Deep Learning Based Classification of Pedestrian Vulnerability Trained on Synthetic Datasets

Abstract

The reliable detection of vulnerable road users and the assessment of the actual vulnerability is an important task for the collision warning algorithms of driver assistance systems. Current systems make assumptions about the road geometry which can lead to misclassification. We propose a deep learning-based approach to reliably detect pedestrians and classify their vulnerability based on the traffic area they are walking in. Since there are no pre-labeled datasets available for this task, we developed a method to train a network first on custom synthetic data and then use the network to augment a customer-provided training dataset for a neural network working on real world images. The evaluation shows that our network is able to accurately classify the vulnerability of pedestrians in complex real world scenarios without making assumptions on road geometry.