In this thesis we present a complete pipeline for real-time traffic stop-sign detection on the STM32H747I-DISCO embedded system equipped with a B-CAMS-OMV camera module. Starting from the pretrained ST Yolo LC v1 person detector, we use transfer learning to derive a single-class stop-sign detector, quantize it to an INT8 representation, and automatically generate C code for the board’s Arm Cortex-M7 core using the ST Edge AI tool. While quantizing the 256^2 model we discover objectness-logit saturation in the YOLO head; we propose a fix that rescales the γ and β parameters of the final batch-normalization layer, raising the INT8 mAP from 32.4% to 42.3%. We evaluate the 192 × 192 and 256 × 256 models on our own field dataset of 118 photographs of Slovenian stop signs. The best model reaches 41.6% mAP@0.5, only 0.7 percentage points below its result on the COCO 2017 validation split (42.3%), indicating a negligible domain gap between the training and target domains.
|
