Fast Accelerometer-Based Motion Recognition with a Dual Buffer Framework

Hubert P. H. Shum, Taku Komura, Shu TakagiInternational Journal of Virtual Reality 2011 · DOI

Abstract

The low-cost gyro-accelerometer based controllers have opened up a potential of using 3D computer games for serious applications such as sports and vocational training. Previous games apply simple template matching for motion recognition, which suffers from poor accuracy and time lag. In order to cope with these problems, we propose a novel dual buffer approach that dramatically increases the recognition rate and shortens the time lag for motion recognition. The system first recognizes the user control signal with a small buffer to minimize the time lag. When more signals arrive from the sensor, an elaborate recognition is performed, and the previously recognized action is switched if necessary. Using boxing as an example, we show that we can control a virtual character to perform 13 different actions using a buffer size of one-tenth of a second. Since our system is computationally inexpensive, it can be used in game consoles. As it is accurate and responsive, it can also be applied for serious sport training.

Introduction

In the world of motion-controlled gaming—think Wii Sports or VR boxing—there is a persistent "tug-of-war" between speed and accuracy. If a system waits for a user to finish a punch to be certain of the action, the game feels laggy. Conversely, if it guesses too early, the character might perform the wrong action, such as throwing a jab when the user intended a hook. Traditional motion recognition relies on template matching, comparing controller data to a library of pre-recorded motions. However, the inherent time lag of large data windows often makes fast-paced sports games unplayable.

System Architecture

The proposed system bridges the gap between raw sensor input and complex character animation through a multi-stage pipeline. It utilizes an interaction graph to manage character behaviors and an action-level motion graph derived from professional motion capture data.

Overview of the motion recognition pipeline, from accelerometer signals to interaction graphs.

The researchers recorded seven minutes of shadow boxing from a professional boxer to create an "Action Level Motion Graph" segmented into 148 semantic actions. For control, 13 specific actions (e.g., straight punches, uppercuts, parries) were selected to form a control signal database.

The Dual Buffer Framework

The core innovation of this work is the Dual Buffer Framework, which operates across two concurrent threads: a Rendering Thread and an Action Recognition Thread.

The dual-threaded architecture showing the interaction between action recognition and rendering via the dual buffer system.

The framework functions in three distinct stages:

  1. The Main Buffer (Fast Guess): A small buffer (6 frames or 0.1 seconds) captures the start of a movement. The system performs a "rough" match using Dynamic Time Warping (DTW) and immediately triggers the corresponding animation.
  2. The Supplementary Buffer (Refinement): As the movement continues, older signals shift into a larger supplementary buffer. The system continuously re-evaluates the entire sequence (Main + Supplementary).
  3. Action Switching: If the refined match suggests the initial guess was incorrect, the system switches the animation mid-stream. To ensure visual smoothness, "temporal skipping" is used to start the new animation from the current frame count, combined with physics-based PD (Proportional-Derivative) control.

Key Results and Insights

The system achieves a response latency of just 0.1 seconds, significantly faster than many commercial alternatives of its era. For experienced players, accuracy reached approximately 95%. Notably, the dual-buffer approach provided a 10-20% accuracy boost for novice players whose movements were often ambiguous.

A critical component for maintaining synchronization is the "null signal" detector. By identifying stabilizing periods between actions, the system prevents the end of one motion from being misidentified as the start of the next. Furthermore, the system is computationally efficient, requiring less than 5ms for recognition on standard hardware, making it highly suitable for home consoles.

Conclusion

By prioritizing immediate responsiveness and allowing for mid-action correction, the Dual Buffer Framework effectively solves the latency-accuracy trade-off in motion recognition. While future work aims to incorporate more nuance into the animations based on specific acceleration profiles, the current system provides a robust foundation for both entertainment and serious sports training applications.

Summary updated

Abstract

The low-cost gyro-accelerometer based controllers have opened up a potential of using 3D computer games for serous applications such as sports and vocational training. Previous games apply simple template matching for motion recognition, which suffers from poor accuracy and time lag. In order to cope with these problems, we propose a novel dual buffer approach that dramatically increases the recognition rate and shortens the time lag for motion recognition. The system first recognizes the user control signal with a small buffer to minimize the time lag. When more signals arrive from the sensor, an elaborate recognition is performed, and the previously recognized action is switched if necessary. Using boxing as an example, we show that we can control a virtual character to perform 13 different actions using a buffer size of one-tenth of a second. Since our system is computationally inexpensive, it can be used in game consoles. As it is accurate and responsive, it can also be applied for serious sport training
BibTeX
@article{shum2011fast,
  author = {Hubert P. H. Shum and Taku Komura and Shu Takagi},
  title = {{Fast Accelerometer-Based Motion Recognition with a Dual Buffer Framework}},
  journal = {International Journal of Virtual Reality},
  year = {2011},
  doi = {10.20870/ijvr.2011.10.3.2831},
}
Figure