Source-linked AI summary
An informative path planning framework for UAV-based terrain monitoring
Marija Popovic, Teresa Vidal-Calleja, Gregory Hitz, Jen Jen Chung, Inkyu Sa, Roland Siegwart, Juan Nieto
TL;DR
Efficient UAV data acquisition requires planning under finite resources when informative regions are unevenly distributed and initially unknown. The paper presents a general adaptive IPP framework that maps terrain online with probabilistic, variable-resolution sensors and optimizes continuous 3-D trajectories. It reports higher efficiency than state-of-the-art methods, validates online photorealistic mapping, and demonstrates real-time agricultural monitoring.
Problem
Finite sensing resources constrain measurements, while terrain information value may be unevenly distributed and unknown; direct Gaussian Process mapping also becomes computationally expensive as dense imagery accumulates.
Method
A modular adaptive IPP framework fuses variable-resolution probabilistic sensor data into terrain maps and optimizes continuous 3-D trajectories initialized by a coarse grid search.
Results
The approach achieved higher efficiency than state-of-the-art methods, supported online photorealistic mapping with a SegNet-based sensor, and ran in real time for an agricultural monitoring proof of concept.
Takeaways & Limitations
The framework supports online information-rich monitoring of discrete or continuous terrain variables across simulation, public-dataset, and outdoor agricultural scenarios.
Takeaways & Limitations
The SegNet model incorrectly mapped two small cars because it was limited in segmenting fine details given its training data.
Abstract
from arXiv · showhide
Unmanned Aerial Vehicles (UAVs) represent a new frontier in a wide range of monitoring and research applications. To fully leverage their potential, a key challenge is planning missions for efficient data acquisition in complex environments. To address this issue, this article introduces a general Informative Path Planning (IPP) framework for monitoring scenarios using an aerial robot, focusing on problems in which the value of sensor information is unevenly distributed in a target area and unknown a priori . The approach is capable of learning and focusing on regions of interest via adaptation to map either discrete or continuous variables on the terrain using variable-resolution data received from probabilistic sensors. During a mission, the terrain maps built online are used to plan information-rich trajectories in continuous 3-D space by optimizing initial solutions obtained by a coarse grid search. Extensive simulations show that our approach is more efficient than existing methods. We also demonstrate its real-time application on a photorealistic mapping scenario using a publicly available dataset and demonstrate a proof of concept for an agricultural monitoring task.
1 Introduction
The paper introduces a modular aerial-robot IPP framework for adaptive terrain monitoring under limited sensing resources. It combines online mapping and continuous 3-D trajectory planning to focus data collection on informative regions.
- Finite energy, time, or travel-distance resources limit measurements, motivating paths that maximize information about unknown environments.
- Existing IPP solutions are often platform- or domain-specific, while many real robots still use passive coverage-based data acquisition.
- The proposed framework adapts plans online as measurements reveal targeted objects or regions of interest.
- It integrates probabilistic mapping of dense visual data with informative trajectory search in continuous 3-D space, balancing sensor resolution, FoV, constraints, and adaptivity.
- The framework was evaluated in simulation, validated on a publicly available dataset, and demonstrated in a real-time agricultural monitoring proof of concept.
2 Related work
Related work spans environment representations, adaptive information gathering, and continuous-space planning. The paper addresses computational and generalization limits by combining efficient probabilistic mapping with greedy initialization and trajectory optimization.
- 2.1 Environment mapping: Occupancy grids efficiently represent spatial sensing with uncorrelated measurements and support discrete active classification tasks.
- 2.1 Environment mapping: Gaussian Processes model interdependencies and uncertainty, but direct use in terrain monitoring becomes computationally expensive as dense imagery accumulates.
- 2.1 Environment mapping: The paper extends multiresolution Bayesian mapping to handle both non-correlated and correlated monitored variables with altitude-dependent sensing.
- 2.2 Informative planning: Adaptive planners change actions as information arrives, unlike non-adaptive coverage strategies with predetermined actions.
- 2.2 Informative planning: Limited-horizon replanning reduces overhead but trades away guarantees on exploration optimality outside the planning horizon.
- 2.2 Informative planning: The method uses a greedy grid-based IPP solution to initialize evolutionary optimization of smooth polynomial trajectories in continuous robot workspace.
3 Problem statement
The IPP problem is to select a continuous aerial trajectory that maximizes an information objective while respecting a resource budget. Measurements are collected along the trajectory and assigned a corresponding cost.
