TKK 12.4.2007 STRS system combining LiDAR and multiple images: Multi-scale template matching and LS- adjustment of a parametric crown model with lidar.

Slides:



Advertisements
Similar presentations
Active Appearance Models
Advertisements

UEF Traditional Christmas Seminar on Forest Inventory Full-waveform LiDAR research at the Hyytiälä site Ilkka Korpela, Docent Outline What is FWF LiDAR.
1 Autonomous Registration of LiDAR Data to Single Aerial Image Takis Kasparis Nicholas S. Shorter
Computer vision: models, learning and inference
IIIT Hyderabad Pose Invariant Palmprint Recognition Chhaya Methani and Anoop Namboodiri Centre for Visual Information Technology IIIT, Hyderabad, INDIA.
Chapter 6 Feature-based alignment Advanced Computer Vision.
Instructor: Mircea Nicolescu Lecture 13 CS 485 / 685 Computer Vision.
FOR 474: Forest Inventory Plot Level Metrics from Lidar Heights Other Plot Measures Sources of Error Readings: See Website.
Workshop on Earth Observation for Urban Planning and Management, 20 th November 2006, HK 1 Zhilin Li & Kourosh Khoshelham Dept of Land Surveying & Geo-Informatics.
Multi video camera calibration and synchronization.
A Study of Approaches for Object Recognition
Multiview stereo. Volumetric stereo Scene Volume V Input Images (Calibrated) Goal: Determine occupancy, “color” of points in V.
Automatic Image Alignment (feature-based) : Computational Photography Alexei Efros, CMU, Fall 2005 with a lot of slides stolen from Steve Seitz and.
CS 223B Assignment 1 Help Session Dan Maynes-Aminzade.
Object Recognition Using Geometric Hashing
Computer Science Department Andrés Corrada-Emmanuel and Howard Schultz Presented by Lawrence Carin from Duke University Autonomous precision error in low-
Multiple View Geometry Marc Pollefeys University of North Carolina at Chapel Hill Modified by Philippos Mordohai.
Smart Traveller with Visual Translator for OCR and Face Recognition LYU0203 FYP.
Airborne LIDAR The Technology Slides adapted from a talk given by Mike Renslow - Spencer B. Gross, Inc. Frank L.Scarpace Professor Environmental Remote.
55:148 Digital Image Processing Chapter 11 3D Vision, Geometry Topics: Basics of projective geometry Points and hyperplanes in projective space Homography.
DR. JOHANNES HEINZEL (Dipl.-Geogr.) University of Freiburg, Department of Remote Sensing and Landscape Information Systems, Freiburg, Germany Use.
EE392J Final Project, March 20, Multiple Camera Object Tracking Helmy Eltoukhy and Khaled Salama.
An overview of Lidar remote sensing of forests C. Véga French Institute of Pondicherry.
Lecture 1 Signals in the Time and Frequency Domains
Computer vision.
Final Exam Review CS485/685 Computer Vision Prof. Bebis.
Digital Terrain Modeling
EADS DS / SDC LTIS Page 1 7 th CNES/DLR Workshop on Information Extraction and Scene Understanding for Meter Resolution Image – 29/03/07 - Oberpfaffenhofen.
Automatic Registration of Color Images to 3D Geometry Computer Graphics International 2009 Yunzhen Li and Kok-Lim Low School of Computing National University.
Mapping forest plots: An efficient method combining photogrammetry and field triangulation/trilateration MARV1 June 2007.
Shape Matching for Model Alignment 3D Scan Matching and Registration, Part I ICCV 2005 Short Course Michael Kazhdan Johns Hopkins University.
3D SLAM for Omni-directional Camera
Mapping of Understory Lichens with Airborne Discrete-Return LiDAR Data Ilkka Korpela Dept. For. Res. Management, Helsinki Faculty of Forest Sciences, Joensuu.
Automated Face Detection Peter Brende David Black-Schaffer Veni Bourakov.
Generalized Hough Transform
