RSO-YOLO: A Real-Time Detector for Small and Occluded Objects in Autonomous Driving Scenarios

Area of Science:

• Computer Vision
• Autonomous Systems
• Machine Learning

Background:

• Detecting small and occluded objects is a critical challenge in autonomous driving.
• Existing models struggle with the complexity and variability of real-world driving scenarios.

Purpose of the Study:

• To develop an enhanced object detection model, RSO-YOLO, for improved performance in autonomous driving.
• To specifically address the limitations in detecting small and occluded objects.

Main Methods:

• RSO-YOLO is based on YOLOv12, incorporating a bidirectional feature pyramid network (BiFPN) and space-to-depth convolution (SPD-Conv).
• A detection head for the P2 feature layer and a feature enhancement and compensation module (FECM) were added to improve small and occluded object detection.
• A lightweight global cross-dimensional coordinate detection head (GCCHead) was developed to balance efficiency and performance.

Main Results:

• RSO-YOLO achieved significant mAP@0.5 improvements: 8.0% on SODA10M, 10.7% on BDD100K, and 7.2% on FLIR ADAS compared to YOLOv12.
• The model reduced parameters by 15.4% and computational complexity by 20%.
• Enhanced detection accuracy and robustness under occlusion were demonstrated.

Conclusions:

• RSO-YOLO offers superior object detection capabilities for autonomous driving systems.
• The model provides a practical solution by increasing accuracy while decreasing computational demands.
• RSO-YOLO shows strong potential for real-world autonomous driving applications.