An Ego-Lane Estimation Method With Sensor Failure Modeling

Abstract

Accurate vehicle localization at the level of individual lanes is crucial to ensure the safe and efficient operation of autonomous vehicles, serving as a cornerstone for the development of future Advanced Driving Assistance Systems (ADAS). Contemporary localization methods relying on Global Navigation Satellite Systems (GNSS) often fall short of achieving the necessary precision, necessitating the involvement of additional systems. These supplementary systems frequently depend on the output of road line detectors, whose performance can be hindered by various factors, including adverse weather conditions and heavy traffic, resulting in noisy or sporadically missing data. This study introduces a probabilistic algorithm designed to precisely estimate the actual lane positioning of a vehicle in the specific context of multi-lane roads, such as highways, without relying on GNSS data. The proposed algorithm is built upon a Hidden Markov Model that exploits the output of a generic line detector, a common component of contemporary driving assistance systems. This model ensures consistent lane localization estimates even when faced with noisy or intermittently missing data. Experiments demonstrate the algorithm’s effectiveness, providing a reliable estimate of the vehicle in-lane position in challenging datasets containing highway scenarios with hundreds of lane changes. This contributes to the enhancement of existing literature, achieving an accuracy of 86.71% over a segment exceeding 50 km. These results, improving by almost 10% over our previous efforts, suggest that our approach has the potential to enable new ADAS functionalities and offer a robust localization scheme for use in the context of autonomous driving scene understanding.

Keywords

• Ego-lane-detection
• multi-lane roads
• lane-localization
• vehicle-localization
• selflocalization
• highway-like scenarios