Answers are grounded in this paper and cite their sources.
Embedded traffic-sign detection must balance accuracy against the compute, memory and power limits of the target platform. This study evaluates YOLOv8n on TT100K at six independently trained input resolutions (640×640, 512×512, 480×480, 320×320, 224×224 and 160×160) and on three platform configurations: an NVIDIA GeForce RTX 3050 laptop GPU, a Raspberry Pi 5 CPU and a Raspberry Pi 5 with a Google Coral USB Edge TPU. Validation accuracy decreased monotonically from 0.812 mAP@50 at 640×640 to 0.088 at 160×160. By contrast, measured Edge TPU throughput remained between 87.70 and 89.33 FPS, so downsampling produced no meaningful throughput gain on the evaluated configuration while substantially reducing accuracy. Normalized sign occupancy was then related to recall for all 7,471 annotations in the test subset. Recall increased from 20.50% below 0.005% occupancy to 98.17% above 0.10%, with a transition region between approximately 0.02% and 0.05% observed for this dataset and model configuration. The combined analysis provides a platform-specific basis for selecting input resolution and camera geometry.
No takes yet. Share an insight, caveat, or question.
Vinayakumar et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: