[Navible: Navigation for Blinds] 1st place in Google Coral-board Challenge.

Ambient AI Bootcamp & Competition

I took the 1st place in Google Coral-board Challenge (Navible: Navigation for blinds) .

[Competetion description]

• Develop a Coral board AI application with a free topic in 3 weeks.
• Host: SNU GSDS.
• Sponsorship: Google Coral AI.

[Proposed Solution]

NaviBl Scenario:

Provide the visually impaired with correct directional guidance and hazard alerts based on real-time traffic conditions, leveraging a Coral board and auditory feedback.

To support a real-time scenario, we adopt the following strategies:

• Light-weight Models: We applied 8-bit quantization to two YOLOv5-based detection models and converted them to EdgeTPU-compatible formats to run on the Coral board.
• Efficient Resource Scheduling: Prioritized scheduling for vehicle detection (2.5 fps) and crosswalk/traffic light detection (1 fps).
• Computation Minimization: Reduce computation for feedback operations based on priority.

[Starategy 1 and 2] Light-weight model and system design for NaviBl with an efficient resource scheduling.

[Starategy 3] Priority based feedback algorithm.

[Prototypes]: Wearable device for Navible

[Results]