The system is built upon

The system is built upon a dual-model hybrid architecture, balancing inference speed with spatial accuracy. The pipeline is heavily optimized to run on local hardware accelerators.

1. Object Detection (YOLOv8) Standardized on Ultralytics YOLOv8 for realtime bounding box generation. Fine-tuned to recognize critical indoor obstacles (doors, benches, people, structural barriers) while maintaining high FPS on edge hardware.

2. Metric Depth Estimation Utilizes Depth Anything V2 (Metric Indoor Small). Unlike relative depth models, this Vision Transformer (ViT) architecture outputs exact physical distances in meters, crucial for accurate collision calculations.

3. Edge Quantization (FP16) To preserve metric distance accuracy, the pipeline avoids INT8 quantization. Models are converted using FP16 (16-bit float) via ONNX/TFLite formats, doubling inference speed without degrading depth measurements.

4. Custom Python Pipeline A unified OpenCV-based tracking and inference loop. It manages frame buffering, selectively triggers the heavier depth model at optimized frame intervals, and computes the Collision Envelope mathematics.