Novel Metric to Evaluate Safety of Environmental Perception in Autonomous Driving
This paper introduces a new safety metric called EPSM to evaluate the safety of perception systems in autonomous driving, focusing on object and lane detection.
Why it matters
Evaluating the safety of perception systems is crucial for ensuring the safety of autonomous vehicles in complex driving scenarios.
Key Points
- 1Conventional performance metrics like precision, recall, and F1-score do not capture safety-critical aspects of perception
- 2EPSM integrates an object safety metric and a lane safety metric to provide a unified measure of perception safety
- 3EPSM can identify safety-critical perception errors that conventional metrics miss
- 4Evaluated on the DeepAccident dataset, EPSM demonstrates the importance of safety-centric evaluation for autonomous driving perception systems
Details
The paper proposes a novel safety metric called EPSM (Environmental Perception Safety Metric) to evaluate the safety of perception systems in autonomous driving. Conventional performance metrics like precision, recall, and F1-score assess overall detection accuracy but do not consider safety-relevant aspects. This can lead to perception systems with high scores that still have critical errors that could cause accidents. EPSM integrates an object safety metric to quantify the potential risk of object detection errors, as well as a lane safety metric that accounts for the interdependence between object and lane detection. The combined EPSM score provides a unified, interpretable measure of perception safety. Using the DeepAccident dataset, the authors demonstrate that EPSM can identify safety-critical perception errors that conventional metrics fail to capture, emphasizing the importance of safety-centric evaluation for autonomous driving perception systems.
No comments yet
Be the first to comment