StickFigure: Enabling Pose-Only IoT Cameras for Privacy-Preserving Edge Vision

Aveline Xiaoyuan Han; Chenshu Wu

cs.CVOct 25, 2026

StickFigure: Enabling Pose-Only IoT Cameras for Privacy-Preserving Edge Vision

Authors: Aveline Xiaoyuan Han, Chenshu Wu

Organizations: The University of Hong Kong

Abstract

Indoor IoT cameras can support assisted living and activity-aware spaces, but exporting raw RGB frames raises privacy concerns by exposing facial, clothing, and household details that body-motion tasks often do not require. We present StickFigure, a pose-only IoT camera system that, during real-time monitoring, keeps raw frames local and ephemeral while exporting only compact skeleton streams sufficient for downstream human activity recognition (HAR). While pose estimation has matured on GPUs, extracting accurate pose streams on resource-constrained IoT cameras remains difficult. To overcome these challenges and enable on-device pose streams on (almost) any IoT camera, we present two core designs in StickFigure: an activation-aware pose backbone that allocates channel capacity after spatial downsampling and a lightweight hierarchical MLP refiner that selectively corrects elbow and wrist coordinates using intermediate visual features and kinematic context. We fully implement StickFigure and evaluate it on public benchmarks and real-world deployment. Compared with state-of-the-art MCU pose models, StickFigure achieves the highest pose accuracy and downstream HAR utility on both NTU RGB+D and Penn Action, while using 96.5% fewer parameters and 99.3% fewer MACs per frame than RTMPose-S-based baselines, achieving 0.107 s p95 latency on a representative STM32H753 MCU. We believe StickFigure establishes a new operating point for privacy-aware edge vision on massive IoT cameras.

Figures & tables

Figure 1. Representation trade-offs for body-centric sensing. Qualitative comparison of visual exposure (lower toward the right) and downstream flexibility (higher toward the top). Pose reduces appearance-rich exposure while retaining a flexible body-motion representation.

Figure 2. StickFigure system overview. During real-time monitoring, RGB remains within the trusted camera boundary. An activation-aware pose backbone and a hierarchical MLP refiner produce a 12-joint pose, which is exported as a 76 B fixed-format skeleton record (SFSK) for downstream HAR. The dashed path shows one-shot site calibration.

Figure 3. Activation-aware pose backbone and hierarchical joint refinement. The backbone delays larger channel expansion until after spatial downsampling, while the joint refiner uses intermediate features and kinematic-parent context to correct elbows and wrists.

Figure 4. Hardware prototype and deployment environments. Left: the NUCLEO-H753ZI prototype with an STM32H753ZIT6 MCU, an OV2640 camera module, and the board-level power-measurement point. Right: views of deployment Envs. A, D, and E used for live evaluation in § 4.4 .

MCU-scale methods
Method	Model resources				NTU RGB+D				Penn Action
	Params	Flash	MACs/ frame	Peak activation RAM	[email protected]	p95 Latency	HAR Top-1	Fine-arm Recall	[email protected]	p95 Latency	HAR Top-1	Fine-arm Recall
Tiny MobileNet Pose	307.4K	313.88 KiB	13.39M	197.05 KiB	66.40%	0.061 s	$54.70{\pm}0.85$ %	38.00%	62.22%	0.067 s	$81.37{\pm}1.19$ %	83.56%
Uniform-Expansion Backbone	435.6K	439.96 KiB	41.27M	262.75 KiB	72.90%	0.121 s	$68.55{\pm}0.30$ %	51.80%	68.81%	0.127 s	$85.80{\pm}0.59$ %	85.28%
Progressive-Expansion Backbone	284.9K	290.94 KiB	25.14M	262.75 KiB	73.70%	0.076 s	$69.43{\pm}0.34$ %	52.70%	64.40%	0.082 s	$84.05{\pm}0.27$ %	85.24%
XiNet Pose	607.5K	603.46 KiB	75.09M	283.63 KiB	72.08%	0.190 s	$60.45{\pm}0.80$ %	41.79%	69.01%	0.209 s	$86.70{\pm}1.06$ %	87.40%

Table 1. End-to-end FP32 pose, p95 pose latency, HAR, and resource comparison on NTU RGB+D and Penn Action. Flash reports compiled INT8 weights and associated constants, while peak activation RAM reports the maximum simultaneously live model activation memory in the profiled NTU compiled graph. Non-MCU models are accuracy references; dashes indicate unavailable exact resource or target-latency measurements.

