Progress 03/01/24 to 02/28/25
Outputs
Target Audience:Members of the Forest Modeling Research Cooperative (FMRC) based at Virginia Tech have been regularly updated on progress of this project. The FMRC is an industrial research cooperative and the methods and results from this project may be widely implemented across many thousands of acres of industrial timberlands. Research talks were given on this material to faculty of the General Robotics Automation Sensing and Perception Laboratory, undergraduate, Master's and doctoral students students at Penn, lab tours and visits. Ankit Prabhu, who is a Research Assistant at the University of Pennsylvania, gave a research seminar at the annual FMRC meeting where he interacted with stakeholders in the forestry industry on how to incoporate robotics technology. PhD student Xu Liu, Ankit Prabhu and Jizhou Li (Master's student at Penn), made regular visits to an apple orchard near Philadelphia to collect data on how the appearance and size of apples changes over time as they grow (early spring through the summer). These efforts have been crucial to teach us the critical problems to focus upon, e.g., mapping unstructured trees and plants. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?This project is currently training the following undergraduate, Master's and doctoral students. 1. Xu Liu (Mechanical Engineering PhD, just graduated, postdoc at Stanford now) 2. Dexter Ong (Computer Science PhD, 3rd year) 3. Ankit Prabhu (Resarch Assistant) 5. Siming He (Computer Science undergrad senior), 6. Fernando Caldera (6th year PhD), 7. Derek Chang (2nd year Masters, just graduated), and 8. Zach Osman (2nd year Master's student). A 1st yearundergraduate student, Yazlin Moujalled from the University of Maryland Baltimore County,was also involved in this research as a part of her REU. She worked on developing a system for identifying different species of trees, plants, weeds using large vision-language models. These students have engaged in research (which includes field experiments in the Virginia and Wharton State forests), gathering and processing data from the hardware platform as well as building the hardware platform for hand-held mapping of trees and autonomous flight. A number of simulation experiments were conducted before these efforts to fine-tune our methods and explore new algorithms. In addition to this, these students have also taken part in weekly group meetings (joint w/ groups of Vijay Kumar, Pratik Chaudhari and Virginia Tech PI Patrick Corey Green). How have the results been disseminated to communities of interest?These results were presented at conferences (ICRA, ACC, CVPR) over Summer 2024, some others will be presented at ICRA 2025. Some of the team members are also organizing a workshp on Robotics for Agricultureand Forestry at ICRA 2025. A major journal paper is being written up on these results. Excerpts from these results were presented during lab tours and visits, e.g., to local high-school, to staff and administrative personnel from Ivy league colleges during their annual visit to Penn. What do you plan to do during the next reporting period to accomplish the goals?The key things that we will focus on in the coming year are: Finish our analysis on the collected forestry data to build a system that can give photorealistic maps of up to 2 sq. km areas, it allows a user to click on any point in this region and visualize the geometry of the tree, obtain statistics such as (volume of wood and species of trees). This system will also be able to summarize these statistics over large regions. We expect to demonstrate this system and open it up to the general public this year. We have built a new drone with better cameras and an onbard GPU. This drone can give us more detailed observations higher up in the forest canopy because we can fly closer to dense branches. Over the last year, we have developed the ability to do experiments with ground and aerial vehicles together. We are currently instrumenting a legged robot that can traverse forest terrain much more easily. This robot will be used for some of the field experiments. We will refine our techniques to model the changes in appearance of plants/trees over time. We have seen some promising results on apple orchard data. We will now translate these results to trees in a forest across seasons.
Impacts
What was accomplished under these goals?
