Deep learning-driven CNN model for detection and classification of dynamic obstacles
Hamza Assemlali, Soukaina Bouhsissin, Nawal Sael
Abstract
: In recent years, advancements in autonomous driving technologies have shown significant potential to improve road safety, although their current impact remains a subject of ongoing research and debate. At the core of this improvement is obstacle detection, a crucial element for ensuring the effective functioning and security of these systems. Despite substantial progress, challenges remain in the detection and classification of dynamic road obstacles, which pose greater risks than fixed ones; both drivers and the obstacles in question can suffer fatal consequences because of collisions with dynamic obstacles. This highlights the urgent need for systems that can alert drivers in advance, thereby preventing accidents. The current paper presents a deep learning-driven approach to detect and classify dynamic obstacles, including pedestrians, vehicles, and animals. By integrating data from various datasets, we developed and adapted CNN architecture for this task, enhancing the reliability and safety of autonomous driving systems. This approach demonstrated significant improvements under various conditions. We have gotten a classification accuracy of 99.5% and a detection precision of 97.1%. The results indicate that our architecture offers a better ability to identify and classify road obstacles, thus contributing to the advancement of autonomous driving technologies.