KotlinYOLO
Overview
KotlinYOLO is a real-time object detection system built for Android-based smart glasses,
focusing on efficient on-device inference using YOLO models in ONNX format.
The project explores the practical limits of edge AI deployment on constrained wearable hardware,
balancing latency, thermal constraints, and usability.
Key Features
- Real-time object detection using YOLOv8n / YOLO11n models
- ONNX Runtime integration using Java/Kotlin APIs
- CameraX pipeline with dynamic resolution handling
- Bounding box rendering on live camera feed
- Support for Vuzix M400 and Blade 2 devices
Architecture
The system follows a lightweight pipeline:
- Camera frame acquisition via
CameraX - Preprocessing into 640×640 tensor format
- ONNX model inference using
OrtSession - Post-processing (NMS, bounding box scaling)
- Overlay rendering to display detections
Technical Challenges
- Achieving usable FPS on low-power wearable hardware
- Handling aspect ratio mismatch without distortion
- Minimising memory allocations in real-time loops
- Balancing CPU vs GPU execution paths in ONNX Runtime
What I Learned
- Practical constraints of edge AI vs cloud inference
- Android performance optimisation at the system level
- Trade-offs between model size, accuracy, and latency
- Designing systems for real-world hardware limitations
Future Work
- Segmentation model support (YOLOv8-seg)
- Hardware acceleration improvements
- Server-assisted inference fallback
- Integration into production AR workflows