Large Scale Mapping (Thrust A) Forest Mapping Current forestry measurements use low-data-rate optical sensors (small LiDARs) and heavy manual involvement (e.g., walking across large areas). We have shown that 360 cameras offer a much cheaper alternative, providing large amounts of rich, visual data. Over the past year, we collected large datasets from a research forest at Virginia Tech (where we have ground truth data from collaborators) and outdoor parks near Philadelphia. We gathered both 3D LiDAR (Ouster) data and 360-degree 4K RGB camera data from a manually flown drone and human-carried sensor rig. We processed large parts of this data into point clouds (representing tree trunks, branches, leaves, etc.) using both LiDAR and RGB data. We developed new techniques for synthesizing new views, semantic segmentation, and built a pipeline to extract individual trees. We combined this data with inertial sensors (accelerometers and gyroscopes) to compute statistics like diameter at breast height and compare it to LiDAR-based and human-collected measurements. Our results suggest we can recover tree diameter with about 15% error using only RGB data. These techniques could shift how forestry data is collected. From NeRFs to Gaussian Splats, and Back In robotics, with limited (typically ego-centric) views, parametric representations like neural radiance fields (NeRFs) generalize better than non-parametric ones like Gaussian splatting (GS) for views different from the training data. However, GS renders much faster than NeRFs. We developed a procedure to convert between the two, achieving the benefits of both: NeRFs provide superior PSNR, SSIM, and LPIPS on dissimilar views, and GS offers real-time rendering with easy representation modification. The computational cost of these conversions is minimal compared to training both from scratch. 4D Metric-Semantic Mapping for Persistent Orchard Monitoring: Method and Dataset Automated, persistent monitoring of orchards at the tree or fruit level maximizes crop yield and optimizes resources like water, fertilizers, and pesticides. We present a 4D spatio-temporal metric-semantic mapping method that fuses LiDAR, RGB, and IMU data to track fruits in an orchard over their growth season. A LiDAR-RGB fusion module segments fruits with a deep neural network and tracks them using the Hungarian Assignment algorithm. The 4D data association module aligns data from different growth stages and tracks fruits spatially and temporally, providing information like fruit counts, sizes, and positions. We demonstrate the accuracy of our method with real orchard data under natural, uncontrolled conditions with seasonal variations. We achieve 3.1% error in fruit count estimation across 1790 fruits and 60 trees, and a 1.1 cm mean error in size estimation. The datasets, including LiDAR, RGB, and IMU data for five fruit species, will be made publicly available at: https://4d-metric-semantic-mapping.org/ Active Perception (Thrust B) An Active Perception Game for Robust Information Gathering Active perception selects future viewpoints using an estimate of information gain. An inaccurate estimate can be detrimental in critical situations (e.g., locating a person in distress quickly). However, true information gain can only be calculated post hoc. We present a method for estimating the discrepancy between the predicted and actual information gain. By analyzing the relationship between active perception and estimation error in a game-theoretic setting, we developed an online approach that reduces sub-optimality and achieves sub-linear regret in estimating true information gain. Our experiments, using a range of environments, robotic platforms, and perception data, show our approach reduces information gain errors by 42%, increases information gain by 7%, improves PSNR by 5%, and enhances semantic accuracy by 6%. In real-world experiments with a Jackal ground robot, the approach demonstrates complex trajectories to explore occluded regions. Active Scout: Multi-Target Tracking Using Neural Radiance Fields in Dense Urban Environments We study pursuit-evasion games in highly occluded urban environments (e.g., tall buildings) where a scout (quadrotor) tracks multiple dynamic targets on the ground. We show that a neural radiance field (NeRF) can be built online using RGB and depth images from different vantage points, allowing for information gain calculation to track targets and explore unknown areas. We demonstrate this with a custom-built simulator using OpenStreetMaps data from Philadelphia and New York City, showing we can locate 20 stationary targets in 300 steps. For dynamic targets hiding behind occlusions, our approach maintains a tracking error of 200m, while a greedy baseline can have a 600m error. Our scout's behavior adapts over time, switching attention between targets as the NeRF representation improves. Forestry Research (Thrust C) PhD student Nitant Rai at Virginia Tech is finalizing his research plan with the following objectives for 2025-2026: Using UAS hyperspectral data and LiDAR point clouds, assess accuracy in detecting seedlings in loblolly pine forests. Investigating the robustness of predictive models for forest structure with varying LiDAR resolutions. Developing new methods to use parametric stem taper models to predict tree volumes from incomplete point cloud profiles. Accomplishments: Dissemination of research findings in the Forest Modeling Research Cooperative 2024 Annual Report. Presentation of research findings at the Forest Modeling Research Cooperative 2024 Annual Meeting. Preliminary data collection for seedling detection in the Appomattox-Buckingham State Forest. Initial development of methods to assess the robustness of forest structure metrics using existing point clouds and plot measurements.
Publications
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2024
Citation:
He, Siming, Zach Osman, and Pratik Chaudhari. "From NeRFs to Gaussian Splats, and Back." CVPR Workshop arXiv:2405.09717 (2024).
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2024
Citation:
He, Siming, Yuezhan Tao, Igor Spasojevic, Vijay Kumar, and Pratik Chaudhari. "An Active Perception Game for Robust Autonomous Exploration." ICRA 2025.
