Computer vision plays an essential role in the perceiving surrounding environment in the field of autonomous driving. One of the main focuses of this field is the reliable detection and classification of trac signs, to be able to abide by trac laws and provide a safe autonomous product. For this detection and classification, many machine learning based approaches exist. In this thesis, a self-supervised method for training a vision transformer, a recent deep learning architecture, called “self-distillation with no labels” is discussed and evaluated on the German Trac-sign Recognition Benchmark dataset. Moreover, the method is evaluated using a small dataset from a prototype vehicle at the Dahlem Center for Machine Learning. In total 3 models are evaluated. A model pretrained on ImageNet1K, a model further trained on the GTSRB dataset using the weights of the first model, and a model trained from-scratch exclusively on the GTSRB dataset. With these models a k-NN classification on the GTSRB dataset containing 43 classes is performed, producing precision and recall averages of 88.61% and 84.06% for the first model respectively. The second model output better precision and recall averages of 97.77% and 96.37%. The third model achieved comparatively worse with precision and recall averages of 77.91% and 72.46%.

The field of Autonomous Driving has witnessed rapid improvement and innovation
in recent years. This has been driven to a large degree by the major advances
in Machine Learning and Object Detection of the previous decade. Deep
Neural Networks enable the vehicle to find solutions to the myriad of complex
and interconnected tasks associated with Autonomous Driving. One such essential
task is the reliable detection and classification of traffic signs, which is also
significant for advanced driver assistance systems.
This thesis will focus on a new approach utilizing the current state-of-the-art
object detection algorithms trained on public datasets. The proposed system
combines YOLOv7 with a custom classifier trained on the German Traffic Signs
dataset. The results are evaluated on the GTSDB dataset and fisheye footage
recorded by the robotic laboratory’s prototype car. The final full system achieves
high results in most metrics, scoring 99.0 on Recall, Precision, and mAP on the
GTSDB test dataset. But it’s limited by the number of traffic sign types the classifier
can recognize.

Extremwetter stellt für das autonome Fahren eine unverkennbare Herausforderung dar. Das Erkennen des Wetters erlaubt es, das Fahrverhalten entsprechend der Wetterbedingung automatisch anzupassen. Im Rahmen dieser Arbeit wurden verschiedene Convolutional Neural Networks zur Klassifizierung des Wetters auf Verkehrsbildern unter unterschiedlichen Trainingsumständen auf dem umfangreichen BDD100K-Datensatz trainiert und evaluiert. Als wesentlich für die Übertragbarkeit der Vorhersagefähigkeiten auf die Bilder des autonomen Fahrzeuges der Freien Universität Berlin, stellte sich dabei der Einsatz von Label Smoothing heraus. Die Evaluation zeigt, dass das System besondere Schwierigkeiten bei Bildern mit leichtem Regen und einer geringen Spiegelung auf der Straße hat und zu Fehlklassifikationen bei Tunnel-Bildern neigt. Die Vorhersagefähigkeit des Systems wird durch den Einsatz eines exponentiell gleitenden Durchschnitts gesteigert und weitere Verbesserungsvorschläge für zukünftige Arbeiten wurden gemacht.

As more deep learning models are deployed for high-stakes use cases, explaining the predictions of a model is becoming more important. One class of methods for explainability are counterfactual examples. A counterfactual modifies a model input in such a way that the model output, for example a classification, changes in a target direction. In this work, we apply an efficient method for generating such counterfactuals (ECINN) to graph node classification. We introduce a synthetic graph dataset with ground-truth explanation labels.  Using this dataset, we quantitatively compare the model-specific ECINN method against a model-agnostic counterfactual generation method by Wachter et al.  on explanation size and correctness. We find that ECINN produces higher-quality counterfactuals and discuss the trade-offs between it and model-agnostic methods.