Our Client:

A defence technology start up.

Situation:

The ability to survey large areas is crucial in many applications, such as disaster response, border patrol, or protecting endangered animals from poachers. Conducting such surveys is challenging and resource-intensive, especially in difficult to access areas or complex terrain.

The objective was to explore the possibilities of using state-of-the-art AI techniques to create an efficient, real-time system for large-area surveillance that can help mitigate the risks associated with complex terrain and limited resources.

Traditional approaches to area surveillance are often inefficient, as they rely on manually controlled, individual drones that can only cover small areas and generate vast amounts of data. In contrast, our solution leverages machine vision, enabling drones to autonomously identify and track specific objects (such as people, animals, or vehicles) of interest in real-time. This technology enables the coordination of multiple drones to explore an area and collect data efficiently, maximising effectiveness and impact with limited resources.

Task:

DSL developed an integrated system that combined state-of-the-art machine vision techniques, multi-agent reinforcement learning and cooperative exploration algorithms. The drone swarm’’ was equipped with cameras and onboard processing hardware to perform object detection and tracking using a state-of-the-art Transformers-based architecture. This enabled the drones to identify and track various objects such as cars,people, and animals, even in challenging environments.

To maximise information gathering during the limited flight range of each drone, DSL developed cooperative exploration algorithms that enabled the swarm to optimally explore the area. The algorithm considered both the information gained from exploring new areas and the cost of exploration, taking into account the remaining battery life of the drones.

The cooperative behaviour between the drones was learned using multi-agent reinforcement learning, which allowed the swarm to efficiently accomplish a broad range of objectives in complex and dynamic environments.The different components of the system were integrated using the Robot Operating System (ROS), a widely used framework for developing robotic applications.

Result:

The system demonstrated a significant improvement over the human baseline in terms of the area surveyed and the amount of information gathered in the same amount of time with the same resources.

The tracking system was able to achieve high accuracy in identifying and tracking objects in real-time, enabling the system to provide valuable intelligence and situational awareness. The use of multi-agent reinforcement learning enabled the drones to work cooperatively and optimally explore the area to maximize information gain while ensuring efficient use of limited resources. The integration of different components using ROS provided a robust and scalable solution that can be applied in a wide range of use cases, from defence, agriculture and emergency response.