Method	NTU RGB+D			Penn Action
	FP32	INT8	$\Delta$	FP32	INT8	$\Delta$
Tiny MobileNet Pose	66.40%	65.87%	$-0.53$	62.22%	62.01%	$-0.21$
Uniform-Expansion Backbone	72.90%	71.84%	$-1.06$	68.81%	68.65%	$-0.16$
Progressive-Expansion Backbone	73.70%	64.20%	$-9.50$	64.40%	64.19%	$-0.21$
XiNet Pose	72.08%	70.45%	$-1.63$	69.01%	68.85%	$-0.16$
Activation-Aware Backbone	78.00%	77.45%	$-0.55$	75.49%	75.33%	$-0.16$

Table 2. FP32-to-INT8 [email protected] retention.

Figure 5. Design analysis of the activation-aware capacity allocation and hierarchical distal-joint refinement. (a) Stage-wise capacity allocation under a fixed largest individual activation tensor. (b) Accuracy–computation trade-off across the three controlled expansion schedules. (c) Joint-wise gains from the refiner, showing improvements concentrated on the elbow and wrist.

Figure 6. Live deployment results. (a) [email protected] after one-shot site calibration. Orange and blue markers denote target-supervised references and one-shot calibration results. The dashed line denotes the dataset-domain reference. (b) Three-seed mean HAR Top-1 for the dataset reference and two independently collected Env A sessions; Env A* changes appearance and lighting.

Figure 7. Motion-triggered adaptive execution. A 100-s power trace with two human-entry periods. MOTION_CHECK performs one motion check per second and switches to full-rate POSE_INFERENCE upon detecting motion, returning to MOTION_CHECK after approximately 5 s of inactivity.

Figure 8. Long-horizon monitoring case study. Left: pose, HAR, and temporal metrics from a separate approximately 40-minute annotated session. Center: three daytime monitoring timelines. Right: day-specific prediction timelines, inter-method disagreement (Dis.), and activity-state switches per hour for the two estimators.

Figure 9. Cross-platform INT8 pose and downstream utility. (a) INT8 [email protected] and (b) downstream HAR Top-1 for the best feasible operating point of each architecture family on ESP32-S3, STM32H753, Raspberry Pi 4, and Jetson Orin Nano. Each point is the highest-performing configuration satisfying the cross-platform real-time gate of p95 inference latency

\leq 125

ms and sustained throughput

\geq 8

FPS. StickFigure achieves the highest pose and HAR utility across all four platforms.

Human Pose Estimation

Human pose estimation recovers people’s body keypoints, 3D skeletons, or meshes from images, video, and other sensor data.

See all 79 papers

Sep 13, 2026

EdgeHAR: An Edge-Native Compact Sensor Foundation Model for Human Activity Recognition

Sensor-based human activity recognition (HAR) is fundamental to ubiquitous and wearable computing, yet existing foundation models are largely designed for cloud-scale deployment and struggle with real-world sensing shifts, including unseen users, devices, sampling rates, and sensor placements. We present EdgeHAR, an edge-native compact sensor foundation model designed for wearable intelligence. Unlike conventional models that entangle activity knowledge with acquisition variations, EdgeHAR learns transferable representations by factorizing sensor signals into three latent codes: an (i)Activity-Semantic Code capturing reusable activity knowledge, a (ii)Motion-Dynamics Code modeling temporal patterns, and an (iii)Acquisition-Context Code representing sensor-specific variations. This disentangled design enables efficient adaptation to new users, devices, placements, and activity classes with limited target-domain data. By incorporating lightweight adaptation modules, EdgeHAR achieves foundation-model-level transferability while satisfying edge constraints in computation, memory, latency, and privacy. Experiments across heterogeneous HAR datasets demonstrate that EdgeHAR maintains competitive recognition performance under distribution shifts with substantially reduced deployment cost. EdgeHAR establishes a practical paradigm for compact, edge-first foundation models for ubiquitous sensing systems.

