Research On Pedestrian Detection In Aerial Video Based On UAV

With the vigorous development of wireless communication technology,UAV image application is a powerful supplement to satellite remote sensing,and the application of target detection technology in UAV aerial photography scene has also become a focus topic in theoretical circles.However,although the combination of UAV aerial photography and computer vision has greatly facilitated life,it still faces the following problems to be solved urgently: how to avoid missing and wrong detection in pedestrian detection in UAV aerial video;How to block each other between pedestrians and detect them under different lighting conditions.Common target detection algorithms have good detection effect in natural scenes,but the target detection effect in aerial photography of unmanned aerial vehicles is not ideal.Based on the above problems,thesis proposes an improved pedestrian target detection algorithm for aerial video,and the main research work has been completed as follows:(1)The production of pedestrian detection data set in UAV aerial video.In this thesis,a dataset of pedestrian target detection in UAV aerial video is produced,which has less pedestrian dataset and uneven sample distribution for UAV aerial video.Through the relevant technology of data enhancement,the number of data sets is expanded,and the features of pedestrians in the processed data sets become richer,which reduces the difficulty of network detection and improves the detection accuracy.(2)The network structure is improved.Multi-scale feature fusion and improved attention mechanism are proposed.It not only optimizes the network structure,but also enriches the feature extraction ability of pedestrians.The optimization of the network structure avoids filtering out small-scale pedestrians as noise in complex background,reduces the missed detection of smallscale pedestrians in aerial video,reduces the parameter quantity of the model by 2.83 M,and improves the average accuracy of pedestrian detection by 3.46%.(3)A fused attention and image pyramid called SPCF structure is proposed.It mainly solves the problems of multi-scale changes of pedestrians and complex background in aerial video,makes the network pay more attention to the areas where small pedestrian targets gather,and at the same time endows feature maps with different scales with dynamic weight information,so that feature maps with different sizes can be better integrated with small target features in aerial video after up-sampling and down-sampling.It not only enriches the characteristics of pedestrian small targets,but also solves the problem of pedestrian small target feature loss when feature fusion,Compared with the unfused model,the average accuracy is improved by 4.93 percentage points.Finally,through the verification of experiments,the improved parts of this paper show good detection effects.The size of the network model is 21.22 MB,and the simple parameters of the network structure are compared with the original network Performance on the pedestrian data set of aerial video produced in this thesis Compared with the original algorithm,the average accuracy is improved by 5.54 percentage points.The detection speed reaches 62.7 FPS.