ZZ-YOLOv11: A Lightweight Vehicle Detection Model Based on Improved YOLOv11.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Urban traffic scenarios present challenges for vehicle detection due to complex lighting and occlusions.
• Existing detection algorithms struggle with accuracy, high error rates, and computational demands.
• A need exists for more efficient and accurate lightweight vehicle detection networks.

Purpose of the Study:

• To propose an improved lightweight vehicle detection network, ZZ-YOLO, addressing accuracy and computational burden.
• To enhance object edge focus and feature fusion in detection models.
• To reduce model parameters while maintaining high detection performance.

Main Methods:

• Developed a Global Edge Information Transfer (GEIT) module for improved feature representation.
• Introduced a Lightweight Detail Convolutional Detection Head (LDCD) for parameter reduction and feature fusion.
• Applied layer-adaptive magnitude-based pruning (LAMP) and model distillation (YOLOv11x + LDCD as teacher).

Main Results:

• ZZ-YOLO achieved 70.9% detection accuracy and 58% mAP@0.5 on KITTI and BDD100K datasets.
• Demonstrated a 5.7% increase in detection accuracy and 2.3% increase in average precision compared to the original algorithm.
• Reduced model parameters by 34% and effectively decreased misdetection and omission rates in real-vehicle tests.

Conclusions:

• The proposed ZZ-YOLO network offers a significant improvement in lightweight vehicle detection for urban traffic.
• The integration of GEIT and LDCD modules, along with pruning and distillation, enhances both accuracy and efficiency.
• ZZ-YOLO effectively addresses limitations of current methods, showing promise for real-world traffic monitoring applications.