Research On UAV Aerial Vehicle Detection Technology

With the development and promotion of computer vision and unmanned aerial vehicles,the research of object detection based on aerial images of unmanned aerial vehicles(UAVS)has become a hotspot in the field of computer vision.The goal of object detection is to determine whether or not there are any instances of objects from the given categories in some given image and,if present,to return the spatial location and extent of each object instance.With the rapid development of deep learning,the field of object detection has made breakthrough progress.With the advantage of high mobility,UAVs have large camera angle.And aerial images taken by UAVs have more objects and more complicated background.Objects sizes are different because of the UAVs flight at different heights.And UAVs usually flight at high altitude,there are more small objects.It brings new challenges to object detection.Object detection algorithm is mainly divided into the traditional object detection algorithms and object detection algorithms based on deep learning,features used by traditional object detection algorithms are designed by artificial,it usually applies to specific application scenarios.Object detection algorithms based on deep learning have strong ability of extracting features and have better detection results than traditional object detection algorithms.By studying the object detection algorithm based on deep learning,this paper improved the SSD and improved the detection effect of UAV aerial images.The main work of this paper is as follows:In object detection algorithms based on deep learning,feature maps with different resolution sizes are usually used to detect object with different sizes.Large feature maps will be used to detect small object due to their small receptive fields,while feature maps of small resolution will be used to detect large objects due to their large receptive fields.This paper analyzes the network structure of SSD and finds that the low-level feature maps used to detect small objects contain more information about small objects but lacks sufficient semantic information,while the high-level feature maps contain rich semantic information but lack information about small object.Therefore,this paper puts forward a improved algorithm based on feature fusion of SSD by fuse high-level feature maps with rich semantic information and low-level feature maps with more small object information.In the process of feature fusion,we proposed a new method to fuse multiple feature maps.In the final detection process,because each channel of the feature map has different effect on object detection,we proposed a feature enhancement method to enhance the effective channel of the detection layer and suppress the invalid channel.Through the experiment,we found that the proposed algorithm on the UAV aerial data set UAVDET have 29.8%mAP,4.2%higher than the SSD’s,and the detection speed is 19fps.Through the analysis of the existing one-stage object detection algorithm,it can be seen that the current object detection algorithms adopt the method of setting the default bounding boxes of fixed size and aspect ratio in each unit of the feature map of different scales and then classify and regress these default bounding boxes.The region based object detection algorithm can generate high quality candidate regions from the network,such as Faster R-CNN can generate 2000 candidate regions through RPN network for object detection.However,the location of the default bounding box is fixed in the region free object detection algorithm,so the default bounding boxes may not match well with the ground truth,especially for small object which results in that the detection algorithm cannot locate the object location well.In addition,this method will generate a large number of bounding boxes,which is far more than region based algorithm.As a result,in the region free object detection algorithm,the distribution of positive and negative samples in the candidate bounding boxes is extremely uneven,so that the train network detection effect is not as good as that of the region based object detection algorithm.Due to the background of UAV aerial images are more complex,and have more small target,one-stage object detection algorithm is not good for small object locating,this paper proposes a two-stage algorithm based on two-level features fusion,first of all,in the first level feature pyramid,filter out part of the background bounding boxes,and adjust the default bounding boxes to make them more close to the ground truth,and then these bounding boxes will be the input of second stage,this will promote the accurate of small object detection.Through experiments,we found that the algorithm proposed in this paper achieves 32.1%mAP on UAVDET,6.5%higher than SSD,and detection speed is 14fps.