Towards Trustworthy Object Detection for Advanced Air Mobility using Deep Ensembles

In the rapidly evolving field of Advanced Air Mobility, ensuring the accuracy and reliability of object detection systems is paramount. Typical object detection approaches consist of black box deep-learning solutions, which downscale high-resolution images to meet the required model input sizes, eliminating crucial information in the process, especially for small objects. This thesis presents a novel approach to (i) increase the detection accuracy of small objects and (ii) enhance the trustworthiness of the overall detection in high-resolution aerial imagery. To accomplish these goals, the technical innovations of this research include (i) cropping patches of the original image to improve the detection of small objects and (ii) the creation of a deep ensemble consisting of multiple single-stage object detectors for inference on these patches. Additionally, a unified dataset from five high-resolution, open-source image datasets representing diverse Advanced Air Mobility scenarios has been accumulated. Experimental validation was conducted by comparing the presented approach to an established multi-stage baseline model. The experiments show a general detection increase (COCO AP metric) of 25.1% and a detection increase for small objects (COCO APs metric with bounding box area < 322 pixel) of 107.8%, compared with the baseline. These results not only validate the proposed methodology but also indicate significant advancements in both the accuracy and trustworthiness of the object detection system. It successfully demonstrates that deep ensembles, when applied to patches from high-resolution images, can substantially enhance object detection capabilities in Advanced Air Mobility applications.