Abstract
This paper analyses latency in a field-edge architecture for transport monitoring. The target use case is real-time vehicle detection and collision-risk monitoring, where inference speed and system throughput are operational requirements rather than secondary metrics.
Authors
Aistis Raudys, Lukas Baltramaitis, Robert Mackevič
What the study evaluates
The study evaluates how edge-side processing, batching, and parallel execution affect latency and frame rate in a YOLOv8-based detection pipeline. It focuses on whether the system can sustain real-time performance without sacrificing detection reliability.
Main findings
The proposed field-edge setup supports high-throughput detection by distributing work across local processing components. Batch processing and parallel threads help maintain frame rates above the level required for practical monitoring.
The results show that system architecture is as important as model accuracy for real-time AI applications. A strong detector still needs a deployment design that can keep latency within operational limits.
Why it matters
Transport monitoring systems need timely decisions. This work provides a practical architecture for running computer vision near the data source, reducing dependence on central processing and improving responsiveness in safety-sensitive environments.