Research And System Implementation Of Building Segmentation Method For Video Surveillance Image

In recent years,China’s urban over-development,illegal land use,illegal construction and other problems become increasingly prominent,urgent need to carry out efficient and accurate building segmentation technology research.In view of the existing problems of building segmentation based on remote sensing images in natural resource monitoring,such as single means,low timeliness and low accuracy of building segmentation based on traditional artificial visual interpretation,this paper is devoted to the research of building automatic segmentation key technology based on video surveillance images.And developed a video surveillance image building segmentation system based on DeepLabv3+model.The main research contents and achievements of this paper are as follows:(1)The process of data set annotation and data set augmentation method are introduced,and the scene building segmentation data set with building label is created.For the problem of small amount of data,open source street View image Camvid is processed into two categories of images,that is,only buildings and backgrounds are included for pre-training of the model.(2)Based on the collected data,the DeepLabv3+ model was used to carry out experimental comparison.The effects of different ASPP modules’ void convolution,deconvolution and bilinear interpolation upsampling methods and Res Net50 and Res Net101 trunk feature extraction networks on building segmentation are compared.Experimental results show that DeepLabv3+ model adopts Res Net50 backbone feature extraction network,bilinear interpolation up-sampling method and expansion rate of(12,24,36)to achieve the best segmentation effect on the collected data set,and the highest Io U reaches 87.51%.(3)Based on the improvement of DeepLabv3+ model,a DeepLabv3+ model integrating CBAM attention mechanism is proposed.The model uses the idea of deep supervision to involve the output of ASPP module in the calculation of loss,and uses UNet’s jump connection idea to merge the layer2 output of backbone network Res Net with the sampling result of ASPP module’s double bilinear interpolation to further integrate the semantic information of high level and the texture information of low level.The experimental results show that the Io U of the improved model increases by 1.09%,which can clearly extract the building and obviously improve the details of the edge of the building.(4)The building segmentation algorithm is servified and deployed by flask technology and Docker technology.How flask servize algorithm and Docker package algorithm into image and deploy algorithm on server side are described.(5)Combined with front end technology,geographic information technology and cameras,real-time display technology research and development of the video surveillance image segmentation system,building the system is integrated in guilin 50 public security cameras and 1 installed on the tower of hd intelligent high-speed camera haeundae network,can real-time playback,click on the camera to the camera captured image segmentation for buildings.