Deep Neural Network based Secured Control of Flying Vehicle in Urban Environment

Aerial transport for daily inter- and intra-urban passenger mobility has been a human dream for decades ago. Now, the dream is near to be realized after the great technological advances achieved recently. As urban populations' size increases, traffic congestion and air pollution remain major threats to economic growth. However, questions concerning safe and secure autonomous flying vehicles remain to be answered, especially in a highly dynamic environment. This research paper presents a Deep Neural Network (DNN) algorithm as an additional component that avoids obstacles and increases the tracking efficiency of Cascaded Proportional Integral Derivative with Feed Forward (PID+FF) controller for flying vehicles. In the current paper, we introduce KaNET: a convolution neural network designed to safely drive and improve tracking performance. Simultaneously, cascaded PID+FF algorithms are applied to regulate the atti-tude/altitude of a flying vehicle. To ensure a safe and secure flight in a dynamic environment, the proposed KaNET generates two outputs for each input image. The first output is the probability of a collision allowing the flying vehicle to recognize and react quickly to hazardous situations. The second output represents a steering angle that helps the vehicle to keep flying and avoiding obstacles. The cascaded PID+ FF is used to control and maintain the desired position and orientation of the flying vehicle. Results have shown that the proposed approach improves tracking accuracy and avoids obstacles in highly dynamic environments.