Research On Real-time Detection Technology Of UAV Aerial Photography Vehicle Based On Target Spatial Distribution Feature

In recent years,the research of intelligent transportation system has been widely concerned by scholars from all walks of life.Among them,accurate and real-time vehicle detection on the road is an essential part of the construction of intelligent transportation system.However,the traditional method of detecting vehicles in road traffic by installing a vision sensor in a fixed position has small coverage,poor flexibility and high cost.With the continuous improvement of UAV technology,the research on road vehicle detection based on UAV carrying vision sensor has become a research hotspot in recent years.Compared to cameras mounted in a fixed position,aerial drone photography can provide a larger perspective,a wider range,and more flexibility,and can be quickly applied to critical road areas and sudden traffic scenarios.However,while giving full play to the advantages of UAV aerial photography with large field of vision and high flexibility,the small-scale vehicle targets with few feature points and the requirements of rapidity in most application scenarios also bring challenges to accurate and real-time detection of vehicles.At present,with the rapid development of deep learning technology,convolutional neural network has a greater advantage in detection performance compared with previous methods based on target motion feature detection and traditional machine learning by virtue of its powerful feature extraction ability.However,in the perspective of UAV aerial photography,due to the large image resolution and the small scale of most vehicle targets,it is difficult to obtain good results in accuracy and speed when the classical convolutional network algorithm is directly applied to vehicle detection.In order to solve the above problems,this paper comprehensively considers the optimization of convolutional network detector and the adaptive segmentation of aerial images,and studies the real-time detection technology of UAV aerial photography vehicle based on target spatial distribution characteristics.The main research contents are as follows:(1)Based on single-stage convolutional neural network,the real-time detection method of UAV aerial photography vehicle is studied.Based on the single phase target detection framework of SSD,the standard convolution operation optimization for the effective depth of separable convolution,to reduce the feature extraction of the aliasing effect caused by continuing drop sampling background noise caused by the introduction of small and medium scale of aerial image vehicle target feature extraction,in front of the pooling layer design can anti aliasing low-pass filter module is to learn to,Build aerial vehicle detection network E-SSD.In this paper,K-means++ clustering algorithm is used to obtain the parameter information of the default candidate box for vehicle detection task in aerial photography scene,and the E-SSD network is tested.Compared with the standard SSD network,the speed of E-SSD network has been greatly improved,reaching71 FPS,and the accuracy has been improved by 4 percentage points,providing an efficient detector for the construction of subsequent detection network.(2)Based on the idea of conditional generative adversarial,the vehicle density estimation method of UAV aerial photography is studied.To get the accurate distribution of aerial images in the vehicle,to generate high quality vehicle density figure,the U coding,decoding structure,and introduced hybrid expansion in the encoding and decoding structure convolution HDC and residual block building density figure estimated network as a generator,at the same time,based on Patch GAN discriminator for density diagram is used to identify true value and the result of the generator,on the basis of the objective function continuously optimized density figure weight parameter to estimate the network,so as to improve the quality of aerial vehicle density diagram,experiments show that of the aerial vehicle density estimation network generation density figure compared with existing methods must be promoted,It lays a foundation for the accurate image segmentation in the future.(3)Based on the spatial distribution characteristics of the target,a vehicle detection method for UAV aerial photography is studied.UAV aerial vehicle density estimation to generate network combined with real-time optimized aerial vehicle detector,the network is estimated based on the density of the space distribution features of vehicle adaptive segmentation study,and the global image block with the original local after segmentation image of the detector,input the optimized based on the Soft-NMS algorithm in the decision-making level fusion detection results,thus realize the aerial vehicle target accurately and efficiently under the perspective of road traffic scene detection.The method improves the input data source by using the target distribution features,and enriches the detailed feature information of the vehicle target,especially the small scale target.Relevant experiments were designed to analyze the detection performance of the whole network.Experiment showed that the DF-NET,aerial vehicle detection network designed in this paper based on the spatial distribution characteristics of targets further improved the detection accuracy of vehicle targets in aerial images and maintained a good real-time performance.