Frequency Dithered Human Pose Analytics for Privacy in Fitness Cameras

Abstract

The proliferation of smart fitness systems equipped with high-resolution cameras has introduced significant privacy vulnerabilities within residential environments. While continuous visual data capture is essential for accurate human pose estimation and biomechanical feedback, it simultaneously exposes users to severe risks of unauthorized surveillance and sensitive data leakage. Existing privacy-preserving methods, such as software-based blurring, federated learning, and hardware-level optical defocusing, often impose substantial computational overhead or unacceptably degrade the accuracy of pose analytics. To address these limitations, this paper proposes a novel framework utilizing frequency-dithered human pose analytics specifically designed for privacy-sensitive fitness camera systems. By introducing controlled, high-frequency spatial and temporal dithering at the sensor level, the proposed system obfuscates identity-revealing biometric features, such as facial details and environmental context, while preserving the low-frequency structural kinematics required for precise skeletal mapping. Comprehensive empirical evaluations demonstrate that the frequency-dithering mechanism reduces facial recognition confidence to near-random probability without significantly compromising the mean per joint position error in pose estimation tasks. Furthermore, the localized nature of the dithering algorithm ensures minimal latency, making it highly suitable for real-time fitness applications on resource-constrained edge devices. This research bridges the critical gap between biomechanical utility and data privacy, offering a scalable solution for next-generation smart home fitness ecosystems.