Description
Date depot: 27 décembre 2021
Titre: Hybrid AI for the design of intelligent radars -Joint recognition and tracking of drones in complex environments
Directrice de thèse:
Amal EL FALLAH SEGHROUCHNI (LIP6)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: In the quest to build autonomous intelligent systems able to form hypotheses, make decisions, learn from data, and adapt their behavior to changes in their environment, several large-scale organizations have begun to align around a concept known as the “Cognitive Era” [1].
To design autonomous intelligent radars, capable of intelligent surveillance capability in complex environments as those of Thales missions, several cognitive functions are necessary, such as the perception and the representation of the environment (including other radars and threats), reasoning about knowledge and ongoing actions as well as learning from past experiences. Hybrid AI is then necessary to combine all these functions; in addition, their collaboration within a radar to achieve the sought intelligent surveillance.
The goals addressed in this thesis are the following. First, we are interested in proposing an architecture based on agent paradigm which is able to embed the necessary cognitive functions in hybrid manner; then, to define an orchestration mechanism in order the take benefit from these functions in a cooperative way.
The agent paradigm is particularly recommended as a suitable receptacle for embedded hybrid AI. Indeed, an agent architecture is modular [2] and it is possible to define an interpreter able to build a life cycle and to orchestrate the agent basic building functions. At the agent level, these various functions can collaborate and produce a coherent behavior. When more radars are involved, e.g in multi-static scenarios, a multiagent system paradigm [9] offers many tools for cooperation, coordination and collaborative problem solving [12, 2].
Case study: Cognitive Radars
Safety, security and defence operations increasing rely on the use of friendly drones (or UAVs) whilst facing the threat of hostile drones. Of all sensors, radar sensors provide the most assurance in sensing drones, particularly non-cooperative drones such as those that are operated autonomously.
Drones have low radar cross section, fly at low speeds in dense clutter environments and they can be easily confused with natural (for example birds) and manmade targets. This represents a significant problem for their effective detection, tracking and recognition.
Effective tracking aids robust recognition, yet in dense target environments, recognition features are important for effective tracking.
In this study we propose to rely on agent paradigm and to develop AI techniques to jointly solve the problem of recognition and tracking of UAVs and drones, using combinations of AI recognition, AI tracking and AI supervisory agents. The scope of these techniques will be further broadened to encompass measurements from networks of radars, to manage handover between coverage zones and include additional information gathered from radars with different aspects of view or fundamental properties.
This work will be supported by real radar measurements and further measurements could be collected using Aveillant Gamekeeper radars.
AI background:
Suitable AI techniques could be integrated into agents: MCTS (Monte-Carlo Tree Search) and Reinforcement Learning (RL) in continuous spaces, e.g. with unknown Markov Decision process, for Multiple hypotheses tracking & recognition where the agent (i.e. tracker) uses a computationally tractable algorithm to perform the two tasks of estimating the system state (e.g. target kinematics state) which can incorporate the target identity (i.e. recognition) with the inherent present uncertainties and the planning of its action (i.e. in this context this is to optimize its strategies for tuning their parameters with respect to target kinematic and identification) within a Bayesian framework. This can be drawn on the prior work in [1]. Addressing the adaptation of the recognition algorithm (e.g. kernels-based) parameters within this formulation (e.g. based on distribution shifts in a simple of case) can be highly beneficial.
Doctorant.e: Alhadhrami Esra