Publications |

Real-Time Foot Height Estimation and Activity Classification Using a Foot-Mounted IMU Implemented on a Smartphone

Sat, 16 May 2026 00:00:00 +0000

Why it matters

Tripping is a leading cause of falls, and it is closely tied to how high we lift our feet while walking. Measuring foot clearance continuously, out in daily life, could flag fall risk early — but doing it well with a wearable sensor is hard:

A single inertial measurement unit (IMU) drifts, so integrated foot height accumulates error over time.
The usual fix (zero-velocity updates) does not fully remove vertical drift.
More accurate methods work only offline and do not tell you what activity the person is doing.

What I built

A real-time, single-IMU system that runs entirely on a smartphone — no cloud, no multi-sensor setup. It reconstructs the foot’s height trajectory stride by stride and classifies five locomotion activities (level walking, ramp ascent/descent, stair ascent/descent) at the same time.

Key contributions:

On-device Android app for stride-by-stride foot-clearance measurement and live activity classification.
Zero Height Change (ZHC) constraint — a biomechanically grounded correction that cancels cumulative vertical drift at each stride boundary.
Heel-strike velocity error used as a compact, physically meaningful feature to adaptively pick the right reconstruction model per activity.
Toe-height estimation from the IMU pose through a rigid-body transformation, with accuracy comparable to the foot itself.

Validating the height measurement

The reconstructed foot trajectories separate cleanly by activity, and the continuous height signal tracks real-world ramps and stairs across a full walking circuit.

To confirm the system measures real clearance, participants stepped over boxes of known heights and the estimated peak stride height was checked against each obstacle.

Results

96.08% overall activity-classification accuracy, with less than one gait cycle of latency.
Level walking stayed at the ground reference (0.0 cm, 95% CI −1.8 to 1.8 cm).
Cumulative height error stayed below 1.1 cm across ramp and stair negotiation (mean absolute error 0.42%).
Toe height was recovered with accuracy comparable to foot height.

Together, this gives a practical foundation for real-time gait feedback and fall-prevention tools that could one day run on the phone already in someone’s pocket.

Real-Time Activity Recognition Using Minimal Biomechanical Features: A Lightweight IMU-Based Classifier for Older Adults

Thu, 04 Sep 2025 00:00:00 +0000

Why it matters

Knowing what activity a person is doing, and how well they are doing it, is the foundation of any real-world health-monitoring tool. But most activity-recognition systems fall short where it matters most:

They lean on deep learning and large datasets, so they are too heavy to run in real time on a phone.
They are trained almost entirely on young, healthy adults, and quietly assume everyone moves the same way.
Older adults with mobility limits often negotiate stairs with a non-alternating leg (NAL) pattern, leading with the same leg every step, and generic models simply misread this as normal.

What I built

A real-time, smartphone-based classifier that uses only two interpretable inputs, the thigh angle and the foot angle from two small IMUs, to recognize five everyday activities: walking, stair ascent, stair descent, standing, and sitting.

Key contributions:

Minimal biomechanical features. Rather than a deep network, a simple and interpretable physics based KNN model classifies activity from just the thigh and foot angles, light enough to run live on a phone.
Adaptive toe-off segmentation. A segmentation step detects toe-off across different activities to cut each stride cleanly and reduce computation.
On-device Android app that collects labeled data and reports activity in real time.
Built for older adults. The model is explicitly trained to recognize the non-alternating leg stair pattern that generic classifiers miss.

See it in action

A short demo of the system running live: the phone screen shows the app classifying activities in real time while a participant walks and negotiates stairs.

Recognizing how older adults move

The hard case is stairs. When someone leads with the same leg on every step, the leading and trailing legs trace very different angle patterns, so the system has to read each leg differently.

Separating the leading leg from the trailing leg in the foot and thigh angle traces is what lets the model distinguish this mobility-limited pattern from normal stair negotiation.

Results

99% activity-classification accuracy for young and older adults with normal movement patterns.
93% accuracy in distinguishing non-alternating leg stair negotiation from normal stair negotiation in older adults.
Runs in real time on a smartphone, with no deep-learning model and no cloud.
Validated on cohorts the model never trained on, including ten older adults.

Together this shows that a small set of well-chosen biomechanical features can power an accessible, real-time health-monitoring system that respects how aging populations actually move.

The Effects of Real-Time Haptic Feedback on Gait and Cognitive Load in Older Adults

Wed, 11 Jun 2025 00:00:00 +0000

Why it matters

Walking speed is one of the strongest predictors of health and independence in later life, and when it drops the cause is often neuromotor, the timing and coordination of movement, rather than weakness or poor fitness. A key culprit is reduced peak thigh extension (PTE), how far the thigh swings back at the end of each step. If we could cue people to extend the thigh a little more, in real time and out in the world, we could lengthen their stride and speed them up, without a treadmill, a therapist, or a lab.

What I built

A wearable, smartphone-based haptic feedback system. Two IMUs track the thigh and foot angles, the phone computes PTE stride by stride, and small vibrotactile cells on the thighs deliver a gentle cue when a step needs adjusting.

How the feedback works:

Each stride’s PTE is compared to a personalized target.
Two feedback strategies were tested: a cue when the step was incorrect (PTE below target, F_IR) and a cue when the step was correct (target met, F_CR).
A third group received only verbal instructions (I_V), as a benchmark for the haptic system.

Measuring the mental cost

Any gait cue is only useful if people can follow it without concentrating so hard that they stop paying attention to their surroundings. To measure that cognitive load, participants did a probe reaction-time task while walking, responding to audio beeps, with reaction time measured precisely from the sent and received audio.

Results

+14% stride length and +18% gait speed from a single session with the haptic system.
Gains were comparable to verbal coaching, but reached through a different mechanism (direct thigh feedback rather than instruction).
Added cognitive load was small: reaction time rose only 27 ms (F_IR) to 74 ms (F_CR).
Thirty community-dwelling older adults, averaging about 80 years old.

Together this shows that real-time thigh haptics can meaningfully improve older-adult gait in a single session while keeping attentional demand low, a promising, low-cost path toward wearable gait training and fall prevention.

Increasing Thigh Extension with Haptic Feedback Affects Leg Coordination in Young and Older Adult Walkers

Sat, 01 Mar 2025 00:00:00 +0000

Trajectory Optimization of a Spot-Welding Robot in the Joint and Cartesian Spaces

Sun, 01 Jan 2023 00:00:00 +0000

Revisiting Screw Theory-Based Approaches in the Constraint Wrench Analysis of Robotic Systems

Fri, 01 Jan 2021 00:00:00 +0000