Multi-view Dense Match for Forest Area

LIDAR(Light Detection And Ranging) is widely used in forestry applications to obtain information about tree density, composition, change, etc. An advantage of LIDAR is its ability to get this information in a 3D structure. However, the density of LIDAR data is low, the acquisition of LIDAR data is often very expensive, and it is difficult to be utilised in small areas. In this article we eavluate different methods by using multi-view to acquire high resolution images of the forest. Using the dense match method a dense point cloud can be generated. Our analysis shows that this method can provide a good alternative to using LIDAR in situations such as these. ISPRS Technical Commission I Symposium, Sustaining Land Imaging: UAVs to Satellites 17 – 20 November 2014, Denver, Colorado, USA, MTSTC1-154


INTRODUCTION
Using point data one can extract an object's 3D information and structure.Point matching means matching corresponding points between two or more images of the same scene and this is an important feature of many computer vision and pattern recognition tasks, including object recognition and tracking and 3D scene reconstruction.This makes point data a very important source for data mapping purposes.Because of the importance of point cloud data for many applications LIDAR data is widely used in many projects.As Bartels and Wei comment "is an important modality in terrain and land surveying for many environmental, engineering and civil applications" (Bartels and Wei, 2010).However, when using LIDAR to extract the parameters and characteristics of forest areas there are a number of problems with the approach namely: data resolution, cost, and data processing requirements.it is a good test by using multi-view match to generate dense match cloud point.
The European spatial data research organization started a benchmark on image based DSM generation in feburay 2013.This test is based on two representative image blocks, which were processed by different groups with different software systems.
The interpretation of the benchmark results identified some scenarios that still can cause some problems during image based surface reconstruction.Some solutions showed decreasing accuracies as shadows were cast.In addition, the representative image blocks don't include the forest area (Haala, 2013).
Dense matching methods can be divided into two types: depth map fusion and object patch.One technique used to reduce the search area of matching processes in aerial images is MVLL (Multi-View Local Learning), which defines a vertical line in the object space and calculates the correlation coefficients of the two image matrices according to a point along this line (Zhang, 2005;Waser, 2008;Baltsavias, 2008;Jiang, 2004;Ming, 2009).
Dynamic programming is a method for efficiently solving optimization problems by caching subproblem solutions rather than recomputing them again, dynamic programming was a subglobal matching strategy where one 1D constraint is considered along image rows, the use of those row-wise 1D constraints results in depth maps that show a "streaking effect"(H.Baltsavias and Waser developed a new image matching software package.They demonstrated its application in 3D tree modelling by comparing this to data obtained by the airborne laser.It showed that photogrammetric DSM (Digital Surface Models) can be denser than a DSM generated by LIDAR.Leberl compared point clouds from aerial and street-side LIDAR systems with those created from images.They show that the photogrammetric accuracy compares very well with the LIDAR method.However the key advantage of the photogrammetric approach is that the density of surface points is much higher from the images than from the LIDAR method.The authors conclude that "throughput is commensurate with a fully automated all-digital approach''.When image capture has been completed the next step is to manage and process the collection of images.
In this article, we try to evaluate the dense match algorithm in forest area.section 2 show the algorithm of the dense match methods, section 3 shows the test area and results, sections 4 gives the conclusion.

ALGORITHM
The Under forest area, it is hard to set control point in the forest.So we get the image orientation parameter by using GPSsupported bundle adjustment.By using multi-view dense match algorithm, we generate dense matched cloud point.
We use SIFT algorithm to generate connection point.
We use calibrated bundle adjustment to calculate the accurate orientation and camera distortion.
as to the dense match method, I use three kind of methods, they are PMVS, Pix4d and SURE.PMVS is a patch based method.this method is implemented as a match, expand, and filter procedure, starting from a sparse set of matched keypoints, and repeatedly expanding these to nearby pixel correspondences before using visibility constraints to filter away false matches (Zhu, 2009; Furukawa, 2008; Furukawa and Jean, 2009; Furukawa et al., 2009b; Shi, Guo and Hu, 2011).
Pix4d is photogrammetry software for UAV.We use SURE algorithm to generate the dense matched cloud point.There are three main modules in the algorithm.image rectification module is to generate epipolar images for the matching process.then it is dense stereo matching module.structure computation module is the algorithm for 3D object point triangulation.
The SGM algorithm aims to estimate disparities across stereo pairs such that the global cost function

