Assessing YOLO models for real-time object detection in urban environments for advanced driver-assistance systems (ADAS)

Advanced Driver Assistance Systems (ADAS) rely on precise and fast object detection to ensure safe navigation in dynamic urban environments. However, achieving an optimal balance between detection accuracy, real-time processing speed, and model efficiency remains a significant challenge. Recent advancements in convolutional neural networks have significantly enhanced the state-of-the-art in computer vision, particularly in the domain of object detection. This progress has been instrumental in advancing the capabilities of ADAS through improved recognition of objects in urban outdoor environments. Among the plethora of object detection frameworks, the "You Only Look Once" (YOLO) series stands out for its ability to provide a delicate balance between detection accuracy and real-time processing speed. This paper endeavours to conduct a comprehensive evaluation of the various iterations of the YOLO models, employing the transfer learning approach to assess their efficacy in urban object detection tasks pertinent to ADAS. Utilizing the BDD100K dataset, a benchmark for urban environment object detection, this research meticulously trained and evaluated several YOLO versions, focusing on their detection performance, processing speed, and model scalability. The findings indicate that YOLO v9 offers an optimal compromise between speed and accuracy, positioning it as a viable model for real-time ADAS applications. Conversely, while YOLO v1 excels in processing speed, it lacks the accuracy provided by its successors. The study concludes that while certain YOLO models demonstrate potential for ADAS implementation, others fall short, emphasizing the importance of selecting models that achieve a harmonious balance among speed, accuracy, and model size for practical ADAS applications in the real world.