Yuyan Wu.
← All research

03 Interpret existing traces

Emotion recognition

We recognize human emotions by estimating valence and arousal from footstep-induced floor vibrations. The main intuition is that emotions affect gait patterns, which in turn shape the footstep-induced floor vibrations captured by our vibration sensors.

A key challenge is the large between-person variation in this emotion–gait–vibration relationship. Our main contribution is personalized learning: we compare each training person’s gait with the target person’s gait and assign higher weights to samples from people with higher gait similarity when fine-tuning the target person’s model.

See how personalization works EmotionVibe · IEEE TAFFC 2026 ↗

How footsteps appear in floor-vibration recordings

MEASURED DATA

Play the walk to follow the vibration traces, then choose a sensor to inspect its recorded signal.

Step 1 of 11
Four sensors occupy the positions shown in the supplied experiment diagram. Choose a walk and channel to explore a real twelve-second excerpt. The four overview traces share a normalized amplitude scale. The selected sensor’s signal is enlarged below, with labeled amplitudes.

Loading recorded footsteps…

Personalized learning from gait similarity

Scroll to explore the figureView full figure
The gait of each training person, A, B and C, is compared with the blue target person’s gait. Higher gait similarity means more weight for every sample from that training person when fine-tuning the target person’s emotion model.

Higher gait similarity to the target person means more weight for that training person’s samples. Every sample from the same person receives the same weight when fine-tuning the target person’s emotion model. Paper ↗

Method & reference

Similarity is computed in the general model’s feature space. Distances between target and training footsteps are averaged for each training person, then inverted and normalized into a Gait Similarity Index. A higher index means greater gait similarity and a larger sample weight. Every sample from that person inherits this weight during fine-tuning.

PAPER

  1. [1]

    Y. Wu, Y. Dong, S. Vaid, G. M. Harari & H. Y. Noh (2026). EmotionVibe: Human Emotion Recognition Through Footstep-Induced Floor Vibrations. IEEE TAFFC, 17(2), 1788–1805.Author preprint · Personalized learning: §IV-D.

← Explore the other projects