Deep learning method for risk identification of autonomous bus operation considering image data augmentation...

Area of Science:

• Autonomous driving technology
• Road traffic safety
• Artificial intelligence in transportation

Background:

• Autonomous buses represent a critical application for self-driving technology.
• Identifying operational risks is crucial for enhancing road safety and facilitating widespread adoption of autonomous vehicles.

Purpose of the Study:

• To identify risks associated with autonomous bus operations.
• To evaluate the effectiveness of deep learning and image augmentation strategies for risk identification.

Main Methods:

• Converted Shanghai autonomous bus operational data into grayscale and radar images.
• Applied the AlexNet deep learning convolutional neural network for image recognition.
• Utilized image data augmentation strategies, including Fourier transform, to address sample imbalance.

Main Results:

• Achieved optimal risk identification accuracy (ACC) of 90.4%, true positive rate (TPR) of 83.7%, and false negative rate (FPR) of 94.58%.
• AlexNet performance was enhanced with Fourier images.
• Grayscale images yielded higher accuracy than radar images for risk identification.

Conclusions:

• Autonomous buses face heightened risks in turning sections and intersections.
• Fourier transform effectively mitigates uneven sample distribution issues.
• Input time series length significantly influences risk identification accuracy.