__________. Introduction Importance – Wildlife Habitat – Nutrient Cycling – Long-Term Carbon Storage – Key Indicator for Biodiversity Minimum Stocking.
SINGLE-TREE FOREST INVENTORY USING LIDAR AND AERIAL IMAGES FOR 3D TREETOP POSITIONING, SPECIES RECOGNITION, HEIGHT AND CROWN WIDTH ESTIMATION Ilkka Korpela.
RASTERTIN. What is LiDAR? LiDAR = Light Detection And Ranging Active form of remote sensing measuring distance to target surfaces using narrow beams of.
DN Ordinate Length DN Difference Estimating forest structure in tropical forested sites.
Senior Project Poster Day 2005, CIS Dept. University of Pennsylvania Surface Reconstruction from Feature Based Stereo Mickey Mouse, Donald Duck Faculty.
Raquel A. Romano 1 Scientific Computing Seminar May 12, 2004 Projective Geometry for Computer Vision Projective Geometry for Computer Vision Raquel A.
1) Single-Tree Remote Sensing with LiDAR and Multiple Aerial Images 2A) Mapping forest plots: A new method combining photogrammetry and field triangulation.
LiDAR Remote Sensing of Forest Vegetation Ryan Anderson, Bruce Cook, and Paul Bolstad University of Minnesota.
Jack Pinches INFO410 & INFO350 S INFORMATION SCIENCE Computer Vision I.
Learning to Detect Faces A Large-Scale Application of Machine Learning (This material is not in the text: for further information see the paper by P.
Bundle Adjustment A Modern Synthesis Bill Triggs, Philip McLauchlan, Richard Hartley and Andrew Fitzgibbon Presentation by Marios Xanthidis 5 th of No.
Features, Feature descriptors, Matching Jana Kosecka George Mason University.
Remote Sensing of Forest Structure Van R. Kane College of Forest Resources.
SGM as an Affordable Alternative to LiDAR
Image-Based Rendering Geometry and light interaction may be difficult and expensive to model –Think of how hard radiosity is –Imagine the complexity of.
MMVAR-Kollokvio Multi-scale template matching and LS- adjustment of a parametric crown model with lidar data in 3D tree top positioning and estimation.
Citation: Kato, A.., L. M. Moskal., P. Schiess, M. Swanson, D. Calhoun and W. Stuetzel, LiDAR based tree crown surface reconstruction. Factsheet.
U NIVERSITY OF J OENSUU F ACULTY OF F ORESTRY Introduction to Lidar and Airborne Laser Scanning Petteri Packalén Kärkihankkeen ”Multi-scale Geospatial.
Puulajeittainen estimointi ja ei-parametriset menetelmät Multi-scale Geospatial Analysis of Forest Ecosystems Tahko Petteri Packalén Faculty.
Integrating LiDAR Intensity and Elevation Data for Terrain Characterization in a Forested Area Cheng Wang and Nancy F. Glenn IEEE GEOSCIENCE AND REMOTE.
Yun, Hyuk Jin. Theory A.Nonuniformity Model where at location x, v is the measured signal, u is the true signal emitted by the tissue, is an unknown.
Active Flattening of Curved Document Images via Two Structured Beams
Factsheet # 27 Canopy Structure From Aerial and Terrestrial LiDAR
Mapping forest plots: An efficient method combining photogrammetry and field triangulation/trilateration MMVAR-Colloquim May 4, 2007 Korpela, Tuomola and.
Automatically Collect Ground Control Points from Online Aerial Maps
Semi-Global Matching with self-adjusting penalties
José Manuel Iñesta José Martínez Sotoca Mateo Buendía
Nearest-neighbor matching to feature database
Motion and Optical Flow
Multi-footprint Airborne LiDAR Data in Forest Vegetation
Structure from motion Input: Output: (Tomasi and Kanade)
Nearest-neighbor matching to feature database
Aim of the project Take your image Submit it to the search engine
--- Range Image Registration
Structure from motion Input: Output: (Tomasi and Kanade)
Presentation transcript:

