Introspective False Negative Prediction for Black-Box Object Detectors in Autonomous Driving

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Object detection is vital for autonomous driving safety.
• Current detectors have limitations and can fail in complex scenarios.
• False Negative (FN) object detection failures pose significant risks but lack sufficient online prediction methods.

Purpose of the Study:

• To develop a general framework for online prediction of FN objects for black-box object detectors.
• To address the critical need for mitigating risks associated with missed object detections in autonomous vehicles.
• To improve the reliability and safety of automated driving systems.

Main Methods:

• Proposed a general introspection framework for online FN object prediction.
• Designed an introspective FN predictor for feature extraction, moving beyond empirical assumptions.
• Introduced a multi-branch cooperation mechanism to handle the foreground-background imbalance specific to FN detection.

Main Results:

• Achieved 81.95% precision and 88.10% recall in predicting FN objects on the KITTI Benchmark.
• Demonstrated the framework's effectiveness in identifying missed detections.
• Showcased significant improvement in overall object detection performance by incorporating FN predictions.

Conclusions:

• The proposed introspection framework effectively predicts FN objects in real-time.
• The method enhances autonomous driving safety by addressing a critical failure mode.
• This approach offers a robust solution for improving the dependability of object detection systems.