Reducing Noise in Label Annotation: A Lane Change Prediction Case Study

Abstract A major challenge in automated driving is to understand what the perception system detects with its various sensors and map it to meaningful behavior. In situation interpretation one of the tasks is to capture patterns in the motion of surrounding vehicles to estimate their future behavior for safely master driving. One basic approach to estimate this future behavior is to predict whether a vehicle is going to perform a lane change or not. Different methods exist to predict lane change maneuvers. Here, the focus is on the domain of machine learning, such as neural networks or support vector machines. Since the systems are trained in a supervised fashion, ground truth information for the learning algorithm is needed. While manual labeling by human experts is time-consuming, expensive, and subjective, an automatic generation of the labels is desired. This paper investigates different methods of providing labels for the lane change prediction task automatically. Further, the focus is on optimizing the labeling strategy to prevent contradicting labels. Additionally, the interactions between different labeling methods and the imbalanced learning issue is discussed based on experimental results.