Design And Implementation Of Pedestrian Accompanying UAV At Night

As the unmanned aerial vehicle(UAV)technology matures gradually,the fusion of computer vision and UAV technology is one of the important research directions at present.Small quadrotor have been widely used in video aerial photography,military reconnaissance and public safety inspections due to their high maneuverability,low consumption and hovering advantages.Using infrared image processing technology and the quadrotor platform to track the target of interest in real time is of great significance to the safety of walking alone at night.In this paper,the tracking system is designed based on the quadrotor and airborne infrared vision sensor.The improvement of the YOLO target detection algorithm and the SiamRPN tracking algorithm are completed according to the application requirements of the aerial engineering scenario.Then verification is performed on the self-built infrared aerial data set.The algorithms are deployed on the UAV vision processing platform to realize the tracking task of specified targets at night.The specific work of the paper is as follows:Firstly,according to the application requirements,the thesis completes the selection and research of the basic components of the vision sensor,the airborne image processing platform and the quadrotor UAV.The thesis builds a platform for nighttime pedestrian UAV systems,and designs the system software architecture and detailed workflow.Secondly,According to the characteristics of infrared aerial images and the application problem of convolutional neural network on mobile airborne platforms,a Slim-YOLOv3 infrared aerial pedestrian detection method is proposed based on YOLOv3.An infrared aerial data set was constructed,which contained a variety of typical scenes and three different scales.The K-Means clustering algorithm was used to analyze the target bounding box in the self-built infrared data.The BN layer scale parameter γ factor in the YOLOv3 network is thinned,and the global threshold pruning and layer pruning strategies are used to streamline the model.Experimental results show that the Slim-YOLOv3 infrared target detection model has smaller memory and faster inference speed than before the improvement,and the detection accuracy can reach 91.5%.Thirdly,the Siamese FC and SiamRPN target tracking algorithms based on the Siamese network structure are studied.Due to the lack of color and texture information of pedestrian targets in infrared images,the SiamRPN algorithm is prone to tracking failure when the target of interest is blocked by other pedestrians.The GPS information is introduced to detect the tracking status,and tracking is re-executed when tracking fails.A concentric circular moving target search strategy is designed to overcome the problem of limited range of infrared camera imaging angle of view.Experiments have proved that the SiamRPN algorithm can effectively predict pedestrian targets in the infrared image sequence.It can run at a speed of 15 fps on the Jetson TX2 computing platform,and the predicted target position error does not exceed 3 pixels.Finally,the thesis performs a system test and analysis on the pedestrian accompanying UAV system at night.The flight stability test was performed on the UAV platform.SiamRPN algorithm and Slim-YOLOv3 model were completed and deployed on the Jetson TX2 platform.Perform tracking flight tests on the system in night scenes and analyze the tracking success rate in different scenes.The results show that the system can track interest pedestrian in night scenes,and the tracking success rate can reach 87%.