This work presents a comparative evaluation of advanced YOLO architectures for object detection, with a specific focus on their performance in traffic light detection for autonomous driving applications. Two deployment strategies were analyzed: a native implementation using PyTorch and a Model Based Engineering (MBE) implementation through automatic code generation. Evaluation metrics included precision-recall curves and confusion matrices across varying Intersection over Union (IoU) thresholds, as well as mean Average Precision (mAP) to assess detection quality and inference time measurements to evaluate computational efficiency on embedded platforms. The evaluation was based on a custom video extracted from the CARLA simulator, which was meticulously annotated by reviewing each frame to ensure the accuracy of the labeling. The study highlights the compromises between model accuracy and computational cost, providing a reproducible framework for performance benchmarking of object detection algorithms in safety-critical environments.

This paper evaluates the current maturity of automatic code-generation workflows for deploying modern CNN-based object detectors on embedded GPU platforms. We compare a native pipeline against a code generation pipeline through a Model-Based Engineering (MBE) approach, using YOLOv8/YOLOv9 inference on NVIDIA Jetson Orin Nano and Jetson AGX Orin as representative edge-GPU workloads. We report detection-quality metrics (mAP, PR curves) and system-level metrics (latency distribution and initialization overhead) under a controlled single-class scenario based on a CARLA-generated sequence with frame-level annotations. Absolute accuracy and latency values are scenario-dependent and may vary under different camera optics, illumination, motion blur, sensor noise, occlusion patterns, and multi-class scene. Results quantify the performance gap between code generation and native pipelines and show that, for the evaluated workloads, the automated pipeline remains less competitive in both latency and accuracy. We discuss the implications of this gap for deployment workflows in safety-oriented domains, and we outline bottlenecks that should be addressed. The study is intended as a controlled traffic-light detection micro-benchmark and does not aim to validate full ADAS perception stacks.