• object recognition
• UAV
• image processing
• situation awareness

The aim of this work is to design, develop and validate methodologies for object detection and recognition using a single colour camera that can be applied to Unmanned Aerial Vehicles (UAV) in airborne autonomous controlled missions.
State-of-the-art techniques for object extraction and description are analysed, discussing for each their qualities and limitations.
Two different methods for object recognition have been designed and validated in virtual scenarios. A first method, based on object shapes, has shown faster evaluation time but lower precision and accuracy than the second one.
The second method, based on object details, has shown complementary performances. Hence, this complementarity has suggested to combine them in a third methodology, composed of the cascade of both methods, which has obtained better accuracy, precision and specificity than the original systems composing the cascade.
An UAV attention system is designed to automatically extract possible objects from the scene, and a detection method is experimented which allows real-time performances. The detector designed demonstrated to be acceptable in virtual environments, but it needs further developments and integrations for a practical application.
The UAV object recognition system is designed with a look on human situation awareness in order to completely automate the UAV comprehension of scenes, limiting to the minimum possible the human supervision and to aid UAV decision systems to take autonomously correct actions for improved mission success probabilities.
The system performances have been compared systematically to commonly adopted approaches based on the percentage of matching features from a database image to an input image, and it has been tested in simulated and real scenarios. In virtual environments, a database of four objects, three vehicles and an airplane, has been loaded and matched against images containing objects in various and mimetic environments, in different light conditions and even partially occluded. The proposed recognition system has recognised almost a thousand of object instances and rejected almost a thousand of negative images with better precision, specificity and accuracy than other adopted approaches. While real-time object recognition performances of the entire system were successful in virtual tests, in real complex scenarios performances could degrade between one or two seconds per frame and to lower accuracy, because of the lack of both effective and fast object extraction and shape estimation techniques. Future works will identify possible improvements for the proposed methodologies, test more objects, experiment new complementary acquisition hardware in order to obtain similar extracting performances in real scenarios and add an higher level decision system which automatically defines the best action to be executed relatively to the analysed situation. The conclusions resulting from this work are that shapes, which are commonly not considered when matching features, determine advantages in terms of performances and accuracy in object recognition problems, and that unmanned aerial systems with situation awareness capabilities improve the mission success and remove the current necessity of human supervision or result in a single human simultaneously supervising a larger number of vehicles.