Buch, Englisch, 212 Seiten, Paperback, Format (B × H): 148 mm x 210 mm, Gewicht: 326 g
Reihe: Informatik
Buch, Englisch, 212 Seiten, Paperback, Format (B × H): 148 mm x 210 mm, Gewicht: 326 g
Reihe: Informatik
ISBN: 978-3-8439-5552-2
Verlag: Dr. Hut
In the field of commercial vehicles, there is a strong need for automation. Their tasks are complex and challenging, and the environments are unstructured, so humans must be present. Therefore, guarantees for the safety of persons in the vicinity are needed. Advancements in machine learning, mainly object detection with Convolutional Neural Networks (CNNs), make this realistic. However, it is unclear when a network fails to detect a person. So, current interpretation approaches can not give guarantees.
The work at hand makes a step towards this by dividing commercial vehicle applications into different categories, first dependent on the sensor setup. Two archetypical application scenarios are taken into account. Other application scenarios can be approximated by combination. Then for each category common disturbances to the images are identified and further categorized. As it is nearly impossible to collect enough data to accommodate all possible disturbances in each of these categories under all circumstances, the creation of simulated data is introduced and the suitability is assessed.
In this work, it is assumed, that an object detection network is provided and the suitability in safety-critical situations should be analyzed. Therefore, different methodologies are used to assess the level of invariance to each of the identified disturbances, such as changing weather conditions. The methodologies leverage the representation of the filters in a CNN as wavelets and utilize the correlation between the robustness of features and their Intrinsic Dimensionality (ID). The assessment of the methods is done on simulated data and on real data from application scenarios.
In the end, a workflow to reduce the influence of common disturbances is proposed and evaluated. In this way, an object detection network can be analyzed with respect to the suitability of being used as a safety-critical part of an automated commercial vehicle and can be improved if necessary.
