Research On Image Semantic Segmentation Algorithm And Its Application In Unmanned Surface Vehicle

Surface unmanned vehicle(USV)is a kind of surface intelligent equipment.Real-time environmental information is necessary in USV while performing maritime tasks or docking.However,there are not only objects such as ships and buoys,but also stuffs such as shore and dock in surface environment.The existing environment perception methods of USV based on target detection or coastline detection cannot parse all the surrounding information.Semantic segmentation is a pixel-level classification of images which is free from the shape of objects.This enables semantic segmentation to parse various objects in the surface environment and provide rich and accurate environmental information for USV.Therefore,based on the application requirements of USV,this article has carried out a series of research on semantic segmentation related technologies.This paper proposes the Multi-Direction Convolution(MDC)for context information fusion.Context is known to be one of crucial factors effecting the performance of semantic segmentation.However,state-of-the-art segmentation models built upon fully convolutional networks inherently lack contextual information because of stacked local operations such as convolution and pooling.Therefore,this paper proposes the MDC to fuse contextual information.Different from standard convolution,MDC no longer fixes the convolution kernel on channel,but fixes on height,width and channel.Rich contextual information is encoded by fusing the convolution in each direction.MDC can always maintain the original size of the feature map when encoding context information,which makes the context features of each position dynamically learned and more adaptable.MDC is also efficient and can be implemented with few standard convolution layers with permutation.Extensive experiments show that MDC outperforms existing contextual modules on two standard benchmarks,including Cityscapes and PASCAL VOC2012.This paper proposes a real-time semantic segmentation network in driving scene.The driving scene is a kind of high-resolution scene and most of the information in the distance is redundant.Most of the existing semantic segmentation networks do not take this situation into consideration and additional structures is usually used to obtain global information or large receptive fields,which will undoubtedly bring more calculations and limit the inference speed.Therefore,this paper proposes the Limited Receptive Field Network(LRFNet)for real-time driving scene semantic segmentation.LRFNet no longer uses additional structure to introduce larger receptive fields or global information,which reduces the amount of calculation and avoids coding too much redundant information.LRFNet has two sub-encoders,which are used to obtain accurate spatial information and rich high-level semantic information respectively,and use a lightweight decoder to restore the details.Experiments show that LRFNet achieves a good balance between speed and accuracy on the two driving scene datasets(Cityscapes and CamVid)which gives the highest accuracy and one of the fastest speed among real-time semantic segmentation networks.Finally,this paper proposes an environment perception system of surface unmanned vehicle(USV)based on semantic segmentation.The system uses a panoramic camera as sensor,which can obtain all the surrounding information.Then the semantic segmentation algorithm is used to parse the panoramic image.Experiments on the surface datasets(MaSTr1325 and Tampere-WaterSeg)show that the semantic segmentation can effectively parse the objects such as ships,water surface,and shore in surface environment,providing rich and accurate environmental information for USV to navigate and perform tasks safely.