Smart environment monitoring through micro unmanned aerial vehicles

In recent years, the improvements of small-scale Unmanned Aerial Vehicles (UAVs) in terms of flight time, automatic control, and remote transmission are promoting the development of a wide range of practical applications. In aerial video surveillance, the monitoring of broad areas still has many challenges due to the achievement of different tasks in real-time, including mosaicking, change detection, and object detection. In this thesis work, a small-scale UAV based vision system to maintain regular surveillance over target areas is proposed. The system works in two modes. The first mode allows to monitor an area of interest by performing several flights. During the first flight, it creates an incremental geo-referenced mosaic of an area of interest and classifies all the known elements (e.g., persons) found on the ground by an improved Faster R-CNN architecture previously trained. In subsequent reconnaissance flights, the system searches for any changes (e.g., disappearance of persons) that may occur in the mosaic by a histogram equalization and RGB-Local Binary Pattern (RGB-LBP) based algorithm. If present, the mosaic is updated. The second mode, allows to perform a real-time classification by using, again, our improved Faster R-CNN model, useful for time-critical operations. Thanks to different design features, the system works in real-time and performs mosaicking and change detection tasks at low-altitude, thus allowing the classification even of small objects. The proposed system was tested by using the whole set of challenging video sequences contained in the UAV Mosaicking and Change Detection (UMCD) dataset and other public datasets. The evaluation of the system by well-known performance metrics has shown remarkable results in terms of mosaic creation and updating, as well as in terms of change detection and object detection.