Engineering of a High-throughput Multi-Camera Scale-invariant Collision Avoidance Pipeline

Kurzfassung

Modern object detection algorithms often work on fixed image sizes for their input and usually rely on scaling the input data to match the required measurements, resulting in issues with input data that significantly differs from these fixed dimensions.

This work illustrates a real-time visual pipeline that processes and leverages SAHI-based patch extraction with batched YOLO inference for scale-invariant object detection for a multi camera system with a centralized computing architecture. The developed solution contains optimization steps which enable high data throughput and is evaluated with regard to the complex dependencies between throughput and input delay.