An object detection algorithm combining self-attention and YOLOv4 in traffic scene

Area of Science:

• Computer Vision
• Artificial Intelligence
• Automotive Engineering

Background:

• Intelligent automobiles rely on environment perception for safety, with object detection being critical for autonomous vehicles.
• Challenges in real-world traffic, including occlusion, small objects, and adverse weather, hinder accurate object detection.
• Existing methods struggle with the complexities of diverse traffic scenarios.

Purpose of the Study:

• To propose an improved object detection algorithm for intelligent vehicles.
• To enhance the accuracy and robustness of detecting vehicles and pedestrians in complex traffic scenes.
• To address the limitations of current object detection methods in special conditions.

Main Methods:

• Developed the SwinT-YOLOv4 algorithm, integrating Swin Transformer with the YOLOv4 architecture.
• Replaced the Convolutional Neural Network (CNN) backbone of YOLOv4 with the Swin Transformer.
• Retained the feature-fusing neck and predicting head components of YOLOv4.
• Trained and evaluated the model on the COCO dataset.

Main Results:

• The SwinT-YOLOv4 algorithm demonstrated significant improvements in object detection accuracy under special conditions.
• Detection precision for cars and persons increased by 1.75% compared to baseline methods.
• Achieved high detection precisions: 89.04% for cars and 94.16% for persons.
• The Vision Transformer proved effective in extracting object features.

Conclusions:

• The proposed SwinT-YOLOv4 algorithm enhances object detection capabilities for intelligent vehicles.
• The integration of Swin Transformer addresses key challenges in traffic scene perception.
• This advancement contributes to improved driving safety in autonomous systems.