Wavelet Statistics for Human Motion Classification

Kevin Quennesson, Elias Ioup, Charles L. Isbell Jr.AAAI 2006

Abstract

Traditional human motion classification often relies on spatial pose analysis, which can fail to distinguish between motions that are visually similar but dynamically distinct. This paper proposes a method that captures the "rhythm" of movement using wavelet transforms. By converting motion signals into histograms of frequency-scaled wavelet coefficients, the researchers create a length-invariant representation that effectively characterizes both the shape and sharpness of motion. Experimental results demonstrate high classification accuracy across motion capture and wearable sensor datasets.

Introduction

Classifying human motion is a fundamental challenge in computer vision and robotics. Most existing approaches focus on the spatial configuration of a person—analyzing the pose in individual frames. However, many motions look nearly identical in a single frame but differ fundamentally in their dynamics. For instance, a fast walk and a slow run may share similar poses, but their temporal frequencies are distinct.

Previous attempts to address temporal dynamics, such as Dynamic Time Warping (DTW) or Hidden Markov Models (HMMs), often require complex signal alignment or suffer from errors during state transitions. This research seeks a robust alternative that captures both the "shape" (low-frequency trends) and the "sharpness" (high-frequency discontinuities, such as foot strikes) of motion simultaneously.

Methodology

The proposed classification pipeline consists of three primary stages: wavelet transformation, frequency scaling/histogramming, and SVM classification.

Wavelet Transform

Unlike the windowed Fourier transform, which uses a fixed analysis window, the wavelet transform employs "Heisenberg boxes" that vary in size. This multi-resolution approach allows the system to analyze high-frequency variations over short time periods and low-frequency shapes over longer durations. This is particularly suited for human motion, where sudden "shocks" like a heel strike provide critical classification cues.

Frequency Scaling and Histogramming

To achieve alignment invariance, the authors convert the time-series wavelet coefficients into a statistical histogram. Because high-frequency coefficients naturally have smaller magnitudes, they are scaled by a constant ($C=4$) as frequency decreases. This ensures that all frequency bands contribute meaningfully to the final feature vector. This histogram-based approach makes the method "length-agnostic," allowing for the direct comparison of motion clips of varying durations.

Visualization of motion signal decomposition and the resulting statistical histogram.

Classification

The histograms from different motion dimensions (e.g., joint angles or accelerometer axes) are concatenated into a single feature vector. This vector is then classified using a Support Vector Machine (SVM). Optimal performance was achieved using biorthogonal wavelets with 6 vanishing points and a histogram bin size of 0.15.

Key Results

The method was evaluated on Motion Capture (MoCap) data from CMU and real-world sensor data (accelerometer and gyroscope) covering various exercises.

Classification accuracy across MoCap and sensor datasets.

As shown in the results table, the approach achieved between 97.3% and 100% accuracy on MoCap data, successfully distinguishing between similar rhythms like "quick walks" and "slow walks." On noisier wearable sensor data, the method maintained a high accuracy of 96.3%, suggesting significant robustness for real-world applications.

Discussion and Limitations

The use of wavelet statistics provides a powerful "fingerprint" for human motion. For example, symmetric motions like marching produce histograms centered around zero, reflecting the underlying signal symmetry.

However, several areas for future work remain. The scaling constant $C=4$ was determined empirically and may need to be adaptive for different sensor types. Additionally, while the method excels at rhythmic motions like walking or exercising, its performance on non-rhythmic or stochastic behaviors requires further investigation. Finally, future research should address the computational latency of building high-dimensional histograms for real-time applications.

Summary updated

BibTeX
@inproceedings{quennesson2006wavelet,
  author = {Kevin Quennesson and Elias Ioup and Charles L. Isbell Jr.},
  title = {{Wavelet Statistics for Human Motion Classification}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure