Object Detection And Human-computer Interaction Based On UAV Platform

With the development of UAV control technology and artificial intelligence,multi-rotor unmanned aerial vehicles have been widely used in many traditional applications such as aerial photography,plant protection and investigation.Nowadays,they are also used to some research areas,like tracking and identifying targets.At present,most of the micro UAVs are applied to civilian entertainment,and the current man-machine interaction of UAV is still mainly depend on remote controller.UAV's development and popularization is limited by the size of controller hinders,so it is necessary to develop a kind of new interaction method that does not rely on remote controllers or other assistant devices.The human body posture information is obvious and friendly to humans.Man-machine interaction could be realized by identifying human body posture information with the application of the computer vision and artificial intelligence technology.Firstly,the mechanical vibration of UAV may cause jitter in videos,and it will impair image processing.The vibration is depressed by using central area template matching method between two approaching frames.Then,human targets in videos are supposed to be detected.Detection will be achieved on the airborne embedded system and ground station respectively.The ground station has abundant computing resources,which is able to support the deep convolution neural network detector Faster R-CNN.It can adapt complex environment with high accuracy,but its large amount of calculation makes it hard to work on embedded systems.To solve this problem,foreground information is used for detection,and a simple CNN is designed to identify whether the target is human.Next,fully-connected neural network is used to identify human body posture information.Users are asked to wave their hands before giving orders to UAV and the control system will check it before detecting human body posture.Four kinds of body postures are designed for man-machine interaction and different postures mean different orders.To simplify the detected posture message,the foreground information of the person is used as features,and the fully-connected neural network is applied for recognition.Finally,those algorithms are transplanted into the airborne embedded systems and flight experiments are taken to validate our system's performance.The results show that our UAV man-machine interaction system can archive high recognition accuracy towards human targets and different body postures with real time processing.