Edge-TSR boosts roadside AI accuracy by 10% on Jetson Orin Nano

A new paper by Aditya Mishra and Haroon Lone exposes a critical blind spot in edge AI evaluation: benchmark-centric testing systematically overstates real-world performance. Their system, Edge-TSR, targets fine-grained roadside perception on the resource-constrained NVIDIA Jetson Orin Nano. Testing three state-of-the-art baselines, the authors observed a consistent 20-30% relative degradation when moving from static-image evaluation to continuous streaming deployment. Causes include temporal instability in video, thermal throttling under sustained load, and workload-dependent variability—effects invisible to standard benchmarks.

Edge-TSR addresses this gap with a lightweight, track-aware temporal stabilization mechanism that integrates detection, tracking, and fine-grained classification. This improves streaming inference consistency with negligible overhead, recovering up to 10.16% classification accuracy over per-frame baselines. In a 55-minute vehicular deployment over a 26-km route, Edge-TSR sustained 16.18 FPS within safe thermal limits on a single embedded device, without any cloud offload. The findings underscore that deployment-aware evaluation and temporal stabilization are essential for continuously operating edge AI in real-world sensing environments.