- Type:
Conference Papers and Presentations
Status:
Awaiting Publication
Year Published:
2024
Citation:
Lei, Jiuzhou, Ankit Prabhu, Xu Liu, Fernando Cladera, Mehrad Mortazavi, Reza Ehsani, Pratik Chaudhari, and Vijay Kumar. "4D Metric-Semantic Mapping for Persistent Orchard Monitoring: Method and Dataset." arXiv preprint arXiv:2409.19786 (2024).
|
Progress 03/01/23 to 02/29/24
Outputs
Target Audience:Members of the Forest Modeling Research Cooperative (FMRC) based at Virginia Tech have been regularly updated on progress of this project. The FMRC is an industrial research cooperative and the methods and results from this project may be widely implemented across many thousands of acres of industrial timberlands. Research talks were given on this material to faculty of the General Robotics Automation Sensing and Perception Laboratory, undergraduate, Master's and doctoral students students at Penn, lab tours and visits. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?This project is currently training the following undergraduate, Master's and doctoral students. 1. Xu Liu (Mechanical Engineering PhD, 5th year) 2. Dexter Ong (Computer Science PhD, 2nd year) 3. Yifei Shao (Computer Science PhD, 2nd year) 4. Ankit Prabhu (Resarch Assistant) 5. Siming He (Computer Science undergrad senior), 6. Fernando Caldera (5th year PhD), 7. Derek Chang (2nd year Masters). Some undergaduate students were also invovled in this research in Summer 2023, Alicia Sun (1st year), Alan Zhu (1st year). As a part of this project, these students have engaged in research (which includes field experiments in the Virginia and Wharton State forests), gathering and processing data from the hardware platform as well as building the hardware platform for hand-held mapping of trees and autonomous flight. A number of simulation experiments were conducted before these efforts to fine-tune our methods and explore new algorithms. In addition to this, these students have also taken part in weekly group meetings (joint w/ groups of Vijay Kumar, Pratik Chaudhari and Virginia Tech PI Patrick Corey Green). How have the results been disseminated to communities of interest?These results will be published and presented at conferences (ICRA and ACC) over Summer 2024. Two journal papers (one published, and one in progress) have resulted from this work. Excerpts from these results were presented during lab tours and visits, e.g., to local high-school, to staff and administrative personnel from Ivy league colleges during their annual visit to Penn. What do you plan to do during the next reporting period to accomplish the goals?In the next year, we will focus on the following action items. - Use high resolution visual data (2x fisheye cameras that together record 8K frames at 30 fps) to build photorealistic maps of trees. Compare these maps to catalogues maintained by the Institute Architect at Penn to undertand the different kinds of trees on campus. - Instrument a UAV with this visual system and collect data using manual and autonomous flights in New Jersey and Virginia forests. - Demonstrate multi-UAV flight and communications in a forest. - Develop a system that can recover fine-grained details of the trees, branching patterns and shrubs in large forested areas using a combination of neural radiance fields and featues of foundation models such as CLIP that are grounded in the physical scene. - Develop procedures to model changes in the appearance of trees across seasons and utilize this information for localization.
Impacts
What was accomplished under these goals?
