Uncertainty Measurement as a Sensor Performance Metric in Adverse Conditions

In autonomous driving, object detection remains an ongoing challenge, as solutions must be reliable and safe to use on the road. Current state-of-the-art solutions provide point estimate predictions, which may be inadequate, as reports of deaths due to autonomous vehicle collisions continue to appear. To increase the safety of these algorithms, uncertainty estimation can be incorporated into traditional object detection pipelines to produce a measure of uncertainty alongside traditional point estimate object predictions. To evaluate uncertainty as a performance metric, signal detection theory is leveraged to generate an uncertainty based classifier. The classifier can be used to track the ratio of false positive to true positive detections, defined as the anomalous detections ratio. Findings from this work indicate that the novel metric is responsive to degraded driving conditions including night time driving and lens obstructions for the RGB camera, while the LiDAR is responsive to snowfall and simulated rain conditions.M.A.S