Infrared object detection via feature interaction and attention-guided fusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Driving Systems

Background:

• Object detection is crucial for autonomous driving environmental perception, but struggles with false positives/negatives in complex scenes.
• Infrared imaging aids low-light conditions, yet conventional methods face challenges with small, occluded objects due to limited semantic information and low contrast.
• Accurate detection of infrared occluded and small objects remains a significant challenge for real-time autonomous driving.

Purpose of the Study:

• To address the ineffectiveness of current object detection algorithms for infrared occluded and small objects in complex road scenarios.
• To propose an improved object detection method for complex infrared scenes (ODMCIS) based on the YOLOv8s architecture.
• To enhance real-time traffic object detection capabilities for all-day autonomous driving.

Main Methods:

• Designed a dual-branch (DB)-spatial pyramid pooling fast (SPPF) module and a dual-residual branch (DRB)-C2f module to improve multi-scale feature enhancement and fusion.
• Developed a novel loss function to accelerate model convergence and boost detection accuracy.
• Implemented a cross-level fusion mechanism and redesigned the detection head within the YOLOv8n network for more precise object localization.

Main Results:

• The proposed ODMCIS-YOLO algorithm demonstrated high precision in object detection.
• Achieved a reasoning speed of 137.6 frames per second (FPS), meeting real-time detection requirements.
• Outperformed state-of-the-art (SOTA) object detection methods in comprehensive performance for complex infrared scenes.

Conclusions:

• The ODMCIS-YOLO method effectively enhances the detection of infrared occluded and small objects in complex road scenes.
• The algorithm's high precision and speed provide a foundation for reliable all-day autonomous driving.
• This improved object detection approach offers a more efficient solution for traffic object recognition in challenging environments.