CoDR: Correlation-based Data Reduction Scheme for Efficient Gathering of Heterogeneous Driving Data.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Advanced driver assistance systems (ADAS) rely on deep learning, necessitating vast amounts of heterogeneous driving data.
• The daily data volume (tens of GB) presents significant storage and network cost challenges.
• Existing methods struggle to efficiently manage and reduce this data deluge.

Purpose of the Study:

• To propose a novel scheme, CoDR (Correlation-based Data Reduction), for reducing heterogeneous driving data volume.
• To address the storage and network cost hurdles associated with large-scale ADAS data collection.
• To exploit correlations among different driving data types for efficient data reduction.

Main Methods:

• CoDR identifies a pivot dataset and relevant data ranges based on collection objectives.
• It analyzes correlations between heterogeneous datasets (videos, OBD-II, GPS).
• Irrelevant data is eliminated from both pivot and other datasets to reduce overall volume.

Main Results:

• CoDR successfully decreases data volume by up to 91%.
• Diverse analytical results reveal significant correlations among the four tested datasets.
• The findings demonstrate the practical utility of CoDR for on-the-spot data reduction in edge computing.

Conclusions:

• CoDR offers an effective solution for managing large volumes of heterogeneous driving data.
• Exploiting data correlations is key to reducing storage and network costs in ADAS.
• The proposed scheme has strong implications for efficient edge computing in intelligent transportation systems.