Fast and accurate object detector for autonomous driving based on improved YOLOv5

Area of Science:

• Artificial Intelligence
• Computer Vision
• Robotics

Background:

• Real-time object detection is crucial for the safety and stability of autonomous vehicles.
• Existing object detection algorithms face challenges in achieving both high accuracy and real-time performance for autonomous driving applications.

Purpose of the Study:

• To propose a fast and accurate object detection model for autonomous driving systems.
• To enhance the performance of the YOLOv5 algorithm for improved vehicle and pedestrian recognition.

Main Methods:

• Structural re-parameterization (Rep) was employed to decouple training and inference, improving model speed and accuracy.
• Neural architecture search was utilized to optimize multi-branch re-parameterization modules, enhancing training efficiency.
• A small object detection layer and coordinate attention mechanism were integrated to boost the detection of smaller objects like pedestrians and vehicles.

Main Results:

• The proposed improved YOLOv5 model achieved a detection accuracy of 96.1% on the KITTI dataset.
• The model demonstrated a high inference speed of 202 frames per second (FPS).
• The enhanced model outperformed several mainstream object detection algorithms in terms of accuracy and real-time performance.

Conclusions:

• The developed object detection method significantly improves accuracy and real-time capabilities for autonomous driving.
• The integration of Rep, NAS, and attention mechanisms offers a promising approach for advanced autonomous vehicle perception systems.