A Lightweight Vehicle Detection Method Fusing GSConv and Coordinate Attention Mechanism

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traffic scenarios pose challenges for target detection due to numerous parameters and high computational costs.
• Existing models often struggle with a balance between detection speed, accuracy, and application cost.

Purpose of the Study:

• To develop an enhanced lightweight real-time detection algorithm for vehicle detection in traffic scenarios.
• To address the limitations of high parameter counts, computational burden, and application costs in current models.

Main Methods:

• Utilized YOLOv7 as the benchmark and integrated MobileNetV3 for feature extraction.
• Designed a lightweight SPPFCSPC-GS module incorporating GSConv and the CA mechanism.
• Employed the MPDIoU loss function for optimized model training.

Main Results:

• Achieved 98.2% mean Average Precision (mAP) on the BIT-Vehicle dataset.
• Reduced model parameters by 52.8% compared to the original YOLOv7.
• Improved Frames Per Second (FPS) by 35.2%.

Conclusions:

• The enhanced YOLOv7 model offers a superior balance between detection speed and accuracy.
• The lightweight design facilitates deployment on mobile devices with limited resources.
• This algorithm provides an efficient solution for real-time vehicle detection in intelligent transportation systems.