Real-Time Evaluation of Perception Uncertainty and Validity Verification of Autonomous Driving

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Systems

Background:

• Deep neural networks excel at object detection but require real-time uncertainty evaluation for safe autonomous driving.
• Assessing the effectiveness and uncertainty of perception in real-time remains a critical research challenge.

Purpose of the Study:

• To propose a novel real-time evaluation method for deep neural network perception in autonomous vehicles.
• To analyze the spatial uncertainty of detected objects and its influencing factors.
• To validate the accuracy of spatial uncertainty estimation against ground truth data.

Main Methods:

• Developed a real-time evaluation method integrating multi-source perception fusion and deep ensemble techniques.
• Evaluated the effectiveness of single-frame perception results in real-time.
• Analyzed spatial uncertainty and its relationship with object distance and occlusion.

Main Results:

• Perception effectiveness evaluation achieved 92% accuracy.
• A positive correlation was observed between estimated uncertainty/error and ground truth.
• Spatial uncertainty was found to be influenced by object distance and occlusion degree.

Conclusions:

• The proposed method effectively evaluates perception effectiveness and spatial uncertainty in real-time.
• The findings contribute to enhancing the safety and reliability of autonomous driving systems.
• Understanding spatial uncertainty is key to improving object detection performance in complex scenarios.