TLDR: ProGait is a novel video dataset designed to improve AI models for analyzing the gait of transfemoral prosthesis users. It comprises 412 video clips from four above-knee amputees testing various prosthetic legs, featuring detailed annotations for video object segmentation, 2D human pose estimation, and gait analysis. The dataset helps fine-tune existing vision models, enabling them to more accurately detect and interpret the unique movement patterns of individuals with prosthetic limbs, thereby supporting better prosthesis design, alignment, and rehabilitation.
A groundbreaking new resource, the ProGait dataset, has been introduced to significantly advance how artificial intelligence (AI) understands and analyzes the movement patterns of individuals using transfemoral prosthetic legs. This multi-purpose video dataset aims to bridge a critical gap in current vision-based machine learning models, which often struggle to accurately detect and analyze people with prostheses due to their unique appearance and movement.
Prosthetic legs are vital for individuals with lower-limb amputations, enabling them to regain mobility and improve their quality of life. Central to optimizing these devices is gait analysis, which provides crucial insights into a user’s walking patterns. Traditionally, this has relied on expensive and intrusive motion capture systems. Vision-based machine learning offers a scalable and non-invasive alternative, but existing models, primarily trained on data from able-bodied individuals, fall short when it comes to prosthetic users.
Introducing ProGait: A Comprehensive Dataset
The ProGait dataset addresses this challenge by providing 412 video clips from four above-knee amputees. These individuals were filmed while testing multiple newly-fitted prosthetic legs during walking trials. The dataset captures the presence, contours, poses, and gait patterns of subjects with transfemoral prosthetic legs, offering diverse scenarios including walking inside parallel bars without assistance and walking in hallways outside parallel bars with safety support.
Each walking trial includes both frontal and sagittal views, providing comprehensive perspectives for analysis. To ensure broad applicability, the trials feature various types and configurations of prosthetic legs, different background contexts, and varied lighting conditions. Importantly, the dataset also covers a range of normal and abnormal gait patterns, each accompanied by detailed textual descriptions from rehabilitation science and human engineering researchers. These descriptions outline specific gait deviations, suggest corrective adjustments for the prosthesis, and explain the reasons behind these recommendations.
Multi-Purpose Applications and Annotations
ProGait is designed to support three key vision tasks:
-
Video Object Segmentation (VOS): Accurately identifying and outlining the human subject, including the prosthetic leg, within video frames.
-
2D Human Pose Estimation (HPE): Pinpointing key anatomical points on the body and prosthesis to track movement.
-
Gait Analysis (GA): Classifying different walking patterns and deviations.
The dataset features high-quality annotations for these tasks, generated through a semi-automated, human-in-the-loop pipeline. This process involved using advanced AI models like GroundingDINO and SAM2 for initial annotations, followed by meticulous manual refinement to ensure accuracy, especially for detecting the prosthesis and distinguishing the subject from healthcare staff.
For gait analysis, researchers provided detailed textual descriptions for each video, categorizing general gait patterns and specific deviations. This rich annotation allows for advanced applications beyond simple classification, such as prosthesis optimization and clinical decision-making.
Benchmarking and Promising Results
To demonstrate the practical application and performance of ProGait, the researchers established benchmark tasks and provided fine-tuned baseline models. These models, including YOLO11 for object segmentation, RTMPose for pose estimation, and a custom LSTM classifier for gait classification, were trained and evaluated on the dataset.
The results are highly encouraging. Fine-tuned models using ProGait significantly outperformed pre-trained vision models in prosthesis-specific tasks. For instance, the RTMPose-ProGait model showed a substantial improvement in 2D Human Pose Estimation, particularly for lower body keypoints, demonstrating its effectiveness in detecting prostheses as part of the human body. In gait classification, the LSTM network performed exceptionally well when analyzing pose sequences from the sagittal view, suggesting it’s a more suitable angle for observing gait patterns.
Also Read:
- Unlocking the Secrets of Human Movement: A New AI Approach to Submovement Analysis
- AI-Powered Framework Enhances Home-Based Physical Rehabilitation
Future Implications
The ProGait dataset represents a significant step forward in making vision-based gait analysis more accessible and effective for individuals with prosthetic legs. By providing high-quality, specialized data, it aims to foster innovations in prosthetic engineering and rehabilitation. Looking ahead, the researchers plan to leverage the detailed textual descriptions of gait patterns with large language models (LLMs) to provide quick and convenient assessments for prosthesis users, potentially reducing the workload for healthcare professionals. Ultimately, this work paves the way for the development of intelligent, user-responsive prosthetic legs powered by microcomputers.
For more detailed information, you can refer to the full research paper: ProGait: A Multi-Purpose Video Dataset and Benchmark for Transfemoral Prosthesis Users.
