Text to Blind Motion

Area of Science:

• Computer Vision
• Robotics
• Human-Computer Interaction

Background:

• Sighted individuals' motion data dominates current 3D human motion models.
• Blind pedestrians exhibit distinct movement patterns not captured by existing models.
• Autonomous systems need to understand diverse human motion for safety.

Purpose of the Study:

• Introduce the first multimodal motion benchmark for blind pedestrians.
• Evaluate the performance of state-of-the-art 3D human motion prediction models on this new data.
• Promote the development of inclusive AI systems that can reason about diverse human movements.

Main Methods:

• Collected 3D motion data from 11 blind participants using wearable sensors in urban environments.
• Developed a multimodal benchmark including rich textual descriptions of pedestrian behavior and environmental interactions.
• Benchmarked existing 3D human motion prediction models on the novel dataset.

Main Results:

• Existing 3D human motion prediction models show poor performance on data from blind pedestrians.
• Off-the-shelf and pre-trained models failed to accurately capture the distinct motion characteristics.
• The BlindWays benchmark reveals significant limitations in current motion modeling.

Conclusions:

• Current 3D human motion models are not sufficiently diverse or robust for blind pedestrians.
• The BlindWays dataset is crucial for developing safer and more reliable AI systems.
• Further research is needed to improve AI's understanding of diverse human mobility.