TKK STRS system combining LiDAR and multiple images: Multi-scale template matching and LS- adjustment of a parametric crown model with lidar data in 3D tree top positioning and estimation of the crown shape Species recognition in aerial images Ilkka Korpela (Morten Larsen)

Contents (Demos included?) Single-tree remote sensing (STRS) Photogrammetric 3D reconstruction and object recognition Airborne lidar-based 3D reconstruction and object reconstruction Coupling allometric constraints to the STRS problems 3D treetop positioning with template matching (TM) “ multi-scale TM Species recognition LS-adjustment of crown models with lidar points Conclusions and outlook

STRS Single-Tree Remote Sensing Air- or spaceborne; active and/or passive sensing 2D or 3D = with or without tree height “Direct estimation” of tree/crown position and species; indirect model-based estimation of height, dbh Restrictions: tree discernibility due to scale (detectable object size), occlusion and shading. Alternative or complement to A) field inventory, B) area-based remote sensing

STRS Single-Tree Remote Sensing Accuracy restricted by “allometric noise” alike in volume functions → tree and stand level bias, tree level impresicion. dbh ~ %. Measurements subject to bias: DTM-errors, lidar does not hit apexes, Dcr underestimation Nothing can be known about quality, only quantity Unsolved issues: species recognition, regeneration stands, calibration and validation in the field, aggregated crowns result in fused trees.

Photogrammetric 3D reconstruction and object recognition Usually from N>1 images (multiscopic) Correspondence problem - ill-posed, perspective errors, reflectance, occlusion, scene complexity Texture in the image functions needed Relies on accurate geometry (camera interior, exterior), ray- intersection

Photogrammetric 3D reconstruction and object recognition Digital revolution (2000→) Aerial images upto 2GB, manageable (I/O, analysis, storage, transfer) Automatic methods in orientation, incl. DSO with GPS/INS, reduction in GCPs. Digital cameras with MS images 2005→ multiple images per target, better radiometry and geometry Automatic DSM production

Photogrammetric 3D reconstruction and object recognition Automation Laborious orientation tasks ± solved DSMs, DTMs using image matching e.g. Building extraction - semiautomatic 3913, 3914, 3915 (triplet matching) → 1946, 1962 Demo 1

Photogrammetric 3D reconstruction and object recognition Photogrammetric STRS scene and object complexity occlusion & shading scale: h = m, Dcr m “BDRF”-effects → automation challenging

Photogrammetric 3D reconstruction and object recognition Demo – manual STRS - NLS (19) treetop 3D, height, Dcr, Sp dbh = f(Sp, h, Dcr) + epsilon Image matching fails for treetop positioning unless we use a feature detector for treetops.

Airborne lidar-based 3D reconstruction and object reconstruction A pulse of short duration (~ 3 m, 1064 nm) Observing returned signal. Discrete data. Upto 128 samples?. Signal is reconstructed into points or samples for later waveform analysis. Intensity of the return/echo. Pros: No texture needed, active → no shading, “real ease of 3D” Cos: discrete sampling, high sampling rates costly, difficult to reconstruct “high- frequency relief”.

Airborne lidar-based 3D reconstruction and object reconstruction Automation DTMs – manual assistance needed – high accuracy – even under canopy Volume estimation of trees – “automatic” – e.g. using regression between lidar features and field observations (e.g. Naesset Suvanto et al. 2005) Lidar and STRS Algorithms that process point clouds directly or interpolated DSMs (CHMs) Underestimation of heights (footprint size, density, crown shape, equipment sensitivity) Species not obtained (so far)

Coupling allometric constraints to the STRS tasks If we know dbh, we have an idea of the height If we know dbh, species and age, we have better idea of height If we know dbh, species, and height we have a good idea of volume If we know dbh, species, height and height of crown, we have a better idea of volume. If we know height, species and crown width we can estimate dbh and volume If we know species we have an idea of the shape of the “crown envelope”. If we know species and height, can we set limits for the variation of crown width f(sp, h) => [Min, Max] of Dcrm, and assume a basic shape?

