The widespread deployment of artificial intelligence in surveillance systems raises many questions about the explainability of these decision models, which are essential for understanding and trusting the information provided by these systems. Other challenges related to model learning, such as energy consumption and access to limited annotated learning data, remain largely unexplored in this research field. This includes the choice of sensors to cope with variable weather conditions and reduce the system’s energy consumption. The integration of these sensors requires adapting the learning models, offering an innovative approach compared to traditional surveillance systems that primarily rely on cameras or intrusive sensors placed directly on the road, which often require costly and difficult-to-implement road infrastructure adjustments.