We have made progress on the three main thrusts in the following directions. UAVs for forestry: Metric-semantic mapping and diameter estimation with autonomous aerial robots (Thrust A, C) To properly monitor the growth of forests and administer effective methods for their cultivation, forestry researchers require access to quantitative metrics such as diameter at breast height and stem taper profile of trees. These metrics are tedious and labor-intensive to measure by hand, especially at the scale of vast forests with thick undergrowth. Autonomous mobile robots can help to scale up such operations and provide an efficient method to capture the data. We present a set of algorithms for autonomous navigation and fine-grained metric-semantic mapping with a team of aerial robots in under-canopy forest environments. Our autonomous UAV system has 3D flight capabilities and relies only on a LIDAR and an IMU for state estimation and mapping. This allows each robot to accurately navigate in challenging forest environments with drastic terrain changes regardless of illumination conditions. Our deep-learning-driven fine-grained metric-semantic mapping module is capable of detecting and extracting detailed information such as the position, orientation, and stem taper profile of trees. This map of tree trunks is represented as a set of sparse cylinder models. Our semantic place recognition module leverages this sparse representation to efficiently estimate the relative transformation between multiple robots, and merge their information to build a globally consistent large-scale map. This ultimately allows us to scale up operations with multiple robots. Our system is able to achieve a mean absolute error of 1.45 cm for diameter estimation and 13.2 cm for relative position estimation between a pair of robots after place recognition and map merging TreeScope: An Agricultural Robotics Dataset for LiDAR-Based Mapping of Trees in Forests and Orchards (Thrust C) Data collection for forestry, timber, and agriculture currently relies on manual techniques which are labor-intensive and time-consuming. We seek to demonstrate that robotics offers improvements over these techniques and accelerate agricultural research, beginning with semantic segmentation and diameter estimation of trees in forests and orchards. We present TreeScope v1.0, the first robotics dataset for precision agriculture and forestry addressing the counting and mapping of trees in forestry and orchards. TreeScope provides LiDAR data from agricultural environments collected with robotics platforms, such as UAV and mobile robot platforms carried by vehicles and human operators. In the first release of this dataset, we provide ground-truth data with over 1,800 manually annotated semantic labels for tree stems and field-measured tree diameters. We share benchmark scripts for these tasks that researchers may use to evaluate the accuracy of their algorithms. Finally, we run our open-source diameter estimation and off-the-shelf semantic segmentation algorithms and share our baseline results. VEMS-SLAM: Versatile, Efficient, Metric-Semantic SLAM for Multi-Robot Navigation and Exploration (Thrust A) In this paper, we propose a generic and versatile metric-semantic SLAM (MS-SLAM) framework for autonomous exploration with heterogeneous robot teams. The proposed MS-SLAM framework supports different types of sensors including LiDARs and RGBD cameras.It enables a heterogeneous team of robots to autonomously explore 3D environments featuring both indoor and outdoor areas without relying on GPS.The MS-SLAM framework utilizes a hierarchical metric-semantic representation of the environment, ranging from high-level sparse semantic maps to low-level dense geometric maps. For the purpose of facilitating autonomous exploration and navigation in large-scale environments, we proposed using a high-level map representation that is sparse and semantically meaningful, consisting of explicitly modeled object landmarks. Building upon such a representation, we use our customized measurement models for each type of semantic objects and incorporate them into a factor graph, so that the robot can efficiently optimize over the semantic map and its poses during the entire mission.We leverage the sparsity, informativeness, and viewpoint invariance of this representation and propose a direct but effective semantic place recognition algorithm to detect both intra-robot and inter-robot loop closures and correct odometry drifts and merge maps from multiple robots.We design a multi-robot collaboration framework that keeps track of the robot's own metric-semantic observations and the observations from other robots within communication range.This framework also incorporates the proposed place recognition algorithm, so that the observations from the other robots can be merged to construct a common metric-semantic factor graph whenever a valid loop closure is detected.Due to the sparsity and informativeness of the high-level semantic map, our system runs in real time on each robot in a decentralized manner and opportunistically leverages communication.We integrate and deploy our proposed MS-SLAM framework on three types of aerial and ground robots that utilize either RGBD cameras or LiDARs, and demonstrate that the team of robots can explore a 3D large-scale indoor and outdoor environments. We show, through both theoretical analysis and empirical evidence, that our proposed MS-SLAM framework has low computational and memory demand. % even with large teams of robots autonomously exploring large-scale environments. In addition, we benchmark our system against state-of-the-art metric-semantic SLAM algorithms. Active Perception using Neural Radiance Fields (Thrust B) We study active perception from first principles to argue that an autonomous agent performing active perception should maximize the mutual information that past observations posses about future ones. Doing so requires (a) a representation of the scene that summarizes past observations and the ability to update this representation to incorporate new observations (state estimation and mapping), (b) the ability to synthesize new observations of the scene (a generative model), and (c) the ability to select control trajectories that maximize predictive information (planning). This motivates a neural radiance field (NeRF)-like representation which captures photometric, geometric and semantic properties of the scene grounded. This representation is well-suited to synthesizing new observations from different viewpoints. And thereby, a sampling-based planner can be used to calculate the predictive information from synthetic observations along dynamically-feasible trajectories. We use active perception for exploring cluttered indoor environments and employ a notion of semantic uncertainty to check for the successful completion of an exploration task. We demonstrate these ideas via simulation in realistic 3D indoor environments.
