Recognition of a Virtual Scene via Simulated Prosthetic Vision

Area of Science:

• Biomedical Engineering
• Computer Vision
• Human-Computer Interaction

Background:

• Developing effective visual prosthetics is crucial for aiding the blind with low-resolution vision and visual recovery training.
• Simulated environments allow for controlled testing of visual prosthetic performance before real-world application.

Purpose of the Study:

• To evaluate the impact of varying low resolutions on pathfinding and object recognition performance in a simulated visual prosthetic scene.
• To determine optimal resolution thresholds for effective visual aid and training for individuals with visual impairments.

Main Methods:

• Virtual scenes (simple and complex) were created using 3DMAX and Unity, then pixelated to resolutions of 32x32, 64x64, and 128x128.
• Twenty subjects performed pathfinding and object recognition tasks within these simulated scenes.
• Recognition accuracy, time required, and number of collisions were recorded and analyzed.

Main Results:

• In simple scenes, increasing resolution from 32x32 to 64x64 significantly decreased recognition time (92.19s to 19.46s) and increased accuracy (51.22% to 96.89%).
• Collisions in simple scenes decreased from 10.00 to 1.00 with increased resolution.
• In complex scenes, increasing resolution from 48x48 to 128x128 improved room type recognition accuracy (65.69% to 85.69%) and reduced time (115.00s to 44.88s).

Conclusions:

• Resolution directly correlates with pathfinding and recognition performance in simulated prosthetic vision.
• A minimum resolution of 48x48 is required for complex scene recognition, while 32x32 is insufficient.
• Optimizing resolution enhances visual prosthetic effectiveness for aiding the blind.