Study area
The test site is located at Zhangye, Gansu Province in Western China (38• 32_ N, 100•15_ E).The elevation of the test site is about 2800 m.The forest is a pure spruce (Piceacrassifolia) stand.The forest floor is covered by moss.Dead branches remain on thelower trunks of some trees.A 100 m × 100 m field plot in relatively flat terrain wasselected for data collection.There were more than 1400 trees with diameter at breastheight (DBH) greater than 2.5 cm within the plot.Range images used in this study wereacquired on 8 June 2008.

LiDAR and aerial imagery
The LiDAR data and the aerial imagery were collected May, 2008, using a LiteMapper 5600 airborne laser scanner and DigiCAM-H/22 (Digital Camera System 22 megapixels), respectively.The LiDAR was operated at a nominal altitude of 2800 m above ground level and recorded the first returns as well as the return intensity in a single pass.A 80% overlap between adjacent strips ensured that density is enough in the surveyed area.The maximum scan angles were ±30° off nadir and the average sampling space is about 1.6 m (0.43/m2 for whole area, 0.57/m2 for broadleaf tress and 0.65/m2 for conifer trees).The nominal accuracy of horizontal (x, y) and vertical (z) is about 0.5 and 0.2 m, respectively.The DigiCAM is a charge coupled device (CCD) camera with 22 mega pixels and each pixel is 9 μ m in size.The DigiCAM imagery has a 50 cm spatial resolution with a 80% overlap along the flying direction and a 80% overlap across flight lines.

Field data
The study area is called super area, because area is covered with all kinds of investigation methods.Super study area is 100m plus 100m, the study area is divided into 16 small area, 25m plus 25m.Ground control point were measured by differential GPS.Reference data for tree heights were measured manullay by total station.Terrestrial laser scanner used in this study.Figure 2 shows the image of the whole area.Figure 3 shows the lidar data of test area.m .In the test area, there are many control point in the test area, so we use them to check the accuracy of the dem of the lidar, we use the height of the trees to check the accuracy of the lidar and matched point cloud.Figure 4 shows the DSM generated by the PMVS. Figure 5 shows cloud point generated by Pix4d.

CONCLUSION
In this paper we have evaluate three methods for dense point cloud extraction from multi-view imagery for the purposes of forestry analysis.Following from the experimental detail outlined in the previous section we show the point clouds generated for the forest area in Figure 4 in both Figure 5 and Figure 6. Figure 5 shows the point cloud for the entire area while Figure 6 shows the point cloud for one of the local areas with the test mosaic.both methods can generate dense matched point cloud than lidar.SURE's result present the whole structure of the forest area.
The accuracy of DSM should be improved for extraction of tree parameter.I think the dense match algorithm and fusion algorithm are two important

Figure 1
Figure 1 workflow of the processing

. The penalty parameters 1 P
Thereby D represents the disparity image holding disparity estimates of all base image pixels x b .T is an operator evaluating to one if the subsequent condition is true and evaluates to zero else.x N denote base image pixels in the neighborhood of x b .The global cost function E is composed of a data term and two terms claiming for smooth surfaces.The data term is computed by pixel-wise similarity measures (x , x ) bm C and 2 P control the gain of surface smoothing(Rothermel et al., 2011; Rothermel et al., 2012; Wenzel et al., 2013).

Figure 2 .
Figure 2. mosaic image of the whole area
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-1, 2014 ISPRS Technical Commission I Symposium, 17 -20 November 2014, Denver, Colorado, USA