Research On Object Detection Algorithm Of Aerial Image Based On Coarse-grained Density Map

The rapid development of drones and satellite remote sensing has made aerial images more and more accessible.Compared to images taken in natural scenes,aerial images can usually contain more feature information,so how to detect in aerial images has become a hot research topic.However,due to the special shooting angle,the scale of the objects that are meaningful for detection in aerial images are usually very small,such as people,cars,etc.Moreover,aerial images also have problems such as complex background,imbalance between positive and negative examples,and imbalance between object classes,this causes great difficulty in object detection.The current aerial image detection algorithm based on convolutional neural network has become the mainstream,this type of algorithm usually divides the detection of aerial images into three stages.The first stage obtains the semantic information of the image,such as object cluster area,object density map,etc.The second stage uses these semantic information to crop the image.In the third stage,object detection is performed on cropped blocks one by one,and the final results are merged.The main reason for the great success of such algorithms in aerial detection is that image cropping can increase the proportion of the object area in the image.But this kind of algorithm also has some flaws,including the unreasonable proportion of the area of the object in the cropped block leads to low detection accuracy,and too many cropped blocks cause low detection efficiency.For the detection difficulties of aerial images and the defects of aerial detection algorithms in the past two years,this paper proposes an aerial image detection algorithm based on coarse-grained density maps.First of all,this paper designs a coarse-grained density map and a density estimation model.Then use the smallest bounding box of the object connected area in the density map as the basis for image cropping,and the closed operation and regression model adjust the number of cluster regions and the scale of the objects in the cluster region.The Mosaic enhancement and adaptive training sample selection mechanism are used in the training phase to solve the imbalance of positive and negative examples,imbalance between classes,and small object training problems.Finally,detect the cropped blocks one by one,and merge the detection results of all cropped blocks in an image.The model is tested on four datasets: VisDrone,UAVDT,DOTA and TT100 K.The experimental results show the superiority of this algorithm.On the VisDrone and UAVDT datasets collected by drones,compared with the aerial image object detection model in the past two years,the algorithm in this paper has a significant improvement in detection accuracy,this shows that the algorithm in this paper is very effective for images taken by drones.For the satellite image DOTA and the traffic signal image TT100 K,the algorithm in this paper has obtained similar effects to the mainstream algorithm,which shows that the algorithm in this paper is robust to different scenarios.