Active headrest combined with a depth camera-based ear-positioning system

Yuteng Liu, Haowen Li, Haishan Zou, Jing Lu, Zhibin LinarXiv 2024 · DOI · arXiv

Abstract

Active Noise Control (ANC) headrests are designed to create localized "quiet zones" around a listener's ears. However, the effectiveness of these systems is highly sensitive to head movement, as the quiet zones are typically only a few centimeters in diameter. This paper proposes a robust ear-positioning (EP) system that utilizes a high-speed human pose estimation model (RTMPose) and a depth camera to track 3D ear positions in real-time. By incorporating a symmetry-based post-processing algorithm, the system maintains tracking accuracy even during significant head rotations. Experimental results demonstrate that the integrated system achieves a Mean Deviation Error (MDE) of less than 4 mm and maintains noise reduction performance of 11.7–13.6 dBA during movement, significantly outperforming static ANC systems.

Introduction

The primary challenge in active headrest design is the spatial limitation of the generated anti-noise field. For high-frequency sounds, the "quiet zone" is extremely small. When a user moves their head—whether through simple translation or rotation—their ears often exit this zone, leading to a sharp decline in noise reduction or even noise amplification. Traditional tracking methods using infrared sensors or thermal imaging often suffer from environmental sensitivity or require invasive user equipment. This research introduces a non-invasive, real-time 3D tracking solution that dynamically adapts ANC filters to the user's current position.

Methodology

The proposed system integrates an Ear-Positioning (EP) pipeline with a feedforward ANC headrest. The process is divided into three primary stages: landmark detection, 3D mapping, and post-processing.

System pipeline for 3D ear positioning.

Human Pose Estimation and 3D Mapping

The system employs the RTMPose framework, a deep learning model pre-trained on datasets such as MS COCO and MPII, to identify five key facial landmarks: the ears, eyes, and nose. These 2D coordinates are captured from a color frame and subsequently mapped to 3D space using an Intel RealSense D455 depth camera. The camera utilizes structured light to provide precise distance measurements for each detected landmark.

Symmetry-Based Post-Processing

A critical innovation of this system is its handling of occlusions. During head rotations (e.g., turning 60 degrees), one ear may become invisible to the camera. To address this, the authors developed an algorithm based on facial symmetry. By defining a perpendicular bisector plane using the coordinates of the eyes and nose, the system can calculate the "virtual" position of an occluded ear by mirroring the visible one. This ensures continuous 3D tracking across a wide range of motion.

ANC Control Logic

The ANC system operates in two phases. In the training stage, optimal control filters are pre-calculated for a grid of potential head positions. In the control stage, the EP system identifies the user's 3D coordinates at 32 FPS and signals a TI DSP controller to apply the nearest pre-stored filter, effectively "moving" the quiet zone with the listener.

Experimental Evaluation

To validate the system, the authors constructed an experimental rig featuring a dummy head mounted on a precision scanning stage. This setup allowed for the collection of ground-truth 3D coordinates to compare against the system's estimates.

Experimental setup featuring the dummy head, depth camera, and ANC components.

The system was tested against broadband noise in the 80–2000 Hz range. Performance was evaluated for both translational movements (shifting along X, Y, or Z axes) and rotational movements (turning the head up to 60°).

Results

The EP system demonstrated high precision and robust noise control:

  • Positioning Accuracy: The MDE remained below 4 mm for translations and under 1.4 cm for rotations up to 60°.
  • Noise Reduction (NR): During translational movement, the system maintained a noise reduction of 11.7–12.2 dBA. In contrast, a static system (with tracking OFF) saw performance drop to -4.8 to -7.5 dBA, indicating noise amplification.
  • Rotational Robustness: At a 60° rotation, the EP system achieved 11.4–13.6 dBA of reduction, whereas the static system's performance varied by over 23 dBA.

Conclusion

This research successfully demonstrates that combining real-time human pose estimation with depth sensing can overcome the spatial limitations of traditional active headrests. The use of facial symmetry to mitigate occlusion is a key advancement for practical, real-world applications. Future work will focus on scaling the system for multiple users and reducing hardware dependency by exploring monocular depth estimation techniques.

Summary updated

BibTeX
@misc{liu2024active,
  author = {Yuteng Liu and Haowen Li and Haishan Zou and Jing Lu and Zhibin Lin},
  title = {{Active headrest combined with a depth camera-based ear-positioning system}},
  year = {2024},
  eprint = {2401.10256},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2401.10256},
}
Figure