Improving Ego-Lane Detection by Incorporating Source Reliability

This paper presents a framework for robust lane detection towards automated driving using multiple sensors. Since every single source (e.g., camera, digital map, etc.) can fail in certain situations, several independent sources need to be combined. Moreover, the reliability of each source strongly depends on environmental conditions, e.g., existence or visibility of lane markings. Thus, we introduce a concept of estimating and incorporating reliability into the fusion. First, a new sensor-independent error metric is applied to assess the quality of the estimated ego-lanes based on the angle deviation. Secondly, we deploy a boosting algorithm to select the highly discriminant features among the extracted information. Based on the selected features, we apply different classifiers to learn the reliabilities of the sources. Thirdly, we use Dempster-Shafer evidence theory to stabilize the estimated reliabilities over time. Using a big collection of real data recordings from different situations, the experimental results support our concept.