Quantification of Reliable Detection Range Using Lidar-Based Object Detection Approaches and Varying Process Parameters

Highly automated or autonomous systems utilize machine learning-based approaches to perceive the environment and make decisions based on the obtained information. These approaches are highly depended on the model used, training data, and environmental conditions. While high performance can be expected in certain and suitable situations, unknown or uncertain situations might lead to undetected underperformance. In case of object detection, the quantification of reliability of a particular prediction is usually given by a detection score predicted by the trained model. While a higher score indicates higher confidence, it does not reflect the actual uncertainty of the prediction considering situational variations. In this paper, the predicted detection score is converted into a true-positive rate to decide whether to accept or reject a prediction. Furthermore, the detection rate is used as a function over distance to determine the dynamic performance expectations using situational knowledge and reliability requirements.