He Zhang, Siyu Yuan, Siyu Liu +2

Northwestern Polytechnical University Xi’an, Shaanxi, China · RPTU University Kaiserslautern-Landau Kaiserslautern, Rheinland Pfalz, Germany · Xi’an University of Technology Xi’an, Shaanxi, China

Human Activity Recognition Wearable Sensing Edge ML Domain Generalization Disentangled Representation Learning

Ask PDF

Sep 8, 2026

ArmPoser: Real-Time, Calibration-Free Arm Pose Estimation from Smartwatch IMU

Arm pose estimation enables applications in fitness, extended reality input, rehabilitation, and life logging. Prior smartwatch-based approaches rely on calibration poses and preprocessing pipelines that transform raw IMU measurements into standardized training formats. These steps hinder deployment in everyday settings and introduce errors due to imperfect calibration and sensor drift. We present ArmPoser, a calibration-free arm pose estimation system using a single smartwatch IMU. Our central contribution is training models directly in the reference frame native to consumer smartwatches, aligning learning with how IMU data is produced by deployed devices. By operating on device-native axes, ArmPoser removes the need for coordinate transformations, explicit alignment, and bone-offset calibration used in prior work. We further augment training with physically grounded variations in watch placement and arm morphology to account for user-specific variability. ArmPoser also includes a wear-configuration module that infers anterior or posterior forearm placement and crown orientation. We evaluate pose estimation on public benchmarks and on a 10-participant, 30-activity study using watchOS and Android smartwatches, where ArmPoser matches or exceeds calibrated baselines without any user calibration.

Bishnu Dev, Vasco Xu, Xi-Aan Loh +3

Mohamed bin Zayed University of Artificial Intelligence · University of Chicago · Northwestern University

Wearable Sensing Human Pose Estimation

Ask PDF

Mar 29, 2026

MoViD: View-Invariant 3D Human Pose Estimation via Motion-View Disentanglement

3D human pose estimation is a key enabling technology for applications such as healthcare monitoring, human-robot collaboration, and immersive gaming, but real-world deployment remains challenged by viewpoint variations. Existing methods struggle to generalize to unseen camera viewpoints, require large amounts of training data, and suffer from high inference latency. We propose MoViD, a viewpoint-invariant 3D human pose estimation framework that disentangles viewpoint information from motion features. The key idea is to extract viewpoint information from intermediate pose features and leverage it to enhance both the robustness and efficiency of pose estimation. MoViD introduces a view estimator that models key joint relationships to predict viewpoint information, and an orthogonal projection module to disentangle motion and view features, further enhanced through physics-grounded contrastive alignment across views. For real-time edge deployment, MoViD employs a frame-by-frame inference pipeline with a view-aware strategy that adaptively activates flip refinement based on the estimated viewpoint. Evaluations on nine public datasets and newly collected multiview UAV and gait analysis datasets show that MoViD reduces pose estimation error by over 24.2% compared to state-of-the-art methods, maintains robust performance under severe occlusions with 60% less training data, and achieves real-time inference at 15 FPS on NVIDIA edge devices.

Yejia Liu, Hengle Jiang, Haoxian Liu +2

Department of Computer Science and Engineering, Hong Kong University of Science and Technology

3D Human Pose Estimation Disentangled Representation Learning

Ask PDF

Abstract

Human Pose Estimation

Explore similar work

EdgeHAR: An Edge-Native Compact Sensor Foundation Model for Human Activity Recognition

ArmPoser: Real-Time, Calibration-Free Arm Pose Estimation from Smartwatch IMU

MoViD: View-Invariant 3D Human Pose Estimation via Motion-View Disentanglement