In the last years, the technology of small-scale Unmanned Aerial Vehicles (UAVs) has steadily improved in terms of flight time, automatic control, and image acquisition. This has led to the development of several applications for low-altitude tasks, such as vehicle tracking, person identification, and object recognition. UAVs are usually equipped with heterogeneous sensors such as thermal cameras and Red Green Blue (RGB) cameras. The latter are the most used for three reasons. The first is that the acquired images are easy to interpret by human eye. The second is their cost, which is more affordable with respect to other sensors. The third is the availability, since also the commercial UAVs mount an RGB sensor. Moreover, recent developments in fields such as Deep Learning (DL) have improved the autonomous capabilities and the precision of the UAVs in tasks such as guidance. Despite this, there are no algorithms optimized for UAVs for object detection task. This project is focused on the latter. In detail, the aim is to propose a modified deep neural network optimized for the detection of people in aerial images acquired at very low altitudes by small-scales UAVs.
UAVs are now used in several civilian and military applications such as Search and Rescue, border controls, and environment monitoring. The latter includes processes such as the detection and classification of objects, vehicles, and people in restricted areas such as airports and military bases. To be able to perform such task, machine learning and deep learning must be used. By exploiting a simple RGB sensor mounted on commercial small-scales UAVs, and a deep learning model, it is possible to detect objects in aerial images. In the current literature, the models are trained on well-known datasets which do not include aerial images. In case of works that use deep learning with aerial images, they just train the models with the needed images without proposing a new or a modified model. Moreover, the actual deep learning models do not consider the perspective of the objects in the aerial images. The main contributions of the proposed project are the following:
1. The study and the development of a new deep learning model, in order to optimize the detection of people in aerial images acquired at very low altitude by small-scales UAVs.
2. Benchmark of the proposed model with respect to the existing models. The training and the testing will be performed on a dataset containing aerial images of people, vehicles, and object. The dataset has created during my PhD experience and has been published on Transaction on System, Man, and Cybernetics: Systems.