Méthodologie d'optimisation de la configuration matérielle-logicielle pour la détection d'objets dédiée aux véhicules autonomes

In recent years, integrating artificial intelligence (AI) into embedded systems has opened new horizons for real-time applications, including autonomous vehicles, smart surveillance, and robotics. However, deploying AI models in these embedded environments presents notable challenges, particularly the trade-offs between computational performance, energy efficiency, and model accuracy. This thesis, titled "Methodology for optimizing the hardware/software configuration of a system embedding artificial intelligence," proposes a comprehensive methodology for optimizing both the hardware and software configurations of embedded systems that implement AI technologies.The research focuses on object detection models, notably the YOLOv4 algorithm, tailored for resource-constrained architectures such as the Nvidia Jetson Xavier AGX and Jetson Nano. Key objectives include enhancing real-time processing capabilities, ensuring high precision, and minimizing resource consumption. The methodology incorporates model compression techniques, such as pruning and quantization, to reduce computational complexity while maintaining significant detection accuracy for critical objects, including cars, bicycles, and pedestrians, essential for autonomous driving applications.The proposed approach was validated through extensive experimental evaluations on the Jetson AGX Xavier and Jetson Nano platforms, demonstrating the feasibility of real-time performance under different optimization strategies. Specifically, the pruning method applied on the AGX achieved a 75 % accuracy at an inference speed of 31.3 FPS, while on the Jetson Nano it reached 5 FPS with an accuracy of 70.62 %. In parallel, post-training quantization (INT8) led to a significant speedup on the AGX, achieving 62.5 FPS, with a trade-off in accuracy reduced to 56.69 %. On the Jetson Nano, the same quantization strategy maintained a stable mean accuracy of 68.5 % at 5 FPS, highlighting its robustness in resource-constrained environments.This work contributes to the state-of-the-art by providing a scalable and systematic framework for AI optimization, paving the way for more intelligent and responsive embedded systems in real-world applications. The findings are particularly significant for developing safe and efficient autonomous vehicles, resonating with global efforts to enhance road safety and advance intelligent transportation systems.