Global Vehicle Localization by Sequence Analysis Using LiDAR Features Derived by an Autoencoder

Global vehicle localization is normally done by GNSS sensors. In case of GNSS outages, such as in urban canyons or tunnels, highly automated cars cannot localize without an initial known position. In this paper we propose a method for absolute localization in a city based on only one 2D laser scanner.In out method, localization is done by matching vertical scan lines captured by a 2D laser scanner mounted on the vehicle with scan lines derived from a reference point cloud of the environment. We use a neural network to derive significant features describing the shape of the scan lines. Every scan line of a reference data set is labeled with a specific cluster-id using a k-means algorithm and stored in a reference graph. The same k-means algorithm is used to label the single scan lines of a test drive. The localization is done via a sequence mining approach, where a sequence with a specific length is matched to the position with the highest correlation in the reference sequence.In our experiments we analyze the effect of several parameters, including the number of features and sequence length. The results show that the algorithm performs with an accuracy of about 1.4 m and a completeness of up to 99%. Even if the input scan is represented by only ten features, the results are betten than those obtained by using the whole range scan in the localization step.