A Hierarchical Real-Time Object Detection and Fine-Grained Classification Framework for Autonomous Driving
Autonomous driving systems require accurate and real-time perception of road users and traffic infrastructure to ensure safe vehicle operation in complex driving environments. This paper proposes a hierarchical YOLO11-2CNN perception framework for real-time object detection and fine-grained classification in autonomous driving applications. The proposed framework detects multiple road entities, including vehicles, pedestrians, cyclists, animals, potholes, speed breakers, traffic lights, and traffic signs using the YOLO11 object detector. However, fine-grained interpretation of small-scale traffic infrastructure objects, particularly traffic lights and traffic signs, remains challenging due to their limited size, visual similarity, and environmental complexity. To address this limitation, a lightweight CNN-based refinement module is integrated with the YOLO11 detector to perform detailed classification of traffic light states and traffic sign categories using the detected regions of interest. The novelty of the proposed YOLO11-2CNN framework integrates real-time YOLO11 object detection with two lightweight CNN refinement modules for traffic light state recognition and traffic sign classification, enabling improved fine-grained semantic interpretation while maintaining computational efficiency suitable for embedded autonomous driving platforms. The proposed framework was trained and evaluated using a custom real-world driving dataset containing diverse traffic participants and road infrastructure elements. Experimental results demonstrate that the enhanced YOLO11m model achieved a precision of 0.965, recall of 0.908, mAP@0.5 of 0.958, and mAP@0.5:0.95 of 0.816. The proposed framework demonstrated promising real-time deployment capability on an NVIDIA Jetson AGX Orin embedded platform. The experimental findings demonstrate that the proposed framework provides an effective balance between detection accuracy, fine-grained classification capability, and embedded real-time deployment suitability for autonomous driving applications.