Research On Vehicles Targets Detection In Remote Sensing Images Based On Deep Learning

Computer vision combined with satellite remote sensing technology has broad application prospects in the field of intelligent transportation.Compared with the traditional traffic monitoring equipment,the target detection method based on satellite remote sensing image has many advantages,such as a long-term application once it put into use,no damage to the road surface,no impact on ground traffic,large coverage area,abundant traffic information acquired and so on.It provides new data and method for the traffic management and the dynamic monitoring of traffic flow,which will bring new changes to the daily traffic monitoring and management model.This paper has two focuses,one is the road segmentation from high resolution satellite remote sensing images,and another one is the dim and small vehicles targets detection and tracking in high resolution video satellite remote sensing images.For the problem of road segmentation in remote sensing images,firstly,five image segmentation networks based on deep learning,such as PSPNet,DeepLabv3,DeepLabv3+,DenseASPP and U-Net,are compared and analyzed.Then these networks are applied to road segmentation task in remote sensing image.Its performances are evaluated and analyzed in many aspects.Finally,based on the defects of U-Net structure and the characteristics of road diversity in large-scale remote sensing scenes,an improved U-Net structure is proposed,which can effectively extract road areas from remote sensing images and has simpler structure.The improved U-Net network has better performance in road segmentation task of remote sensing image.For the problem of detection and tracking of dim and small moving vehicle targets with less texture features in video satellite remote sensing images,a two-step scheme of road segmentation and moving vehicle targets detection and tracking on the road is proposed,in which road regions in remote sensing images are extracted by improved U-Net network firstly,and then moving vehicle targets are detected only in the road areas.Moving vehicle targets detection is divided into two steps: targets pre-extraction and precise extraction.In the preextraction step,the suspected moving vehicle targets is extracted preliminarily by combining the improved inter-frame difference algorithm and other algorithms.In the precise extraction step,the region growing algorithm and area threshold method are used to accurately extract the position of each vehicle target and effectively remove false targets.Finally,motion estimation and position correlation are combined to achieve synchronous tracking of all moving vehicle targets.The test data are based on 1-meter resolution video satellite remote sensing images.The experimental results show that the proposed methods had high detection accuracy for the dim and small moving vehicle targets,and achieve near real-time tracking processing.