Research Of UAV Object Detection And Localization Technology For Indoor Scenarios

During modern rescue and search tasks,unmanned aerial vehicles(UAV)can quickly fly into buildings in a formation to search and locate designated objects with the carried equipment,which can provide great assurance for the downstream material delivery and personnel rescue.However,due to the occlusion and complex indoor environment,positions of object and UAV cannot be obtained by conventional wireless positioning means such as satellites and base stations,UAV needs to have both autonomous positioning and relative positioning capabilities to the target.In addition,indoor rescue tasks usually require the cooperation of multiple machines:After confirming the mission target,the host needs to guide the slave machine moving forward to the target and carrying out rescue operations.But the conventional wireless methods cannot perform simple and fast guidance for the slave machine.In order to solve the above problems,this paper selects the UAV target detection and positioning technology for indoor scenes as the research.Taking the binocular vision sensor as the core,this paper is aim to systematically solve the problem of autonomous positioning,relative positioning of the mission target and fast guidance between master and slave UAVs.This paper mainly involves the following work:First of all,aiming at localizing the indoor object,especially the object personnel under the condition of low computational,this paper designs a fast target relative localization algorithm combining target detection and binocular vision.Firstly,the image semantic region of the observation object is deduced by using target detection network.Secondly,the object is presented based on the detection area and the edge features,and the fast feature search and registration are implemented on both sides of the center based on the binocular epipolar constraints.Finally,the registration points are triangulated and post-processing optimized,and an iterative search mechanism is introduced.Taking the personnel as the positioning target,the algorithm is tested on the real collected data set,and compared with the traditional sparse point cloud reconstruction and dense point cloud reconstruction.The experimental results show that the speed of this algorithm is 2-3 times faster than the traditional algorithm,and the deviation between the positioning accuracy and the real position value of the target is within 0.6m.Second,aiming at fast guiding the UAV to fine-tuning,a guidance method based on visual laser detection is designed.The host emits the Laser to the target to build a guidance point.And at the slave observation end,the image is roughly filtered based on HSV color space transformation,and the background is also separated from the differential observation image,finally the light spot area is accurately detected and extracted combined with brightness threshold filtering.Based on the optical flow tracking,the optical domain moves between frames can be detected,and the position of light points is displayed by the method of center of gravity summation.Collect the laser spot image in the real environment,and the method proposed in this paper is compared with the classical method.The experiment shows that the detection method in this paper is very stable,and the frame rate per second can reach 30fps.The UAV guidance simulation test is carried out on Prometheus platform,and the test proves the feasibility and convenience of the guidance method.Finally,the UAV hardware experimental platform is built for the realization of indoor UAV target detection and localization algorithm system.The UAV’s own pose is obtained by the binocular visual inertial odometry,whose front-end optical flow is accelerated by the graphics processing unit.The embedded processor achieves a stable running rate of 15fps,which is about 50%higher than the original running speed.The deviation of the odometry trajectory per hundred meters is within 1%.Based on the Mavros communication protocol,the positioning data can be transmitted through the serial port,the flight control parameters are able to be debugged,and the positioning results can be poured into the flight control.The test is carried out indoors,and the UAV can fly stably by obtaining the positioning result.At the same time,the observation targets are screened and localized quickly based on the visual laser and binocular relative localization algorithm.In the end,the UAV is guarded to move towards the target object.The power consumption of the system is very tiny,just about 18w and the operation can run in real time.