The development of deep learning and its application in 2D object detection in the last decade is undoubtedly astonishing. However, almost all models and architectures require extensive computational resources, including in the training phase and deployment. This issue has led to long inference times, especially on embedded systems, and limited the wide deployment of deep learning-based applications. This paper presents an expersimental study on adopting quantization and optimization to produce lightweight 2D object detection models that significantly reduce the inference time of the model while still maintaining its performance, with or without a little trade-off in accuracy. Moreover, some experiments have been conducted and presented as proof of the effectiveness of measuring Jetson Xavier NX embedded systems.