Onboard Smart Surveillance For Micro-Uav Swarm: An Experimental Study

In this paper, a smart approach is proposed and developed to enable micro-UAV surveillance with extremely limited onboard computation resources. Recently, a few algorithms have been proposed to detect ground vehicles from UAV aerial vision views. But most of them are processed via air-ground communication and abundant ground computing platforms. However, this study concentrates on the onboard processing scheme of detection and classification on ground moving vehicles by one flying micro UAV. A unified scheme of saliency detection and shallow convolution neural network classification makes a compatible surveillance performance with the only onboard processor. Under such circumstances, an experimental study is conducted on the developed micro-UAV onboard surveillance approach. The experimental results show that the proposed approach can definitely detect and classify at less six kinds of ground moving vehicles up to similar to 14 fps with only limited onboard computation. This work demonstrates practical feasibility for online vision processing of micro-UAV swarm surveillance on ground objects.