- IPP seeks an optimal continuous trajectory above terrain that maximizes an information-theoretic measure under resource constraints.
- The measurement function collects finite measurements along a 3-D trajectory, while the cost function must remain within predefined budget B.
- The information objective I(·) quantifies the utility of the measurements acquired along the trajectory.
4 Mapping approach
The framework supports discrete and continuous terrain mapping from probabilistic, variable-resolution measurements, then uses online maps for adaptive information-driven monitoring. Discrete targets use occupancy-map updates, while continuous targets use Gaussian-process priors and recursive Bayesian fusion that accounts for spatial correlations, noise, and resolution.
- Framework overview: The framework maps either discrete or continuous terrain variables from sensor measurements and uses the resulting maps to support adaptive planning.The map representation is selected according to the received data, and planning is performed during the mission using online terrain maps.
- Discrete variable mapping: Discrete targets are represented as independent Bernoulli variables in a 2-D occupancy map, with altitude-dependent inverse sensor likelihoods updating each observed cell.The update accumulates the previous log-likelihood with the likelihood from the current measurement and pose.
- Discrete variable mapping: Binary-classifier sensor models account for poorer classification at higher altitudes, and multiple labels can be handled with separate occupancy-map layers.The altitude-dependent curves can be estimated from classifier accuracy analyses using true and false positives at different altitudes.
- Continuous variable mapping: Continuous targets are modeled with Gaussian Processes that encode spatial correlations and initialize recursive filtering with probabilistic measurements at different resolutions.The GP prior is characterized by a mean and covariance, while recursive filtering reduces the computational burden of applying GP regression directly.
- Gaussian-process mapping: The continuous-map prior is inferred at selected prediction points using a kernel-based covariance model, with kernel hyperparameters controlling environmental relations.The framework supports known or unknown environments and allows inference at the same or different resolutions.
- Sequential data fusion: Recursive Bayesian fusion incorporates altitude-dependent measurement variances and a measurement model that represents lower-resolution observations as scaled averages of the maximum-resolution map.The measurement model selects observed state elements at full resolution and combines multiple map locations into one measurement at higher altitudes.
5 Planning approach
The planner builds adaptive fixed-horizon UAV trajectories by combining coarse 3-D grid-search initialization with evolutionary optimization, using information gain to value measurements. Utility definitions support discrete occupancy maps and continuous-variable maps, while grid density trades computational cost against initial-solution accuracy.
- Adaptive replanning: The planning scheme alternates replanning and execution until a time budget B is exceeded, producing fixed-horizon trajectories from the current environment model.The environment model Z can represent either a discrete or continuous target variable.
- 3-D grid-search initialization: A coarse 3-D lattice search greedily selects candidate measurement sites using utility and travel cost, simulates measurements, and assembles initial control waypoints.The lattice provides a low-accuracy initial solution under constant-velocity travel assumptions.
- 3-D grid-search initialization: Grid-point density trades solution accuracy against computation: the denser lattice produces better initial solutions but requires longer evaluation times.Figure 5 contrasts 14-point and 30-point lattices in a 40 m×40 m×30 m workspace.
- Trajectory optimization: CMA-ES optimizes the coarse-search waypoints using the information objective, with the initial UAV position clamped to the current pose and points constrained to the feasible workspace.CMA-ES samples candidate solutions from a multivariate Gaussian distribution and is used for nonlinear, non-convex continuous optimization.
- Utility definition: For discrete maps, exploratory utility is the reduction in Shannon entropy between prior and posterior maps, with occupancy-cell probabilities determining entropy.The same principle extends to complete trajectories by iteratively fusing measurements and computing overall information gain.
- Utility definition: For continuous maps, the planner maximizes covariance-trace reduction as an A-optimal-design criterion measuring total map-cell variance.This avoids the computational expense of calculating entropy through the covariance determinant.
6 Experimental results
Experiments evaluate the framework for continuous and discrete terrain mapping, adaptive replanning, benchmark comparison, and optimizer selection. Across these settings, the method rapidly reduces uncertainty and error, focuses measurements on target regions, and outperforms or complements the evaluated baselines, with stated short-mission and classifier limitations.
- Continuous mapping: The framework produced a fairly complete continuous-variable map in a short mission, while identifying most uninteresting regions.CMA-ES shifted initial measurement sites to high altitudes for rapid low-resolution data collection before map refinement.
