MIO-TCD: A new benchmark dataset for vehicle classification and localization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning models require large labeled datasets for effective training.
• Accurate motor vehicle classification and localization are crucial for intelligent transportation systems.

Purpose of the Study:

• Introduce the MIOvision Traffic Camera Dataset (MIO-TCD), the largest dataset for motorized traffic analysis.
• Evaluate the performance of deep learning architectures for vehicle classification and localization using MIO-TCD.

Main Methods:

• Developed MIO-TCD with 786,702 annotated images across 11 traffic object classes.
• Dataset comprises a localization set (137,743 frames) and a classification set (648,959 crops).
• Benchmarked state-of-the-art deep learning models on the MIO-TCD dataset.

Main Results:

• Top models achieved >96% accuracy and Kappa score on classification, and 77% mAP on localization.
• Demonstrated the viability of deep learning for single-frame vehicle analysis in real-world traffic.
• Identified limitations of current methods in specific traffic scenarios.

Conclusions:

• MIO-TCD is a valuable resource for advancing research in traffic analysis and deep learning.
• Further research is needed to address identified challenges in vehicle classification and localization.
• The dataset and challenge results are publicly available to foster further development.