Research And Application Of Deep Learning Algorithm For UAS Scene Image Segmentation

Unmanned Autonomous System(UAS)is an important support system for unmanned autonomous devices(including unmanned vehicles and unmanned aerial vehicles)to perform tasks in moving scenes.Given that image segmentation is a key prerequisite for environment perception,path planning and task execution in such systems.Therefore,the research on the image segmentation theory and related technologies for UAS scenes holds great theoretical significance and applicable value.This paper proposes a real-time image segmentation method based on deep learning for vehicle and airborne application scenarios.Furthermore,the paper designs and implements an image segmentation system applied to UAS natural scenes,the main studying contents of this paper are as follows:(1)To solve the problem of road and sky segmentation of UAS natural scenes,this study builds data sets of natural scenes images,and adopts Labelme technology to label roads and sky in the natural scenes images,then establishs the suitable l semantics abel for the scheme in this paper.(2)Depending on the characteristics of the UAS natural scene images and whether the network is smooth,an image segmentation method based on deep learning is proposed in the paper.The method is composed of cloud remote image segmentation algorithm and intelligent mobile terminal image segmentation algorithm.The cloud remote image segmentation algorithm uses Kitti Seg data set to train FCN8-VGG16 network architecture.The simulation results show that the algorithm can effectively segment the images of road and sky in natural scenes,in which F1 value increases by over 4% compared with Marvin Techmann.While the mobile terminal adopts the u-net network architecture,image segmentation algorithm is realized in the intelligent mobile terminal.The simulation results indicate that the accuracy reaches 96% above and its F1 value is above 95%,which reflect that the architecture can achieve more efficient image segmentation.(3)The design based on intelligent mobile terminal is studied and the system architecture and the functions of natural scene image segmentation for UAS are implemented.The system architecture consists of mobile computing subsystem and cloud service subsystem.The mobile computing subsystem includes four modules: image acquisition,mobile terminal network communication,mobile terminal image segmentation and image display,while the cloud service subsystem consists of model training module,cloud service network communication module and cloud service image segmentation module.The system designs an adaptive mechanism of network bandwidth,depending on which can realize image segmentation automatically using different network models.When the network is unblocked,cloud service is used to segment the image,and then the segmented image is downloaded to the mobile terminal;as the network is blocked,the mobile terminal is directly used for image segmentation.This thesis designs and implements an image segmentation and application system for UAS scene,which can effectively solve the problem of image segmentation based on cloud service in the situation of different network bandwidth while ensuring the high accuracy of image segmentation.The results of this study has a certain practical significance for improving the practicability of UAS image segmentation.