Gathering to train different machine learning models was not as easy as today. Analyzing and labeling these data is time- and human-consuming, particularly if data from several sensors are gathered. Furthermore, the labeling process gets more complex if the data representation is complex. In this Master thesis, an Automated Data Labeling Pipeline is proposed, implemented, and evaluated. The pipeline takes driving recordings of the Dahlem Center for Machine Learning and Robotics as input. A ROS bag reader extracts the most relevant data. Two stages process the stored camera images (Image Processing Stage) and the LiDAR point cloud data (LiDAR Processing Stage. The last stage consists of a Sensor Fusion, where the image and LiDAR data processing results are fused to form a sensor fusion dataset. Each pipeline stage is evaluated on its own. Furthermore, several publicly available datasets and a newly created one are used to train different object detection models.
By completing this Master thesis, an Automated Data Labeling Pipeline to support a human labeler is realized.