@thesis{Wollschlager2020, author = {Wollschlager, Marinus}, title = {Validation of a CNN classifier for RADAR in real and simulation domain, exclusively trained in simulated RSI model}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-8727}, pages = {41, XVII}, school = {Technische Hochschule Ingolstadt}, year = {2020}, abstract = {The current developments in driver assistance systems enabled OEMs to offer vehicles, ready for highly automatic driving (Level 3). A key factor for this development is the environmental perception, based on sensor-systems. Currently, LIght Detection And Ranging (LIDAR), Cameras and RAnge Detection And Ranging (RADAR) are the mainly used sensor-systems, to generate a perception of the vehicular environment. Due to different used spectrums, the performance and reliability of a sensor depends on environmental conditions, like darkness, rain or fog. RADAR is known as a robust sensor, working well also in rain or light conditions. A key feature for an object, surrounding the own car, is a classification, independently of a certain sensor-system. It enables the adjustment of the own trajectory to specific situations, defined by the surrounded objects. Nevertheless, it is challenging to classify an object with RADAR. Common sophisticated automotive classification algorithms rely on camera. But a precise classification of objects, detected by RADAR, gained on importance due to the reliable characteristics of RADAR mentioned before. In addition to real world domain, simulation software provides tools for RADAR measurements and deployment. Even physical sensor models, near to reality, are implemented in software, enabling the improvement of signal processing techniques in a reproducible, convenient and fast manner. RADAR-specific effects, like multipath or clutter, are also implemented. The effects cause so-called "Ghost-Targets", pretending fake objects. This thesis will make a contribution to the question how the simulation of RADAR can be used to classify objects in real world. The classification task is taken over by a CNN classifier. Due to the fact that it is a supervised learning method, the classifier gets trained. Current published papers used real but quit static data for the training. This thesis uses data out of IPG Carmaker 8 simulation model. The proposed methods cluster the data with a state of the art cluster algorithm for RADAR and label, based on ground truth, with a novel approach. A multi-class CNN, implemented in Tensorflow, is enabled to differ between a car, bicycle or pedestrian. The evaluation is concentrated on a confusion matrix. Beside simulation data, also real world data is used to evaluate the classifier. Therefore, a scenario in real world is created. The measurement is performed with INRAS RadarLog, specialized for research and raw data analyze. Both domains, simulation and real world, are scaled on a common frame size and amplitude. The comparison should indicate the potential of a classifier to classify objects with a certain accuracy. The appropriate classifier is exclusively trained in simulation domain and gets applied in the real world and simulation domain. Weather phenomenons like rain, fog or extreme temperatures are excluded in the simulation.}, language = {de} }