RHM Dataset

A comprehensive multiview visual dataset for human activity recognition designed for assistive robots and HRI scenarios

Complete Dataset Overview: All Activities & Views

26,804 videos across 14 activities4 synchronized viewpoints • RGB + Skeleton data

Front
Back
Robot
Omni
Back to Projects

Introduction

The Robot House Multi-View (RHM) Dataset represents a significant advancement in human activity recognition research, specifically designed for assistive robotics and ambient intelligence applications. Created in a typical British home environment, this comprehensive dataset addresses the critical need for understanding human activities in domestic settings where companion robots and smart home systems must operate effectively.

What sets RHM apart is its focus on essential daily living activities that are crucial for independent living and home care scenarios. The dataset features synchronized multi-view recordings from four strategically positioned cameras, including a unique mobile robot perspective that provides dynamic viewpoints not available in traditional fixed-camera setups.

26,804
Total Video Clips
4
Synchronized Views
14
Activity Classes
6,701
Videos per View

Multi-Camera Setup

The RHM dataset employs a sophisticated four-camera configuration designed to capture comprehensive activity information from complementary perspectives. This multi-view approach enables robust activity recognition that can handle occlusions, varying lighting conditions, and different spatial relationships.

FrontView Camera

  • Type: Static wall-mounted
  • Resolution: 640 × 480 @ 30 FPS
  • Advantage: Elevated position captures table surfaces, stairs, and full room area
  • Performance: Most reliable for pose extraction

BackView Camera

  • Type: Static wall-mounted
  • Resolution: 640 × 480 @ 30 FPS
  • Position: Opposite perspective to FrontView
  • Coverage: Complementary angles for activity verification

RobotView Camera

  • Platform: Fetch Mobile Robot
  • Resolution: 640 × 480 @ 30 FPS
  • Unique Feature: Dynamic following and head tracking
  • Strength: Excels in vertical movement activities

OmniView Camera

  • Type: Ceiling-mounted fish-eye
  • Resolution: 512 × 486 @ 30 FPS
  • Coverage: Complete room omnidirectional view
  • Note: Excluded from analysis due to pose extraction challenges

Activity Classes and Distribution

The 14 activity classes in RHM were carefully selected based on their importance for independent living and their relevance to assistive robotics applications. Each activity represents a fundamental daily task that companion robots and ambient systems must recognize to provide meaningful assistance.

Lifting Objects 700
Reaching 696
Putting Objects Down 649
Cleaning 448
Sitting Down 437
Stretching 433
Stairs Climbing Up 430
Walking 426
Opening Can 421
Closing Can 421
Stairs Climbing Down 413
Carrying Objects 412
Drinking 408
Standing Up 407

Activity Selection Rationale

These activities span fundamental categories essential for home care: mobility (walking, stairs), object manipulation (lifting, carrying), self-care (drinking, stretching), and household tasks (cleaning, reaching). This comprehensive coverage ensures the dataset addresses real-world scenarios where assistive robots must demonstrate understanding and appropriate response capabilities.

Dataset Organization and Splits

The RHM dataset follows rigorous machine learning best practices with standardized train-validation-test splits that ensure reliable evaluation and comparison across different research approaches.

65%
Training Set
4,278 videos per view
17,112 total videos
15%
Validation Set
1,076 videos per view
4,304 total videos
20%
Test Set
1,347 videos per view
5,388 total videos

Technical Specifications

Clip Duration: Variable 1-5 seconds
Synchronization: All 4 camera views perfectly aligned
Environment: Typical British home living room
Subjects: Single person per video clip

RHM Skeleton: Skeleton Extension

Building upon the RGB foundation, the RHM Skeleton dataset provides skeleton-based pose data extracted using state-of-the-art human pose estimation. This multi-modal extension enables researchers to explore both appearance-based and pose-based approaches to activity recognition, offering complementary perspectives for robust system development.