Publications
- Type:
Journal Articles
Status:
Published
Year Published:
2024
Citation:
Prabhu, Ankit, Xu Liu, Igor Spasojevic, Yuwei Wu, Yifei Shao, Dexter Ong, Jiuzhou Lei, Patrick Corey Green, Pratik Chaudhari, and Vijay Kumar. "UAVs for forestry: Metric-semantic mapping and diameter estimation with autonomous aerial robots." Mechanical Systems and Signal Processing 208 (2024): 111050.
- Type:
Conference Papers and Presentations
Status:
Accepted
Year Published:
2024
Citation:
He, Siming, Christopher D. Hsu, Dexter Ong, Yifei Simon Shao, and Pratik Chaudhari. "Active Perception using Neural Radiance Fields." American Control Conference, 2024.
- Type:
Conference Papers and Presentations
Status:
Accepted
Year Published:
2024
Citation:
Cheng, Derek, Fernando Cladera Ojeda, Ankit Prabhu, Xu Liu, Alan Zhu, Patrick Corey Green, Reza Ehsani, Pratik Chaudhari, and Vijay Kumar. "TreeScope: An Agricultural Robotics Dataset for LiDAR-Based Mapping of Trees in Forests and Orchards." ICRA 2024
- Type:
Conference Papers and Presentations
Status:
Accepted
Year Published:
2024
Citation:
Shao, Yifei Simon, Yuwei Wu, Laura Jarin-Lipschitz, Pratik Chaudhari, and Vijay Kumar. "Design and Evaluation of Motion Planners for Quadrotors." ICRA 2024.
|
Progress 03/01/22 to 02/28/23
Outputs
Target Audience:Members of the Forest Modeling Research Cooperative (FMRC) based at Virginia Tech have been updated on project progress. The FMRC is an industrial research cooperative and the methods and results from this project may be widely implemented across many thousands of acres of industrial timberlands. Research talks were given on this material to faculty of the General Robotics Automation Sensing and Perception Laboratory, undergraduate, Master's and doctoral students students at Penn, lab tours and visits, e.g., to 28 students from a local high-school. Changes/Problems:Patrick Corey Green assumed role of the Principle Investigator at Virginia Tech. This was approved from the prime awardee University of Pennsylvania and the Program Manager at USDA/NIFA. What opportunities for training and professional development has the project provided?This project is currently training the following undergraduate, Master's and doctoral students. Xu Liu (Mechanical Engineering PhD, 4th year) Dexter Ong (Computer Science PhD, 1st year) Yifei Shao (Computer Science PhD, 1st year) Ankit Prabhu (Robotics Master's 2nd year) Yingtian Tang (Computer Science Master's, 2nd year) Daiwei Chen (Electrical Engineering Master's, 2nd year) Siming He (Computer Science undergrad) Alan Zhao (Computer Science undergrad) As a part of this project, these students have engaged in research (which includes field experiments in the Virginia and Wharton State forests), gathering and processing data from the hardware platform as well as building the hardware platform for hand-held mapping of trees and autonomous flight. A number of simulation experiments were conducted before these efforts to fine-tune our methods and explore new algorithms. In addition to this, these students have also taken part in weekly group meetings (joint w/ groups of Vijay Kumar, Pratik Chaudhari and Virginia Tech PI Patrick Corey Green) How have the results been disseminated to communities of interest?Excerpts from these results were presented during lab tours and visits, e.g., to 28 students from a local high-school, to staff and administrative personnel from Ivy league colleges during their annual visit to Penn. What do you plan to do during the next reporting period to accomplish the goals?We do not anticipate any changes to the intellectual agenda discussed in the original proposal. In the next year, we will focus on the following action items. Improve the performance of LiDAR odometry to enable robust autonomous flight in forests over long time horizons. Build solutions to combine the data obtained from multiple drones flying simultaneously to map a large part of the forest; this involves research on scene understanding, active exploration and bundle adjustment. Develop a pipeline for species identification that can work for the vegetation expected in the Virginia State Forests and develop deep learning-based object detection that can accurately identify the species of the tree using multiple images of the leaves and tree bark across a broad range of lighting condtions. This also includes building a pipeline for annotation of such data. Apply the methodology for mapping and estimation of volume broadly across planted loblolly pine forests. In addition to continued efforts at accurately characterizing the main stem, begin to assess tree crown metrics including crown ratio, crown length, crown volume, and crown area. Additionally, develop methods to quantify competing understory vegetation in commercial timberlands.
Impacts
What was accomplished under these goals?
