VV-YOLO: A Vehicle View Object Detection Model Based on Improved YOLOv4.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Machine Learning

Background:

• Vehicle view object detection is critical for autonomous driving safety.
• Complex scenes (dim light, occlusion, long distance) pose challenges for current models.
• Existing YOLOv4 models require improvements for robust environmental perception.

Purpose of the Study:

• To propose an improved YOLOv4-based model, VV-YOLO, for enhanced vehicle view object detection.
• To address challenges in complex driving environments.
• To improve the accuracy and robustness of object detection for autonomous vehicles.

Main Methods:

• Implemented an anchor frame-based approach with an improved K-means++ algorithm for stable anchor clustering.
• Introduced a CA-PAN network with a coordinate attention mechanism in the neck for better feature extraction.
• Reconstructed the loss function using a focus mechanism to handle imbalanced training data.

Main Results:

• VV-YOLO achieved 90.68% precision and 80.01% average precision on the KITTI dataset, outperforming YOLOv4 by 6.88% and 3.44% respectively.
• The model demonstrated comparable computation time to YOLOv4.
• Validation on BDD100K and field-collected data confirmed the model's validity and robustness.

Conclusions:

• The proposed VV-YOLO model significantly improves vehicle view object detection in complex scenarios.
• The integration of coordinate attention and a focus-based loss function enhances model performance and training stability.
• VV-YOLO offers a robust and efficient solution for environmental perception in autonomous vehicles.