Résumé:
Owing to their rapid response capabilities and maneuverability, extended range and improved personnel safety, drones equipped with vision systems have great potential for forest fire monitoring and detection. Over the past decade, the demand for drone-based forest fire detection systems has increased, as forest fires have become
a growing concern with climate change, as indicated by several environmental studies, and a reality in some parts of the world. Despite this, existing drone- based forest fire detection systems still present many practical problems for their use in operational conditions. In particular, successful detection of forest fires remains difficult, given the very complicated and unstructured forest environments, the movement of UAV-mounted cameras . These negative effects can seriously cause false alarms or faulty detection. In order to perform this mission, meet the corresponding
performance criteria and overcome these increasing challenges, it is essential to investigate ways to increase the probability of successful detection and improve the adaptation capabilities to various circumstances in order to improve the accuracy of the forest fire detection system. Based on the above requirements, this master
thesis focuses on the development of reliable and accurate forest fire detection algorithms applicable to drones. These algorithms provide a number of contributions, which include: (1) a learning-based forest fire detection approach is developed by considering the color characteristic of the fire; (2) a forest fire localization scheme
is designed by combining both stereo vision and perspective projection; and (3) a design and control of a quadcopter using the PIXHAWK autopilot.
