BLUEASSIST

The main activities and objectives of the collaborative project are:

• Enhanced exchange with potential users
• Conceptual design of an overall system consisting of UAV, tuned payload for data acquisition and transmission and attached ground station
• Development of a fuel cell-powered, hybrid-electric Antares E2
• Concept for the integration of a multi-sensor system, a data transmission unit and a ground control station

In this project, Fraunhofer IOSB and its SZA department are participating in:

• Investigation of methods for automated data (pre)evaluation, filtering and compression onboard with regard to efficient data transmission
• Development of application-specific evaluation methods and design of the necessary hardware components for the ground station.
   

Involvement of the Scene Analysis Department SZA:

Multisensory data processing and analysis (onboard + ground station)

Widespread aerial surveillance camera systems are offered by companies such as Trakka Systems with the TC-300 model. The swivelling and stabilised design with integrated thermal and HDTV camera, with continuous zoom, ensures sufficient flexibility. To test the performance of this camera system in the field of aerial ship detection, artificial intelligence (Mask R-CNN networks pre-trained with MS COCO) was applied to a video section of the camera at Fraunhofer IOSB.
Video Source: Trakka, modified by Fraunhofer IOSB

Mask R-CNN are state-of-the-art instance segmentation methods in which incoming images are processed by different networks (FPN, RPN), and features are segmented and classified. Due to its modularity and multi-phase structure, Mask R-CNN is comparatively easy to train on new scenarios.

A comparable scenario is the automatic detection of persons. For this, the AI framework already described was not trained on ships, but on people. Changing the perspective (now terrestrial) does not pose a problem.

Similarly, AI in the form of Convolutional Neural Networks (CNN) is used in the step-by-step detection of vehicles from the air. Here, the network consists of two branches: the classification branch, due to its four pooling layers, leads on the one hand to a high classification performance and on the other hand to a strong reduction of the spatial resolution. The segmentation branch contains only one pooling layer and compensates for the reduction caused by the classification branch with its correspondingly high-precision resolution. Despite the use of graphics card programming, this algorithm has to wait for every fifth image to keep up due to the enormous computational effort. Due to the large overlap of the individual image areas, this method can nevertheless be described as real-time. Video Source: SELSAS

This project is funded by the research initiative mFUND (Modernity Fund) of the Federal Ministry of Transport and Digital Infrastructure BMVI and its call for ideas and funding "UAS and air taxis".