Research on target detection based on improved YOLOv7 in complex traffic scenarios.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Target detection is critical for intelligent vehicles and driver assistance systems.
• Existing algorithms struggle with real-time detection in complex road scenarios, necessitating a balance between efficiency and accuracy.

Purpose of the Study:

• To enhance the YOLOv7 target detection algorithm for improved performance in complex traffic environments.
• To balance computational efficiency and detection accuracy for real-time applications.

Main Methods:

• Utilized YOLOv7 as a baseline, incorporating deformable convolution and an attention mechanism module.
• Integrated a lightweight network module to accelerate computation and enhance feature expression.

Main Results:

• The enhanced model demonstrated improved detection capabilities in complex scenes.
• Achieved a 3.7% increase in average accuracy on the SODA 10M dataset compared to standard YOLOv7.
• Reached a mean average precision (mAP) of 75.9%.

Conclusions:

• The proposed modifications effectively improve real-time target detection in complex traffic scenarios.
• The integration of attention mechanisms and lightweight modules offers a superior balance of speed and accuracy.