Research On Algorithms For Object Detection In Aerial Images

With the rapid development of aerial photography equipment,intelligent systems based on drone and aerial photography are being deployed more and more widely in various fields,such as security surveillance,traffic management,search and rescue,etc.The object detection in aerial images is an important fundamental component of these intelligent systems.In recent years,detectors based on convolutional neural network have made significant improvements in general object detection,but aerial images are more challenging compared to general scene images,with a larger range of object scale variations and a greater proportion of tiny objects.This paper analyzes the main problems of the existing detectors and proposed improvement methods to address these problems,and finally obtain a single-stage anchor-free detector with high accuracy.The main work of this paper consists of the following sections:(1)In response to the problem that existing feature pyramid construction methods do not balance the multi-layer complementary features of backbone network,this paper proposes a framework to construct feature pyramids based on Up-scale Feature Aggregation networks.It can balance the complementary multi-layer features information of backbone network,making the final constructed feature pyramids more representational and improving object detection accuracy.With ablation experiments on the Vis Drone2018 dataset,this paper demonstrates that the proposed Up-scale Feature Aggregation network can effectively improve the detection average precision by 4.9% compared to the benchmark detector.(2)Up-sampling is used to change the length and width of features during the multi-scale feature fusion.Because existing up-sampling methods are not efficient enough,we propose a novel learning-based up-sampling method called Cascade Feature Up-sampling.Experiments on the Vis Drone2018 show that the proposed Cascade Feature Up-sampling can effectively improve the detection accuracy,which is 1.4% AP higher than the bilinear interpolation.(3)In anchor-based detectors,the hyperparameter tuning of anchor is difficult and the huge number of anchor boxes leading to high model parameters and computational overhead.This paper adopts an anchor-free detection head that does not require anchor hyperparameter tuning,reduces the number of model parameters and computational overhead.The experimental results show that the detection accuracy improved by 1.3%,0.5% and 1.0% on the Vis Drone2018,UAVDT and DIOR datasets,respectively.The runtime speed reaches 11.8FPS on the Nvidia GTX 1080 Ti graphics card,which improved by 1.2FPS compared to the anchor-based detector.