Research On Multi-sensor Fusion Autonomous Localization Technology For Field Search And Rescue UAV

Search and rescue drones are increasingly used in outdoor rescue activities and rescue missions such as floods and earthquakes.Efficient navigation and positioning system is one of the core factors that affecting drones in performing search and rescue tasks.However,most commonly used drone navigation and positioning technologies are strapdown navigation systems composed of inertial navigation systems and global satellite navigation systems.This navigation method relies on GPS signals,so it is not applicable in the field of GPS-denied environment.Visual navigation and inertial navigation are very complementary.Fusion of the two navigation methods can effectively improve the reliability and stability of state estimation.Therefore,this article conducts research on the key issues of autonomous positioning technology for search and rescue drones.Firstly,a search and rescue drone hardware platform was built.According to the search and rescue mission requirements,the Xiaomi camera with integrated camera and IMU is used as the main input sensor,Jetson TX1 is used as the image processing platform,and Pixhawk is used as the flight controller.For the main tasks of search and rescue drones,the software architectures of flight control system,target detection system and autonomous positioning system are designed respectively.Secondly,the improved Gaussain-YOLOv3 based on YOLOv3 is studied,which uses Gaussian distribution to express the uncertainty information of the detection frame to improve the accuracy of network detection.Aiming at the real-time requirements of the embedded platform,TG-YOLOv3(Tiny-Gaussain-YOLOv3)was proposed based on Gaussain-YOLOv3.According to the search and rescue drone operation requirements,a drone aerial dataset was constructed,and a TG-YOLOv3 target detection model was trained.Experiments show that the TG-YOLOv3 model detection accuracy is improved by 7% compared to Tiny-YOLOv3.By using the model to perform target detection experiments at different altitudes,the drone flight altitude is determined.Then,we researched the autonomous positioning algorithm of UAV.This paper selects the feature extraction algorithm SLC(Sub-pixel-accurate Low Complexity)-Harris as the visual front end.To process the IMU data between key frames of the image,a visual IMU coupled initialization method is proposed.Aiming at the problem of data frame redundancy in the sliding window,a marginalization method is proposed to remove the old state frames,thereby achieving computing resources and balance of accuracy;finally,closed-loop detection is used to eliminate cumulative errors,and re-alignment of key frames that repeatedly occur in the scene to obtain more accurate trajectory estimation.By comparing the algorithm in this paper with ORB-SLAM and VINS_MONO on the EUROC dataset,it is proved that the algorithm in this paper has higher positioning accuracy.Finally,the target detection algorithm designed in this paper is merged with the autonomous positioning algorithm of the drone,and transplanted to the search and rescue drone platform,and the drone search and rescue experiment is performed in the field and multiple scenarios. The experimental results show that the search and rescue drone can detect the rescued personnel in multiple scenarios such as open space and grassland,and has high positioning accuracy.