In the automotive industry, the use of biometric technologies has provided safer and more reliable access to vehicles compared to traditional methods such as keys. However, state-of-the-art biometric systems rely on complex deep neural networks that demand significant computational resources, both in terms of memory and processing power. This creates a critical challenge when deploying such systems on resource-constrained automotive hardware with strict real-time requirements. This thesis is part of a larger project, developed by Politecnico di Torino in collaboration with Stellantis, whose objective is to recognize vehicle owners through a multimodal biometric system based on face and gait recognition. This thesis work aims to adapt the pipeline to the capabilities of embedded automotive platforms, using optimization techniques to reduce both memory footprint and inference time, all while preserving accuracy.