UAV Detection Using Split-Parallel CNN For Surveillance Systems

초록

Commercial drones have become available to everyone with different sizes and shapes. Many are equipped with cameras and some with signal sabotage devices, the scariest scenario is that there are websites that offers weapons which can be attached to the drone. All those security threats either for privacy matters or people's safety, encouraged the researchers to find an intelligent system that can be implemented into the surveillance systems to classify unauthorized UAVs that are flying in a restricted area. This paper proposes a system that detects UAVs by acquiring RGB images via sensor then apply them to a convolutional neural network (CNN) that behave as an object classifier. Proposing Split-Parallel Cross Stage Partial DenseNet (PCSPDensenet) that is built from a modified CSPDenseNet. By splitting the feature map in two parts. Then, make each part flow in different side of the parallel network. The proposed network shows simulation results of an increment in the precision and showed higher AP(50) and AP(75) at higher frame rate on the UAV dataset With lower computational complexity.