Intelligence, RecognItion and Reactive SurveillancE – IRISER

The members of the research federation Math-STIC (FR CNRS 3734), in particular the members of the laboratory of Information Processing and Transport (L2TI, UR 3043) and of the Computer Science Laboratory of Paris Nord (LIPN, UMR 7539 CNRS), of the University Sorbonne Paris Nord (USPN) and the company COSE wish to structure their currently informal collaboration through a common laboratory dealing with intelligence, pattern recognition and surveillance. The IRISER LabCom aims to propose, master from end to end the behavior and performance of intelligent or even embedded systems designed for computer vision for the rapid and automated analysis of images/videos (very large, georeferenced multispectral of high resolution) captured on board COSE's aircraft based on visual information processing and machine learning strategies. The work planned is articulated around five main axes, namely

Axis 1: Object detection on very large images. The aim will be to improve the research tracks already explored and to develop essential tools for research but also for the deployment and integration of the solutions developed within the CAMELEON system of COSE. Knowledge on the deployment of similar solutions is essential for the transition from the prototype state to a real application. In parallel, a monitoring activity will allow to stay abreast of new theoretical and technological advances in the field of object detection.

Axis 2: Object classification. The existing solutions today already reach high performances. The work will focus on the optimal use of CNN architectures to allow learning with relatively few samples. This is an important point, because for some uses, the generation of a database for a specific class can be difficult and require a too long time for a user wishing to integrate a new class.

Axis 3: Dynamic object tracking and trajectory prediction in (real) satellite image streams. In addition, it will be necessary to develop tracking evaluation metrics to be able to compare the performances of different implementations.

Axis 4: Learning and autonomous recognition. The aim is to propose a modular recognition model that can be interpreted and adjusted by analysts with knowledge of the field. Nevertheless, we would like them to operate only on interpretable parameters. For this reason, we wish to automate the selection of variables and hyperparameters. The preferred approach is Bayesian optimization, as it allows the exploration of continuous spaces, in contrast to grid search, and tolerates noise in the objective function evaluations. Nevertheless, current implementations work well when the dimension of the parameter space is less than 20. We therefore wish to scale up this type of algorithm in order to apply it to larger tasks.

Axis 5: The optimization of the proposed algorithms in order to be executed in real time or slightly delayed on embedded systems by taking advantage of specific available computational components and the parallelization that chips such as Nvidia's Tegra Xavier can offer.