Optimizing YOLO Inference for Hardware Constraints Through Quantization Techniques

This thesis investigates the application of YOLO (You Only Look Once) models for object detection tasks, with a particular focus on the quantization of such models to enable efficient deployment on edge devices and resource-constrained hardware platforms. Model quantization plays a critical role in reducing memory footprint and computational cost while aiming to preserve the accuracy and robustness of the original floating-point networks. The work focuses on integrating a complete training and evaluation pipeline, including data pre-processing compliant with widely adopted standards (e.g. YOLO) and the integration of automated tools for ground truth visualization and validation. Various training strategies were explored to enhance model performance, includ- ing hyperparameter tuning, architectural modifications, and data augmentation techniques. A central contribution of the work is the design of a modular quantization workflow, leveraging tools compatible with ONNX and tailored for deployment with hardware-accelerated inference platforms. The methodology includes model export, transformation, optimization, and performance validation of the quantized networks. Experimental results, obtained from both standard benchmarks and domain- specific datasets, demonstrate that the proposed approach achieves a favorable trade-off between model compactness and detection accuracy. These findings support the feasibility of adopting quantized YOLO architectures in real-world, real-time applications across diverse environments. The workflow proposed, along with the solid results obtained by this work, offers a valid stepping stone for future improvements and optimizations.