Research On Small Object Detection And Group Object Tracking In UAV Ground Reconnaissance Images

Autonomy is the basic feature of UAV systems.Object detection and tracking in complex environments is the basis for UAVs to achieve autonomous capabilities.However,there are still many unsolved problems in UAV object detection and tracking.In this paper,a small object detection algorithm based on deep learning is designed for the problem that the flight height of the UAV and the physical performance constraints of the reconnaissance load lead to smaller object in the reconnaissance image.Aiming at the problem that the current detection algorithm is difficult to accurately detect the object subclass problem,a metric-based subclass object detection algorithm is designed.The group object trajectory extraction method is designed to solve the problem that it is difficult to extract the group object in group object tracking and the tracking trajectory does not match the actual trajectory.The main research contents and innovations of the thesis are as follows:(1)Facing the application background of UAV object reconnaissance,this paper combines the UAV’s airborne visible light sensor to design a small object detection and tracking framework.The framework covers three parts: UAV-to-ground small object detection,UAV-to-ground subclass object detection,and UAV-to-ground group object tracking.Considering that the dataset taken by UAVs flying at high altitude is difficult to obtain,this paper uses the data obtained by UAVs shooting small military model objects at low altitudes to simulate the data from drones flying at high altitudes.The paper constructs dynamic and static comparison dataset,small object detection dataset and subclass object dataset,and compare the algorithm proposed in this paper with the classic object detection algorithm on these datasets.(2)Design a small object detection algorithm based on the characteristics of the UAV reconnaissance image.The algorithm is improved from three aspects: data augmentation,frame design,and loss function calculation.In order to improve the adaptability of the detection model to the scene,two data augmentation methods,background replacement and add noise,are designed to enrich the training set images.In order to improve the speed and accuracy of small object detection,the DCSPDarknet framework is designed based on the CSPDarknet.First,the network layer in the CSPDarknet that has little effect on improving the detection effect of small objects is deleted,and then the image downsampling network module is added to obtain more small object details without introducing too many calculations.In order to more accurately distinguish between positive and negative samples and increase the number of positive samples,a new method for selecting and defining positive samples is designed when calculating the loss function.Experiments are performed on the small object detection dataset.Compared with YOLO v4,the algorithm proposed in this paper increases the AP50 by 24.5% without increasing the amount of calculation.(3)Aiming at the problem that the current object detection algorithm is not effective in detecting object subclass,a subclass object detection algorithm based on deep layer positioning and low layer classification is designed.The subclass object detection algorithm is improved based on the small object detection algorithm.It first obtains all object positions through deep features,then extracts object features from the low layer features according to the object position,and finally classifies the objects based on the distance metric.When the algorithm is trained,the triple loss is used to maximize the distance between different categories of objects.After the training is completed,the labeled images are input into the model to obtain standard object features to construct an object classification database.The classification database includes two parts: object class and object subclass.When the algorithm is tested,the distance between the object vector and all the features in the classification database is calculated and the object is divided into the category closest to it.The object class is determined first,and then the object subclass is determined.Detect the object class,it is found that the subclass object detection algorithm improves the AP50 by 13.6 compared with the small object detection algorithm.Detect untrained objects,AP50 reached more than30%.Experiments prove that the subclass object detection algorithm need not fix the object category and improve the detection effect.(4)Design a group object trajectory extraction algorithm for the problems existing in group object tracking.First,use corner point extraction and optical flow method to obtain the motion trajectories of all corner points in the image.Then,clustering algorithm is used to cluster the motion vectors of the corner points to distinguish the moving foreground points and the background points.Extract the actual motion trajectory of the moving foreground object according to the difference between the moving foreground points and the background points.According to the actual trajectory of the moving foreground corner points,the trajectory matching algorithm can be used to obtain the movement range of the group object.Finally,combined with the object detection result,the object corner points that do not belong to the group object and the redundant object corner points can be filtered to obtain the actual group object trajectory.Through experiments,it is found that the group object trajectory extraction method proposed in this paper can better extract the group object and obtain the true trajectory.