Methodology for an Analysis of Influencing Factors on 3D Object Detection Performance
This work addresses the need for safety assurance in automated driving by providing a methodology to identify performance gaps in black-box deep learning detectors, though it is incremental as it builds on existing statistical and machine learning techniques.
The paper tackles the problem of understanding how object- and environment-related factors influence the performance of LiDAR- and camera-based 3D object detectors in automated driving, using statistical univariate analysis and a Random Forest model with Shapley Values to analyze pedestrian detection errors and feature dependencies.
In automated driving, object detection is crucial for perceiving the environment. Although deep learning-based detectors offer high performance, their black-box nature complicates safety assurance. We propose a novel methodology to analyze how object- and environment-related factors affect LiDAR- and camera-based 3D object detectors. A statistical univariate analysis relates each factor to pedestrian detection errors. Additionally, a Random Forest (RF) model predicts errors from meta-information, with Shapley Values interpreting feature importance. By capturing feature dependencies, the RF enables a nuanced analysis of detection errors. Understanding these factors reveals detector performance gaps and supports safer object detection system development.