Traffic sign recognition algorithm: a deep comparison between Yolov5 and SSD Mobilenet

One of the most intriguing and fascinating challenges of our century is the realization of the dream of making personal transportation a totally effortless experience. In this context, the reliability and accuracy of selected algorithms play a crucial role, as they enable vehicles to make immediate and accurate decisions, opening the door to a future of innovative transportation. The technical support provided by the MCA company has proved crucial to this project. The ultimate goal is the creation of a rover capable of moving autonomously, emulating fully autonomous driving, especially over rough terrain. This collaboration has allowed our research project to deepen the development of dedicated software for an embedded system that leverages artificial intelligence to perform object recognition and tracking, bringing with it promising prospects for transportation. In the initial stages, we conducted a thorough analysis of existing code to identify common features and establish a solid foundation on which to further build our work. Next, we performed a detailed examination of the selected neural networks and the chosen dataset, diligently seeking a balance between the GPU processing resources available on the Nvidia Jetson Nano card and achieving satisfactory performance results. The first objective of this research, therefore, focused on training two leading object recognition neural networks, namely SSDmobileV2 and YOLOv5. This challenge required the use of a large dataset, including a wide range of distinct traffic signals, divided into four main categories: Traffic Light, Stop, Speed Limit, and Crosswalk. Subsequently, this allowed us to test the best models resulting from both networks in order to compare them in detail. This evaluation phase aims to assess the performance of these neural networks in the context of object recognition in realistic and challenging scenarios, such as those that a neural network in a vehicle must handle effectively. This study, therefore, represents a significant step in researching the main strengths of the two algorithms applied to object detection, using a data set specifically created for the project. It makes it possible to simulate real-world situations in which a neural network in a vehicle must demonstrate its efficiency and reliability, opening up new perspectives for the future of autonomous transportation.