| The real-time monitoring of local collapse of landslide can not only provide early warning information about the revival of old landslide,but also provide important information for understanding the revival deformation characteristics of old landslide,evaluating its stability and predicting its evolution process.Aiming to realize the low-cost,highfrequency,high-precision,fully automatic long-term continuous monitoring of local collapse in the early revival of old landslides,by combining the methods and theories of computer vision,time-lapse close range photogrammetry,deep learning,image processing,computational geometry,etc,this paper carried out the theoretical method research around the three key scientific issues of: close range photographic vegetation detection,automatic detection and acquisition of local collapse signal of landslide mass,and 3D spatial information recognition of collapse mass.The intelligent visual monitoring system for local collapse of landslide is developed based on the established theoretical method and follow the basic technical route: first separating vegetation-then acquiring collapse signal-then extracting 3D collapse information,and it has been applied to Perarolo landslide in Italy.The main research contents and results are as follows:(1)By analyzing the change characteristics of vegetation in the natural state and the collapse characteristics of the vegetation area on the surface of the landslide,this paper defines the specific role of close range photographic vegetation detection in the visual monitoring of the local collapse of the landslide mass.In order to realize the high-precision detection of vegetation pixels in the monitoring image,a deep learning model for vegetation detection is proposed.By combining with deeplabv3+ and resnet18 network,the basic architecture of the model is designed.Based on the residual learning principle and hole convolution method,the backbone network aiming at multi-scale vegetation feature extraction and fusion is designed.The specific functions,component types and network parameters of each operation layer in the backbone network are defined,and the operation methods of each network component and residual unit are derived,which make a complete backbone network operation method established.On this basis,by combining with the principle of supervised learning,a training method for the depth learning model of vegetation detection is established,and an image size adjustment method suitable for migration learning is developed based on the bicubic interpolation algorithm.With the use of the data set established in this paper,the training of the network model is completed.The training results show that the detection accuracy of the model for the vegetation pixels in the monitoring image has reached 99.1%,which can realize the accurate separation of the geotechnical layer and the vegetation layer in the monitoring image,and create good conditions for the subsequent collapse detection.(2)In view of the different visual characteristics when collapse happening on the rock-soil area and the vegetation area on the surface of landslide mass,this paper establishes the collapse detection methods for the two types of areas respectively.Aiming at the problem of collapse signal detection in the exposed area of rock-soil,a collapse detection method based on image sequence non-similarity detection is proposed in this paper.By introducing the image structure similarity algorithm(SSIM),the non-similarity detection of two continuous monitoring images(rocksoil layers)can be realized,and the structure similarity map that can reflect the collapse information can be generated;The influence of uniform illumination change and shadow change on the calculation results of dissimilarity detection is analyzed.Combined with the theory of median filter,Gaussian filter and pre learning threshold filter,an image synthesis filter is established to eliminate the false detection caused by these two factors in the structure similarity map;The calculation formula of each step in collapse detection is derived,and a complete automatic generation method of collapse signal in geotechnical region is established.(3)Aiming at the problem of collapse detection in the vegetation covered area,based on the collapse characteristics of the vegetation area,a collapse detection method based on the deep learning model of vegetation detection is proposed in this paper.The initial collapse mask of the vegetation layer can be determined by tracking the boundary changes of the vegetation area.By analyzing the source of false detection in the collapse mask,combining with the natural dynamic characteristics of vegetation,a suitable filter is established to filter the possible false detection areas in the collapse mask,so as to establish a complete automatic generation method of collapse signal in the vegetation area,which will combine with the automatic generation method of collapse signals in rock-soli area,and make verification by the theory of time lag analysis.Based on the template matching method of control points and the principle of non-reflective similar geometric transformation,the paper establishes a method and model to repair the vibration deviation of monitoring images,which can avoid the impact of camera micro vibration on collapse detection.(4)Based on the principle of binocular vision,combined with Bouguet epipolar rectify,census transformation and semi global dynamic programming,a 3D reconstruction method of landslide surface is established to meet the requirements of low-cost engineering applications.Based on the special data structure of point cloud from 3D reconstruction,combined with image filtering,region expansion,adjacent term filling,Moore boundary tracking and other algorithms,a method for extracting surface point cloud of collapse mass by image processing on depth map is established in this paper.An adaptive alpha shape algorithm is developed for the construction of 3D geometric model and volume calculation of collapse mass,so as to establish a complete 3D spatial information recognition and volume calculation method of collapse mass.(5)Based on the theory and method established in this paper,and combined with digital snapshot,image transmission,computer graphics,database programming,matlab visual programming and other technologies,this paper developed an intelligent visual monitoring system for landslide local collapse,which has the functional modules of monitoring image acquisition,monitoring information operation,monitoring data visual output.Through the engineering application of perarolo landslide in Italy for one year,and comparing the monitoring results with the mapping results of high-cost laser scanner,the reliability of the monitoring system and the correctness of the theoretical method established in the theoretical research of this paper are proved. |