- Comparison against benchmarks: Both the proposed algorithm and RIG-tree outperformed spiral and random benchmarks, which do not use IPP objectives to select waypoints.The comparison used uncertainty, RMSE, and MLL during a 200 s mission budget.
- Comparison against benchmarks: The proposed algorithm achieved lower uncertainty and error than RIG-tree under the same budget, while RIG-tree’s fixed step size limited later map refinement.Step sizes suitable for initial ascents constrained incremental navigation during subsequent refinement.
- Comparison against coverage: More than 50% time savings were obtained for missions longer than 500 s when matching the coverage planner’s map uncertainty.The coverage path required lower altitude and greater travel distance in these missions; coverage performed better only at 100 s.
- Optimization methods: CMA-ES achieved the lowest uncertainty and error among evaluated optimization methods, while large step sizes and Bayesian optimization slowed or worsened performance.Large CMA-ES steps caused random fluctuations, and Bayesian optimization was difficult to tune for the nonlinear domain.
- Adaptive replanning evaluation: Adaptive planning reduced uncertainty in interesting regions more rapidly, especially in clearly separated environments, without disadvantages when the field was uniformly dispersed.Early in missions, adaptive and non-adaptive approaches behaved similarly while exploring the initially unknown map.
- RIT-18 mapping scenario: In the RIT-18 scenario, targeted planning reduced uncertainty and error fastest in areas of interest, whereas non-targeted planning performed better on overall map uncertainty.The targeted planner initially explored and then concentrated on high-probability BRV regions, where low-altitude measurements improved completeness.
- RIT-18 mapping scenario: The RIT-18 deployment misclassified two small cars because the SegNet model was limited in segmenting fine details in its training data.The authors identify alternative classification methods and target classes as future work.
7 Field deployment
The framework was deployed in real time on a UAV for agricultural vegetation monitoring, using onboard imaging and online mapping. The field trial showed decreasing uncertainty and increasingly complete maps of weeds and crop rows as measurements accumulated.
- Deployment setup: The field trial ran the mapping and informative-planning algorithms onboard and in real time during UAV vegetation monitoring.The DJI Matrice M100 monitored a 20 m × 20 m field area between 8 m and 21 m altitude.
- Online mapping: The system mapped normalized Excess Green Index from RGB images using a Gaussian-process method for continuous variables.ExG was computed as ExG = 2g − r − b from normalized color channels, with measurements collected at 0.20 Hz.
- Planning configuration: The planner used polynomial trajectories with three waypoints, maximum reference velocity of 5 m/s, and acceleration of 3 m/s2.Planning optimized uncertainty reduction without an interest-based threshold, using a coarse 14-point lattice and 4.5 m initial CMA-ES step sizes.
- Field results: The total map uncertainty decreased over time, providing the quantitative validation because normalized ExG ground truth was unavailable.The uncertainty curve was offset by approximately 100 s because recording began before takeoff.
- Field results: The estimated maps became more complete as images accumulated, identifying weeds at field edges and crop-row details in the central area.Bluer regions represented lower ExG crop areas, while yellower regions represented higher ExG weed areas.
8 Conclusion and future work
The paper presents a general aerial-robot IPP framework that maps discrete or continuous terrain variables and optimizes continuous-space trajectories from coarse 3-D initial searches. Simulations and real-world evaluations show higher efficiency, online photorealistic mapping, and real-time agricultural deployment, while future work targets larger environments and temporal dynamics.
- Contributions: The framework maps discrete or continuous terrain variables from variable-resolution probabilistic-sensor data for informative path planning.Online terrain maps support information-maximizing trajectories in continuous space.
- Contributions: Continuous-space trajectories are optimized from initial solutions generated by a coarse 3-D search.This combines informed initialization with trajectory optimization for the planning stage.
- Evaluation: Simulations with synthetic and real-world data showed higher efficiency than state-of-the-art methods and lower uncertainty in value-dependent regions of interest.The framework was also evaluated for active classification using a publicly available dataset.
- Evaluation: The system was demonstrated online in photorealistic mapping with a SegNet-based sensor and in real time for an outdoor agricultural monitoring task.The agricultural deployment served as a proof of concept for onboard operation.
- Future work: Future work will address scaling to larger environments, temporal dynamics, persistent monitoring, and improved practical map building.The paper also releases the planner implementation for community use and further development.