EDA has accomplished a two-years study (2020-2021), called ARTINDET, into how Artificial intelligence (AI) applications can be used to improve the automatic detection, recognition, identification and tracking of small, fast-moving targets in a complex battlefield environment. In such a hostile and difficult context, high-performance electro-optical (EO) imaging systems together with high spatial resolution radars seem to be best solution to efficiently detect and mitigate these new threats. One of the main conclusions of the study, is that AI significantly enhances the performance of those two technologies related to image pre-processing, fusion and inference. Another finding of the study points to the future: additional research efforts should be put into these promising technologies which can make a difference for defence capabilities. Hence EDA’s proposal to launch a dedicated project in 2022.
The study developed and analysed new image processing techniques of imaging systems relying on AI based on deep learning paradigm. For that purpose, images captured by high-resolution cameras and Synthetic Aperture Radar (SAR) operating with Wide Field of View (WFOV) fed the different algorithms tested. The different techniques implemented and tested are designed to be integrated in a dedicated HW/SW architecture for an Unmanned Aerial Vehicle (UAV). ARTINDET analysed all the required steps for the deployment of this kind of system.
Urban area & open sea
Concretely, the study used two scenarios for object identification: one in an urban area and one focused on ship detection/recognition at open sea. For each of the two scenarios, two data sets were created, composed of both EO and radar images, and two AI algorithms were developed: one for the segmentation in urban scenario (mainly critical building identification) and one for the detection/segmentation of ships. New AI-based image fusion and resource management techniques were also developed.
The study revealed that the usage of the new AI-based algorithms leads to a considerable improvement of the identification and detection performances, also due to the automatic and ‘intelligent’ choice of the images supported by machine learning and neural networks.
Additional work to be done
The study also highlights the necessity of further work on topics such as:
- The creation of an EU military image data base for AI-based system training and testing;
- Further algorithm developments;
- Standardisation, certification and validation of AI algorithms;
- Hardware architecture implementation;
- Extended measurement campaigns for AI processing performance analysis and validation.
EDA project in the starting blocks
That’s why the Agency has proposed to Member States to launch a dedicated EDA Cat B project, called AIDRIT (Artificial Intelligence for Automatic Detection Recognition, Identification and Tracking of Difficult Target) which, if accepted, could start in 2022. Along others, it would look into the afore-mentioned additional work highlighted in the ARTINDET study. Both the study and the potential AIDRIT project idea are perfectly aligned with EDA’s action plan on AI. There is also realistic possibility that this topic could be the subject of one of the next calls for proposal under the European Defence Fund (EDF).
Small targets coming from different directions and new intelligent and sophisticated weapons operating in complex scenarios represent nowadays the new asymmetric threats in the battlefield. In this hostile and difficult context, the new high-performance electro-optical (EO) imaging system allows to efficiently detect and contrast these new threats. High resolution multidimensional (multiband/multispectral, hyperspectral, multiresolution) EO sensors are designed to have enough diversity for improving detection, recognition, identification and tracking (DRIT) of difficult targets. The main challenge stems from the huge quantity of data produced. On the one hand, this is good because it gives a certain level of completeness in the information; on the other hand, it represents a raising of computational costs and an increase of the image processing complexity. Low contrast distributed targets make their identification and visualization by the operators still really demanding and sometime impossible. Against this backdrop, automatic processing would be helpful with a drastic reduction of reaction time for decision, often crucial in military missions.
New techniques based on AI making use of deep learning and/or machine learning, seem particularly useful for image processing of high-resolution camera, when operating with wide field of view (WFOV) for the detection of difficult targets (low contrast, low signature, small size and operating in degraded visual environment). Moreover, applications on camera mounted on unmanned air vehicle (UAV), where full images cannot be streamed to the ground station, or when multiresolution imagery is needed, are new challenges where AI can also help. AI can have applications and provide benefits on data fusion coming from homogenous or heterogenous sensors, particularly from imaging radar and cameras, for a better situational awareness picture, and on the implementation of aid decision making tools and missions’ planning. AI can also be efficiently exploited in modelling and simulation for data generation and user training.