: Enable fair comparison of different gait recognition algorithms, reducing reproducibility issues in the field. The system can serve as a reference for future research, as the dataset is the largest of its kind (200,000+ steps vs. 20,000 in SFootBD).

  • Improvement: Use the raw 3D tensors to generate synthetic footsteps via spatial/temporal perturbations (rotation, scaling, interpolation) as described in the normalization section.

  • Capability: Improve the performance of AI models trained on small pressure datasets by augmenting with realistic variations from StepUP-P150, reducing overfitting and improving generalization.

These improvements leverage the dataset’s unique strengths: high spatial resolution (4 sensors/cm2), large sample size (150 participants), diverse covariates (4 speeds × 4 footwear types), and rich metadata (demographics, anthropometrics, labels). The resulting AI systems can operate with higher accuracy, robustness, and real-world applicability compared to those trained on existing datasets.

Abstract

Gait refers to the patterns of limb movement generated during walking, which are unique to each individual due to both physical and behavioral traits. Walking patterns have been widely studied in biometrics, biomechanics, sports, and rehabilitation. While traditional methods rely on video and motion capture, advances in plantar pressure sensing technology now offer deeper insights into gait. However, underfoot pressures during walking remain underexplored due to the lack of large, publicly accessible datasets. To address this, we introduce the UNB StepUP-P150 dataset: a footStep database for gait analysis and recognition using Underfoot Pressure, including data from 150 individuals. This dataset comprises high-resolution plantar pressure data (4 sensors per cm-squared) collected using a 1.2m by 3.6m pressure-sensing walkway. It contains over 200,000 footsteps from participants walking with various speeds (preferred, slow-to-stop, fast, and slow) and footwear conditions (barefoot, standard shoes, and two personal shoes), supporting advancements in biometric gait recognition and presenting new research opportunities in biomechanics and deep learning. UNB StepUP-P150 establishes a new benchmark for plantar pressure-based gait analysis and recognition.

Related papers