Mapping unstructured environments and autonomous flight (Thrust A) We fine-tuned the autonomous platform to fly autonomously using LiDAR-based odometry over large time-horizons. Experiments were conducted on the Penn campus, at the Virginia State Forest and the Wharton State Forest and in our longest experiments, we were able to fly (a) non-autonomously in open areas for close to 5km with a drift in the state estimate of less than a meter, (b) autonomously in uncluttered areas with few branches, (c) non-autonomously in more cluttered areas with more branches, and (d) flying at both day and night times. These experiments and the ensuing analysis suggests that (a) our current autonomy stack can effectively handle the requirements of this project, e.g., mapping large regions with low-drift, flying close enough to large trees to gather actionable measurements, (b) the platform is stable enough to obtain precise measurements as compared to a hand-held sensor rig. We have developed a hand-held sensor rig (with a LiDAR, RGB cameras, an IMU and a GPS, all connected to an on board computer) to help with mapping. These sensors are the same as those on the autonomous platform. The sensor rig gives us the flexibility to move around the trees (by carrying it over the shoulders in a backpack) and obtain more precise and exhaustive measurements than those allowed by the autonomous platform. The estimates obtained using data from this sensor rig will be compared to those obtained from the flying platform. Active semantic mapping and perception (Thrust A) We developed algorithms to perform semantic mapping, i.e., identity different objects in the environment and categorize them as such (instances of cars, pedestrians and trees primarily). This has been evaluated extensively, e.g., near the Penn campus, at West Point on ~2 km loops with multiple objects (e.g., a row of cars parked along the road) etc. The highlights of this approach are that we can robustly detect objects and categorize them using partially occluded viewpoints, estimate the orientation (we have tested this for cars) precisely using PCA. For active perception, we have builtupon a theory in computational neuroscience called "slow feature analysis" (SFA) which defines objects as sets of features that vary slowly in time. SFA, and its variants, are about maximizing the mutual information between the representation of the input. Our formulation of active perception revolves around the idea of choosing actions that ruin this invariant of SFA, i.e., they minimize the mutual information between the representation of successive images across time while minimizing a control cost. In simulation experiments using 2D scenes, we instantiated this principle to demonstrate trajectories for agents that explore the scene and show a "curious but busy bee" behavior, i.e., the agent explores an object until it learns enough invariant features about it and moves on to the next object after this... Estimation of tree volume (Thrust B/C) We have developed a procedure to estimate the volume of tree trunks in LiDAR data. This involves A procedure to annotate point clouds corresponding to tree trunks quickly from a top-down view. A deep network that segments tree trunks in a depth image that is created from the LiDAR point cloud. This network is trained on data created from the annotation. We have developed a procedure to estimate the diameter of the tree trunk (by filling a circle to the LiDAR points) at various heights. As an initial experiment, trees representative of the Appalachian Mountain region in Virginia were selected for destructive sampling. Prior to felling, trees were scanned with the robotic sensors. Following this, trees were felled and detailed stem profile data were collected. These data include diameters at every meter along the stem, the diameter at breast height (1.37 m or 4.5 ft), diameter at 5.27 m (17.3 ft) to construct Girard form class, bark thickness, and total height of the tree. Species of these tress was noted down along with any abnormalities regarding the stem. These detailed stem profile data are being used as ground-truth to evaluate the effectiveness of robotic height and diameter measurements. These data are also used to calculate the stem volume. Known volumes have been compared with parametric prediction models and non-parametric methods using robotics data. The following table demonstrates this comparison on the initial experiment. DBH (cm) Height (m) Parametric Volume (m^3) Measured Volume (m^3) 15.7 17.4 0.158 0.178 17.9 17.8 0.211 0.236 42.5 21.9 1.233 1.742 30.5 20.3 0.590 0.803 28.2 22.7 0.619 0.747 26.1 21.2 0.497 0.610 31.3 21.2 0.706 0.750 We compared the estimates obtained the robotic platform to those in the above table estimated by the Virginia Tech team. Across 8 trees that were felled in the first stage of the experiment, we found that Median error in cm of estimates (with the table above as ground truth) Hand-carry Flight DBH Profile DBH Profile 0.55 1.1 1.05 1.05 This is remarkable on two counts: It indicates that estimates of these quantities (both the DBH which is a key metric in forestry and the profile of the diameter with respect to the tree height which is used for parametric volume estimation) are within 2% of the estimates obtained by experts. The estimates obtained from the flying platform are roughly similar in terms of accuracy to the ones obtained using the hand-carried sensor rig.
Publications
|
|