Background technique
Variation detection is used as land cover pattern monitoring, land use monitoring, Disaster Assessment, hazard prediction, geographic information data
One of key technologies in fields such as update, have been a concern, and variation detection includes Changing Area Detection and change type identification,
Traditional variation testing process is to generate disparity map and region of variation classification, and the acquisition methods of disparity map include image difference, figure
As ratio etc., these methods pixel-based are only applicable to large scale satellite image or low resolution aviation image, for answering
With increasingly frequent high resolution image, then a large amount of fragment easy to form is unfavorable for so that generating excessive pseudo- region of variation
Later data processing, traditional classification method is divided into supervised classification and unsupervised classification, but they are based on image, merely with shadow
The colouring information of picture, classification foundation is excessively single, and the accuracy rate of classification is not high.
With the rapid development of unmanned air vehicle technique, it is excellent that unmanned plane image is low, high-efficient with its procurement cost, resolution ratio is high etc.
Point is more and more applied to geography information Source Data Acquisition, utilizes the outdoor scene three-dimensional modeling data of unmanned plane video generation
As important one of geographic information data, outdoor scene three-dimensional modeling data has color and geological information simultaneously, is applied to variation
In detection, while being made according to the higher characteristic of unmanned plane image resolution using Object--oriented method with the object of segmentation
It is changed detection for basic unit, is greatly improved variation detection accuracy.
Deep learning is a new field in machine learning research, and motivation is that foundation, simulation human brain are divided
The neural network of study is analysed, it imitates the mechanism of human brain to explain data, such as image, sound and text, and deep learning passes through
Construct the training data of the machine learning model and magnanimity with many hidden layers, it is intended to the more good feature of study, and then improve
Classification accuracy is traced to its source, and the concept of deep learning is derived from the research to artificial neural network, by combining low-level feature shape
It is special to solve image point to find that the distributed nature of data indicates at more abstract high-rise expression attribute classification or feature
Class identifies that the convolutional neural networks (Convolution Neural Networks, CNN) of problem are a kind of with convolutional coding structure
Deep learning network, CNN can automatically extract space characteristics from image, and pixel to be sorted and its neighborhood pixel are made together
For the input of convolutional neural networks, and then be converted to the feature efficiently used by machine learning task, in recent years, neural network side
Method is for solving the problems, such as that image classification has reached its maturity, and application field is also extending, more traditional classification method, depth
Learning method has the powerful ability from a few sample focusing study data set substantive characteristics, can greatly improve variation type of ground objects
Recognition accuracy.
In conclusion utilizing outdoor scene three-dimensional the invention proposes a kind of change detecting method based on outdoor scene threedimensional model
Model data detects region of variation with Object--oriented method, and is identified by deep learning method to change type, greatly
Variation detection accuracy and change type recognition accuracy are improved greatly.
Summary of the invention
(1) the technical issues of solving
In view of the deficiencies of the prior art, the present invention provides a kind of change detecting method based on outdoor scene threedimensional model, solutions
It has determined traditional change detecting method based on image, merely with the colouring information of image, in variation detection accuracy and variation
The problem of reaching promising result is difficult in type identification accuracy rate.
(2) technical solution
In order to achieve the above object, the present invention is achieved by the following technical programs: a kind of based on outdoor scene threedimensional model
Change detecting method, specifically includes the following steps:
S1, the overlapping region according to the range of outdoor scene threedimensional model, before and after calculating variation;
S2, more threedimensional models carry out texture resampling and elevation resampling, generate digital positive photograph as DOM and digital ground
Model DSM;
S3, by the digital positive photograph in step S2 as DOM and digital terrain model DSM image conversion processing after, carry out towards
The segmentation of object generates figure spot object set;
S4, changed according to the elevation in each figure spot object, judge whether figure spot is region of variation;
S5, the sample data for acquiring different types of ground objects respectively;
S6, sample training;
S7, classifier is generated;
S8, input figure spot color and elevation information, obtain the type of variation atural object.
Preferably, overlapping region is calculated in the step S1 specifically: read in the outdoor scene threedimensional model of variation front and back first
Data calculate variation front and back zone boundary range, obtain the overlapping region of variation front and back, block length and width are then arranged, to overlapping
Block is repartitioned in region, carries out subsequent processing to each block.
Preferably, the model resampling in the step S2 is divided into texture resampling and elevation resampling, according to block
Bounds, generate level sampling grid, and size delta x, the Δ y of grid are according to the resolution setting of model, it is known that grid rises
Beginning coordinate (x0,y0), the horizontal coordinate of mesh point (i, j) can be obtained are as follows: x=x0+ i* Δ x, y=y0+j*Δy。
Preferably, the horizontal coordinate according to mesh point takes the texture face of the model points on corresponding position on model
Color value produces digital positive photograph as DOM, then according to the horizontal coordinate of mesh point, corresponding positions is taken on model as z value
The height value for the model points set generates digital terrain model DSM as z value.
