Self-localization for mobile robots in dynamic environments is a challenging task, especially when relying solely on odometry and 2D LIDAR data. When operating in fleets, mutual detection and the exchange of position information can support each robots localization. Detecting and classifying different robot types in a heterogeneous fleet, however, is non-trivial if only 2D LIDAR data is used. Object shapes detected with this sensor in the real-world can vary heavily due to irregularities of the environment. In this thesis a novel approach for mutual robot detection and relative position estimation based on a combination of convolutional and ConvLSTM layers is therefore presented in order to solve this issue. The algorithm learns an end-to-end classification of robot shapes using 2D LIDAR information transformed into a grid-map. Subsequently multiple hypotheses for each robot type are extracted out of the heatmap output of the network using a hierarchical clustering algorithm combined with a centroid calculation for each hypothesis. These position hypotheses can then be used in an overall multi-robot localization in order to increase the localization of the robots. Due to the similarities that many robot shapes in 2D LIDAR data share with static objects in the environment (rectangles, circles), an additional pre-processing of the data is performed. Three different end-to-end approaches for semi- and fully-dynamic object detection are introduced. The first one is using stacked laser-scans with a Convolutional Network in order to detect spatio-temporal features of moving objects. The second one transforms the sensor data into a 2D grid-map and accumulates multiple consecutive maps to create a map with spatio-temporal features for moving objects. This map is then used as an input for a ConvLSTM network. Both approaches only detect semi-dynamic objects, since the spatio-temporal features in both cases are only visible for robots moving at the currently viewed time-span. Another simple approach is presented for extracting fully-dynamic objects by calculating the difference between the static grid-map provided by the robots Localization & Mapping algorithm and the grid-map calculated from the laser-scan. We conduct an in-depth evaluation in order to determine the networks’ ability to generalize to different environments and cope with uncertainty in the input data. Furthermore the benefit of using simulation data for pre-training real-world models is evaluated together with the improvement of pre-processing the data by the dynamic object detection networks. The results of our evaluation show, that the classification network is able to achieve a precision of 94% on real-world data with a position estimation error of 13cm. Pre-processing on the other hand does not improve the networks performance overall, although the two 2D approaches for dynamic object detection are able to identify most of the moving robots correctly. Using 1D scan data on the other hand does not lead to any promising results for dynamic object detection. In order to determine the benefit of the mutual robot detection and position estimation, the system will be integrated into a multi-robot localization subsequent to this thesis.