Towards Efficient Risky Driving Detection: A Benchmark and a Semi-Supervised Model.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Transportation Engineering

Background:

• Risky driving is a primary cause of traffic incidents, necessitating advanced detection methods.
• Intelligent Transportation Systems (ITS) require robust algorithms for real-time monitoring.
• A significant challenge in developing these algorithms is the scarcity of labeled data from traffic surveillance.

Purpose of the Study:

• To introduce Bayonet-Drivers, a novel benchmark dataset for risky driving detection.
• To propose DGMB-Net, a semi-supervised network architecture to overcome data limitations.
• To enhance the accuracy and generalizability of risky driving detection algorithms.

Main Methods:

• Development of the Bayonet-Drivers dataset using intelligent monitoring systems, capturing diverse traffic scenarios.
• Introduction of DGMB-Net, a semi-supervised network utilizing a teacher-student model to reduce reliance on labeled data.
• Integration of an Adaptive Perceptual Learning (APL) Module and Hierarchical Feature Pyramid Network (HFPN) within DGMB-Net to improve feature extraction and spatial perception.

Main Results:

• DGMB-Net demonstrated remarkable performance in detecting risky driving.
• The proposed methods effectively addressed the challenge of limited labeled data.
• Experiments on State Farm and Bayonet-Drivers datasets validated the network's effectiveness.

Conclusions:

• The Bayonet-Drivers dataset and DGMB-Net offer a significant advancement in risky driving detection.
• Semi-supervised learning combined with advanced network modules can overcome data scarcity issues.
• The developed system enhances the potential of Intelligent Transportation Systems for traffic safety.