Pose Extraction Pipeline

  • Model: HRNet (COCO + MPII trained)
  • Output: 17 body keypoints per frame
  • Format: X, Y coordinates + confidence scores
  • Processing: JSON → 5D Tensor conversion

Dataset Structure

  • Videos: 6,700 synchronized across views
  • Tensor Shape: [Sample, View, Frame, Pose, Coord]
  • Frame Sampling: Fixed lengths (34, 64, 128)
  • Quality Control: Confidence-based filtering

Comprehensive Quality Analysis

Extensive quality assessment reveals critical insights about pose extraction reliability across different camera views and activity types, providing valuable guidance for multi-view system design.

Camera Performance Analysis

  • Front View: Most reliable overall performance
  • Robot View: Excels in vertical movements (stairs)
  • Back View: Challenges with most pose types
  • Omni View: High failure rate (excluded)

Joint Reliability Ranking

  • Most Reliable: Left/Right Shoulders
  • Moderate: Head region (nose, eyes, ears)
  • Variable: Arms and hip joints
  • Challenging: Ankle joints (especially robot view)

Key Research Insights

  • Activity-View Correlation: Vertical activities (stairs, sitting) show different reliability patterns across camera positions
  • Distance Effect: Mobile robot's proximity creates both advantages (detailed tracking) and challenges (occlusion)
  • Elevation Advantage: Higher-positioned cameras provide more consistent pose extraction
  • Joint Hierarchy: Core body joints (shoulders, hips) significantly more reliable than extremities

Research Impact and Applications

The RHM dataset addresses critical gaps in human activity recognition research, particularly in domestic environments where assistive technologies must operate reliably and safely.

Assistive Robotics

Enable home care robots to understand daily activities, predict user needs, and provide appropriate assistance while maintaining safety and privacy.

Ambient Intelligence

Develop smart home systems that adapt to user behavior patterns, optimize energy usage, and provide proactive support for independent living.

Healthcare Monitoring

Create non-intrusive monitoring systems for elderly care, rehabilitation tracking, and early detection of health changes through activity pattern analysis.

Multi-View Learning

Advance computer vision research in multi-perspective analysis, view fusion techniques, and robust activity recognition under varying conditions.

Related Publications

The RHM dataset evolution represents a comprehensive research trajectory spanning foundational work to advanced multi-modal analysis:

Citation Required

Please cite the relevant papers below if you are using the datasets in your research.

RHM: Robot House Multi-view Human Activity Recognition Dataset

Comprehensive documentation of the multi-view RGB dataset, detailing collection methodology, synchronization techniques, and baseline performance evaluations.

RH-HAR-SK: A Multi-view Dataset with Skeleton Data for Ambient Assisted Living Research

Advanced extension incorporating skeleton pose data, quality analysis across multiple views, and applications to ambient assisted living scenarios.

Robot house human activity recognition dataset

Original dataset introduction presenting the Robot House platform and initial activity recognition capabilities, laying groundwork for multi-view expansion.

Human activity recognition in robocup@home: inspiration from online benchmarks

Foundational exploration of HAR challenges in domestic environments, establishing the conceptual framework that inspired the RHM dataset development.

Access the Datasets

Both the RHM RGB dataset and RHM Skeleton dataset are available for research purposes. Please cite the relevant publications when using these datasets in your research.

RHM RGB Dataset

Multi-view RGB video dataset with 26,804 videos across 4 synchronized camera views covering 14 daily activities.

RGB Dataset

RHM Skeleton Dataset

Skeleton pose dataset with 17 keypoints extracted from 6,700 synchronized videos using HRNet, stored in 5D tensor format.

Skeleton Dataset

GitHub Repositories & Source Code

Access the complete implementation code, processing tools, and analysis frameworks developed for the RHM dataset. Each repository contains specific components of the dataset creation and analysis pipeline.

RHM OneView

Single-view activity recognition implementation and baseline experiments.

Feature Extractor

Tools for extracting features from video frames for activity analysis.

DualStream C3D

Dual-stream C3D implementation for multiview activity recognition.

Video Editor

OpenCV-based tools for multiple video editing and preprocessing.

Repository Collection: Complete toolkit for dataset processing, feature extraction, model training, and evaluation. Includes baseline experiments, preprocessing scripts, and analysis tools for multiview human activity recognition research.

Contact for Details
Back to Projects