Preferably, DOM and DSM is divided in the step S3, and generates figure spot object set, is just being taken the photograph using the number of generation
Image is split image using effective dividing method based on figure, Image Segmentation is specific at several, have it is only
The region of property values, it is able to maintain the details of low region of variation, while can ignore the details in High variation region, to reduce tiny
The generation of fragment obtains a good segmentation effect, and using digital terrain model DSM, mesh point height value is stretched to 0-
255 ranges generate grayscale image, with the method for Threshold segmentation, by Image Segmentation at several regions not overlapped, by two kinds points
It cuts result to merge, final figure spot object set can be obtained.
Preferably, it is to count the height in figure spot to each figure spot object that region of variation is judged whether it is in the step S4
Poor mean value, given threshold are considered as candidate change region, are otherwise considered as unchanged region if height difference mean value is higher than this threshold value, raw
At initial region of variation.
Preferably, the generation classifier in the step S7 is generated using the method for deep learning, deep learning network
Flocked together by multiple neural units and construct layering as a result, simplest network by an input layer, an output
Layer, a hidden layer form, and have multiple neurons on each layer, and do not have nerve of the neuron all with next layer on one layer
Member links together, and upper one layer of the input exported as next layer, such network is also known as fully-connected network.
Preferably, prediction variation type of ground objects is input atural object figure spot and classifier in the step S8, into classifier into
Row calculates, and exports the probability of each classification, and probability soprano is the classification for changing atural object.
(3) beneficial effect
The present invention provides a kind of change detecting methods based on outdoor scene threedimensional model.Have compared with prior art following
The utility model has the advantages that should change detecting method based on outdoor scene threedimensional model, specifically includes the following steps: S1, according to outdoor scene three-dimensional mould
The range of type, calculates the overlapping region of variation front and back, and S2, more threedimensional models carry out texture resampling and elevation resampling, generate
Digital positive photograph as DOM and digital terrain model DSM, S3, by the digital positive photograph in step S2 as DOM and digital terrain model
After the processing of DSM image conversion, the segmentation of object-oriented is carried out, generate figure spot object set, S4, according to the elevation in each figure spot object
Variation judges whether figure spot is region of variation, S5, the sample data for acquiring different types of ground objects respectively, S6, sample training, S7,
Classifier is generated, S8, input figure spot color and elevation information obtain the type of variation atural object, it can be achieved that by utilizing outdoor scene three
The color and geological information of dimension module, are split with Object--oriented method, are changed detection by basic unit of object,
After determining region of variation, using the method for deep learning, atural object change type is identified, substantially increases variation detection essence
Degree and change type recognition accuracy, while color and geological information is utilized, it is provided further to reject pseudo- region of variation
Advantage improves detection accuracy, and enriches classification foundation.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
Referring to Fig. 1, the embodiment of the present invention provides a kind of technical solution: a kind of variation detection based on outdoor scene threedimensional model
Method, specifically includes the following steps:
S1, the overlapping region according to the range of outdoor scene threedimensional model, before and after calculating variation;
S2, more threedimensional models carry out texture resampling and elevation resampling, generate digital positive photograph as DOM and digital ground
Model DSM;
S3, by the digital positive photograph in step S2 as DOM and digital terrain model DSM image conversion processing after, carry out towards
The segmentation of object generates figure spot object set;
S4, changed according to the elevation in each figure spot object, judge whether figure spot is region of variation;
S5, the sample data for acquiring different types of ground objects respectively;
S6, sample training;
S7, classifier is generated;
S8, input figure spot color and elevation information, obtain the type of variation atural object.
The present invention calculates overlapping region in step S1 specifically: the outdoor scene three-dimensional modeling data of variation front and back is read in first,
Variation front and back zone boundary range is calculated, the overlapping region of variation front and back is obtained, then block length and width is set, to overlapping region weight
It is new to divide block, subsequent processing is carried out to each block.
The present invention, the model resampling in step S2 is divided into texture resampling and elevation resampling, according to the boundary of block
Range, generates level sampling grid, and size delta x, the Δ y of grid are according to the resolution setting of model, it is known that the starting of grid is sat
Mark (x0,y0), the horizontal coordinate of mesh point (i, j) can be obtained are as follows: x=x0+ i* Δ x, y=y0+j*Δy。
In the present invention, according to the horizontal coordinate of mesh point, the texture color of the model points on corresponding position is taken on model
Value is used as z value, that is, produces digital positive photograph as DOM, then according to the horizontal coordinate of mesh point, corresponding position is taken on model
On model points height value as z value, generate digital terrain model DSM.
The present invention, DOM and DSM is divided in step S3, and generates figure spot object set, is to utilize the digital positive photograph generated
Picture is split image using effective dividing method based on figure, Image Segmentation is specific at several, have uniqueness
The region of property, it is able to maintain the details of low region of variation, while can ignore the details in High variation region, to reduce tiny broken
The generation of piece obtains a good segmentation effect, and using digital terrain model DSM, mesh point height value is stretched to 0-255
Range generates grayscale image, is divided two kinds by Image Segmentation at several regions not overlapped with the method for Threshold segmentation
As a result it merges, final figure spot object set can be obtained.
