Multi-source Fusion Using Neural Networks and Genetic Algorithms Towards Ego-Lane Estimation

An important task of automated driving is to keep the vehicle in the host lane. For a robust lane estimation, the information of multiple data sources needs to be combined to overcome the drawbacks of each individual sensor []. For that reason, this paper proposes a new mediated perception approach towards lane keeping function. Additionally to artificial neural networks (ANNs) used by Nguyen et al. [], genetic programming (GP) is applied to estimate the parameters of an approximation of a clothoid, which is used to represent the ego-lane. Therefor, a set of lane marking detections, information about the ego-vehicle and the leading vehicle and information about the current location are used as input for the estimators. Compared to the reference, the resulting ANN and GP estimators mostly achieve an angle deviation of smaller than 2° at a distance of 30 m. Hereby, both approaches achieve an overall availability of around 90%. Surprisingly, GP surpasses all compared approaches based on lane markings or a deep learning approach, which directly estimates the ego-lane from camera images, with an overall availability of 0.91. For that reason, ANNs and GPs can be applied to solve the lane keeping task so that future research can lead to further increases of the availability.