Small Object Detection in Traffic Scenes Based on YOLO-MXANet.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• General object detection algorithms struggle with small objects in traffic scenes, exhibiting low accuracy, high complexity, and slow speeds.
• Addressing these limitations is crucial for advanced driver-assistance systems and autonomous driving.

Purpose of the Study:

• To develop an improved object detection algorithm (YOLO-MXANet) for enhanced performance on small objects in traffic scenarios.
• To reduce model complexity and increase detection speed without compromising accuracy.

Main Methods:

• Utilized Complete-Intersection over Union (CIoU) loss function to improve small object positioning accuracy.
• Introduced a lightweight backbone network (SA-MobileNeXt) with channel and spatial attention, incorporating Shuffle Channel and Spatial Attention (SCSA) into SandGlass Blocks (SGBlock).
• Employed Mosaic and Mixup data augmentation, a Multi-scale Feature Enhancement Fusion (MFEF) network, and SiLU activation functions within Convolution-Batchnorm-Leaky ReLU (CBL) and SGBlock modules.

Main Results:

• Ablation experiments on the KITTI dataset confirmed the effectiveness of each proposed improvement.
• The YOLO-MXANet algorithm demonstrated reduced model complexity and faster detection speeds.
• Achieved improved object detection accuracy for small objects in traffic scenes.

Conclusions:

• The YOLO-MXANet algorithm effectively addresses the challenges of detecting small objects in traffic scenes.
• The proposed enhancements lead to a more efficient and accurate object detection system.
• Comparative experiments on KITTY and CCTSDB datasets show YOLO-MXANet's advantages over existing algorithms.