The present invention, it is to each figure spot object that region of variation is judged whether it is in step S4, and the height difference counted in figure spot is equal
Value, given threshold are considered as candidate change region if height difference mean value is higher than this threshold value, are otherwise considered as unchanged region, generate just
The region of variation of beginning.
The method that pseudo- region of variation is rejected has: 1) using vegetation index reject the unessential region of variation such as vegetation, according to
The rgb value of image calculates vegetation index EGI=2G-R-B or nGEI=(2G-R-B)/(2G+R+B) of figure spot object, if becoming
The vegetation index for changing front and back all exceeds threshold value, then is considered as pseudo- region of variation;2) tiny isolated region, according to the change of variation detection
Change region area is generally large, and tiny isolated area can be considered pseudo- variation;3) have the unconventional geometry such as long and narrow, height is recessed special
The figure spot of property is considered as pseudo- region of variation, if figure spot is judged as region of variation, thens follow the steps S8, otherwise terminates.
The present invention, the generation classifier in step S7 are generated using the method for deep learning, and deep learning network is by more
A neural unit flock together and construct layering as a result, simplest network by an input layer, output layer, one
A hidden layer forms, and has multiple neurons on each layer, and does not have neuron of the neuron all with next layer on one layer to connect
It is connected together, upper one layer of the input exported as next layer, such network is also known as fully-connected network, deep learning one
As use multilayer neural network, be made of three parts: 1) input layer, be responsible for data acquisition;2) by n convolutional layer and pond layer
Group is combined into extract feature, i.e. hidden layer, externally invisible;3) the multi-layer perception (MLP) classifier structure that output layer is connected entirely by one
At.
The last layer of disaggregated model is usually Softmax regression model, its working principle is that will can be determined that as certain class
Feature be added, it is this kind of other probability that these features, which are then converted to judgement, feature is described as:
featuresi=∑jWi,jxj+bi
I represents the i-th class, and j represents j-th of pixel an of image, biIt is bias (tendency for representing data itself), W generation
Table weight parameter, x indicate the image data of input.
Next as a result as follows to all feature calculation softmax:
Softmax (x)=normalize (exp (x))
Where it is determined that the probability for the i-th class can be obtained by following formula.
It for training pattern, needs to define the nicety of grading that a loss function carrys out descriptive model to problem, loses
Function is smaller, and the classification results of representative model and the deviation of true value are smaller, i.e., model is more accurate.To more classification problems, usually
Use cross entropy (Cross-entropy) as loss function, Cross-entropy is defined as follows, and wherein y is point of prediction
Cloth probability, y ' are true probability distribution (i.e. the one-hot of Label are encoded), are distributed come judgment models to true probability with it
The order of accuarcy of estimation.
It is exactly reversed that stochastic gradient descent (Stochastic Gradient Descent, SGD), which is applied in neural network,
Propagation algorithm, optimizes loss function using common stochastic gradient descent optimization algorithm, and gradient descent method is exactly to utilize negative ladder
Direction is spent to determine the new direction of search of each iteration, and each iteration is enabled to gradually reduce objective function to be optimized,
By known input value (image) and true output valve (prediction probability), the most suitable weight parameter of perceptron is found out.
Deep learning is applied in variation detection as a result, is broadly divided into 3 steps: 1) sample collection, base area species are other
Such as road, building, soil, vegetation, each classification choose certain amount, different resolution, different perspectives, difference on the diagram
The atural object figure spot of bright-dark degree, is put into sample database;2) sample training, defining classification algorithmic formula and loss function, then define
Optimization algorithm, then it is iterated training, in each round iteration, undated parameter is lost to reduce, and is finally reached global optimum's ginseng
Number, in order to preferably complete task, this method uses two different network structures, i.e. Google Inception Net
V3 network structure, SegNet network structure, identify image and are divided;3) classifier is generated, by the model of training output
Parameter saves, and is loaded when to predict.
The present invention, prediction variation type of ground objects is input atural object figure spot and classifier in step S8, is counted into classifier
It calculates, exports the probability of each classification, probability soprano is the classification for changing atural object.
It should be noted that, in this document, relational terms such as first and second and the like are used merely to a reality
Body or operation are distinguished with another entity or operation, are deposited without necessarily requiring or implying between these entities or operation
In any actual relationship or order or sequence.Moreover, the terms "include", "comprise" or its any other variant are intended to
Non-exclusive inclusion, so that the process, method, article or equipment including a series of elements is not only wanted including those
Element, but also including other elements that are not explicitly listed, or further include for this process, method, article or equipment
Intrinsic element.
It although an embodiment of the present invention has been shown and described, for the ordinary skill in the art, can be with
A variety of variations, modification, replacement can be carried out to these embodiments without departing from the principles and spirit of the present invention by understanding
And modification, the scope of the present invention is defined by the appended.