A Real-Time Object Detector for Autonomous Vehicles Based on YOLOv4

Area of Science:

• Computer Vision
• Autonomous Driving Technology
• Deep Learning

Background:

• Real-time and accurate object detection is critical for autonomous driving safety.
• Balancing detection speed and accuracy remains a significant research challenge.

Purpose of the Study:

• To develop a one-stage object detection algorithm that enhances both accuracy and real-time performance for autonomous vehicles.
• To optimize the YOLOv4 architecture for improved efficiency and effectiveness in object detection tasks.

Main Methods:

• Modified the YOLOv4 backbone by increasing residual block stacking in CSPDarkNet53.
• Replaced SPP with RFB in the neck, enhanced PAN for feature fusion, and integrated CBAM and CA attention mechanisms.
• Reduced network width to 75% of the original to decrease parameters and boost inference speed.

Main Results:

• Achieved a 2.06% average accuracy improvement on the KITTI dataset and 2.95% on the BDD dataset compared to YOLOv4.
• Increased inference speed by 9.14% while maintaining comparable detection accuracy.
• Demonstrated real-time detection capabilities exceeding 58.47 FPS.

Conclusions:

• The proposed algorithm effectively balances speed and accuracy for autonomous driving object detection.
• The modifications to the YOLOv4 architecture offer a viable solution for real-time, high-performance vehicle perception systems.