Coupling allometric constraints to the STRS tasks If we know species can we have an idea of the shape of the “crown envelope”?  Timo Melkas

Dcrm (min,max) / Shape | (Sp, height) → - Consistency of measurements (rule out impossible observations) - Initial approximations for iterative approaches in finding true Dcrm & crown shape E.g Short trees have small crowns (adjust search space accordingly, or look for small crowns from a low height) Coupling allometric constraints to the STRS tasks

3D treetop positioning with template matching (TM) Demo D treetop positioning using TM 1) Use, for each of the N views, a model image (template) of a crown. 2) Compute N normalized cross- correlation images (template matching). 3) Form a Cartesian 3D grid in the canopy – in the search space. 4) Aggregate 3D correlation to the grid points. 5) Process the 3D correlation into “hot-spots” – 3D treetop positions. Fine, but not invariant to object variation.

Cartesian grid / Search space in the upper canopy

Correlation > threshold

Multi-scale TM – Treetop positioning Can we assume that the optical properties and the “relative shape” of trees are invariant to their size? I.e. small trees appear as scaled versions of large trees in the images? (inside one species and within a restricted area)

Multi-scale TM – Treetop positioning Maxima at different scales, take global → (X,Y,Z)

Multi-scale TM – Crown size Demo 04403_19, 18, 20, 06214_3900, 3901 Near-nadir views are best for manual measurement of Dcr (crown width)

Species recognition Spectral values Texture Variation: - Phenology - Tree age and vigour - Image-object-sun geometry => reliable automation problematic => bottleneck

LS-adjustment of a crown model with lidar points Assume that 1) Photogrammetric Multi-scale TM 3D treetop position is highly accurate 2) Trees have only moderate slant 3) Crowns are ± rotation symmetric 4) We know tree height and species which give a reasonable approximaion of the crown size and shape → LiDAR hits are “observations of crown radius at a certain height below the apex” Assume a rather large crown and collect lidar hits in the visinity of the 3D treetop position, down to relative height of apprx. 60 %. Use LS-adjustment to find best set of parameters for the crown model.

“LiDAR hits are observations of crown radius at a certain height below the apex?”

LiDAR hits are observations of crown radius at a certain height below the apex – what if crowns are interlaced?”

Example - a 19-m high spruce: Solution in three iterations. Final RMSE 0.31 m Note apex! LiDAR did not hit the apex and the “crown width at treetop” (constant term) is negative.

Example - a 22-m high birch: Solution in six iterations. Final RMSE 0.47 m For some reason RMSEs are larger for birch in comparison to pine and spruce. Convergence?

Conclusions and outlook A - Multi-scale TM works in a manual  semi-automatic way for treetop positioning Possible to automate? Computation costs? (NCC now tried everywhere) - Multi-scale TM in crown width estimation needs comprehensive testing (Image scales, required overlaps) -Species recognition was “overlooked” here, still I think that good 3D treetop positions can be used for the purpose. - Matching LiDAR points using LS-adjustment works only if the exact treetop position is known. Aggregated crowns are problematic, but these cases are known from the tree map

Conclusions and outlook B - If we have a system that can be operated so that a tree measurement takes 2-3 seconds and the measurement inaccuracies are: h ~ 0.6 m Dcr ~ 10% d13 ~ % XY-position ~ 0.3 m Sp ~ 95% Is this fast and accurate enough for sample-plot based STRS? Can we afford the images and LiDAR?

Calibration and validation of results?

Presenting method and early results in ISPRS workshop in Hannover, June 1, 2007 Presentation at Silvalaser, Espoo September 2007? Article about Multi-Scale TM in 3D treetop positioning (ISPRS JPRS?) Article about Multi-Scale TM combined with crown modeling using LS-adjustment and allometric constraints (Silva Fennica 100-yr issue?) Species recognition study remains of the Academy 3-yr resrach project to be completed before VIII/2008