An exercise in sensor fusion and accessible design.
The problem of navigation for the visually impaired is typically treated as a clear-path problem. However, the greater challenge is context awareness. A cane can detect a wall, but it cannot read a sign or detect a silent gesture. The hypothesis was that a multimodal system (Vision + Sonar) could offer a "semantic cane" experience.
The ultrasonic walking stick proved robust, providing reliable < 2m obstacle detection. However, the Smart Cap (Vision) faced significant latency issues on the Pi 4 when running simultaneous object detection and OCR.
We optimized the model using quantization, reducing inference time by 40%. The final system achieved 85% accuracy in real-time sign language translation under controlled lighting.
The project successfully secured ₹335,000 in funding, validating the market need. The "semantic" layer is viable but requires dedicated AI accelerators (like a Coral stick) for true real-time comfort.