A Small-Scale Object Detection Algorithm in Intelligent Transportation Scenarios.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Intelligent Transportation Systems

Background:

• Object detection models struggle with small-scale targets in complex intelligent transportation scenarios.
• Cluttered backgrounds and redundant information in traffic images hinder accurate detection.
• Data imbalance and suboptimal prior bounding box adaptation affect custom traffic datasets.

Purpose of the Study:

• To improve the detection accuracy of small and medium-sized targets in intelligent transportation.
• To enhance the model's focus on relevant traffic targets amidst cluttered backgrounds.
• To address data imbalance and improve bounding box adaptation for custom traffic datasets.

Main Methods:

• A fusion method based on the YOLOv4 tiny framework, enhancing shallow and mid-level feature utilization with Feature Pyramid Network (FPN).
• Integration of the Convolutional Block Attention Module (CBAM) to refine model attention to traffic targets.
• Implementation of an improved Copy-Paste data augmentation method and a K-means algorithm with modified distance measurement for dataset enhancement.

Main Results:

• The proposed algorithm demonstrated a 4.9% improvement in mean Average Precision (mAP) compared to the standard YOLOv4 tiny model.
• The enhanced model maintained real-time performance, crucial for intelligent transportation applications.
• Significant improvements were observed in detecting small and medium-sized targets within complex traffic environments.

Conclusions:

• The proposed fusion method effectively enhances feature utilization and model attention for improved small-scale target detection in intelligent transportation.
• The data augmentation and clustering techniques successfully address challenges related to custom traffic datasets.
• The algorithm offers a viable solution for real-time, accurate object detection in intelligent traffic systems.