Source-linked AI summary
Online Planning for Sparse Ground Target Search from a High-Altitude UAV under Partial Observability
Ashik E Rasul, Hyung-Jin Yoon
TL;DR
High-altitude UAVs can have a target in view that remains too small for reliable detection, while urban flight constraints can limit physical movement. This paper plans PTZ sensing with POMCP, scale-conditioned observation reliability, and selective ensemble detection; simulation results report stronger performance at higher altitudes across camera resolutions than baseline methods.
Problem
At high altitude, small target scale and image-resolution limits challenge detectors, while urban flight regulations can make physical movement to improve observations infeasible.
Source: §I · p. 2Method
The method formulates PTZ search as a POMDP, solves it with POMCP, conditions observation reliability on target scale, and selectively applies ensemble detection.
Source: §I · pp. 2–3Results
The method outperformed single-scan, traditional slicing-aided, and coarse-to-fine methods at higher altitudes beyond the fine-tuning range and across camera resolutions.
Source: §VII · p. 7Takeaways & Limitations
PTZ-based belief-space planning provides an alternative for high-altitude sparse-target search when physical movement is constrained and richer image features are needed at fixed resolution.
Source: §I · p. 2The method discretizes state, action, and observation spaces although these are continuous in reality, and its scale-conditioned observation model assumes conservative domain knowledge.
Source: §VII · p. 7
Abstract
from arXiv · showhide
Unmanned aerial vehicles (UAVs) searching for ground targets from a high altitude face a unique challenge, particularly when the target is already within the field of view but effectively unobservable because of its small apparent scale. Standard object detectors often underperform in such scenarios because of resolution downscaling and limited context. In contrast, active object search frameworks address this challenge by directing the agent to a suitable pose to gather richer visual information. However, flight regulations in urban airspace often restrict such physical movements for UAVs. As an alternative active sensing approach, the UAV can leverage the pan-tilt-zoom (PTZ) mechanism of the onboard camera to dynamically adjust its field of view and sequentially gather enhanced visual information from specific regions of interest. Once the candidate locations of the target are identified, it can deploy more expensive object detection schemes, such as an ensemble of multiple models, to get better reasoning at a fixed scale. In this work, we formulate the sequential exploration with PTZ operation as a partially observable Markov decision process (POMDP), in which the agent maintains a belief state over the target's true location. To solve the POMDP, we deploy partially observable Monte Carlo planning (POMCP), where we condition the sensing reliability on target object scale and deploy selective ensemble detection as an additional reasoning step. We validate our methodology with experiments in a photorealistic simulator under different environmental conditions and vehicle states, showing detection of ground targets at variable scales with significantly fewer steps and minimal dependence on sensor resolution compared to baseline methods.
I. INTRODUCTION
High-altitude UAVs may see a target that is too small for reliable detection, while physical movement and exhaustive sensing can be constrained. The paper uses PTZ-based belief-space planning, scale-conditioned sensing reliability, and selective ensemble detection for sparse target search.
I. INTRODUCTION
- At high altitude, small targets and resolution downscaling can defeat detectors, while physical movement may be infeasible and exhaustive sensing costly.Slicing and coarse-to-fine methods can produce uncalibrated confidence scores and false positives.Source: §I · p. 2
- The search is partially observable: noisy observations vary with PTZ settings, environmental conditions, and target pose, requiring a belief state to select actions.Wide scans cover area quickly but yield noisier observations; precise scans are slower but more reliable.Source: §I · p. 2
- The paper formulates high-altitude PTZ target search as a POMDP and uses POMCP to select informative actions while updating target-location beliefs.Observation uncertainty is modeled conditional on target appearance, with a Beta-Bernoulli model incorporating prior knowledge.Source: §I · pp. 2–3
- Selective ensemble detection is deployed at fixed scale when belief concentration is high, adding a reasoning step to the POMCP search.The method also proposes iteratively updating the model as deployment experience accumulates.Source: §I · p. 3
- In photorealistic simulation, the method finds targets in significantly fewer steps with minimal dependence on camera resolution compared with traditional tiny-object detection methods.The simulation includes cluttered urban scenes, varying altitudes, camera attributes, and dynamic lighting.Source: §I · p. 3
II. BACKGROUND
POMDPs represent sequential decisions under partial observability by maintaining beliefs over states. POMCP is an online approximate solver that uses sparse belief sampling and Monte Carlo tree search.
II. BACKGROUND
- A POMDP models state transitions, observations, rewards, and a policy that chooses actions from the agent’s belief over states given its history.The belief is updated after each action and observation, and the objective is to maximize expected future discounted cumulative reward.Source: §II · p. 3
- Exact POMDP solutions are often intractable, and offline approximate solvers can become slow in larger domains.The passage notes exact approaches are generally tractable only for toy problems.Source: §II · p. 3
- Online POMCP scales through sparse belief sampling and Monte Carlo tree search, using an unweighted particle belief and a black-box simulator.In this application, observation likelihood is explicitly modeled as part of the simulator.Source: §II · p. 3
III. RELATED WORKS
Prior work studies attention-driven visual search, active sensing, and small-object detection, but the paper identifies limitations for sparse high-altitude UAV search. It positions PTZ control with online POMCP and scale-conditioned observation uncertainty as an alternative.
III. RELATED WORKS
- Attention research describes spotlight and zoom-lens models, feature integration, guided search, and suppression of irrelevant information as influences on visual-search models.Guided search relates deployment order to saliency and signal-to-noise ratio rather than distractor count alone.Source: §III · p. 3
- Human attention models motivate sequential visual search, but digital-fovea approaches retain resolution in acquired images without active enhancement before acquisition.Their offline-learned policies require extensive training and lack interpretability.Source: §III · p. 3
- Indoor robotic search methods use physical movement or sensor assumptions that do not capture high-altitude UAV scale variation and urban movement constraints.Related approaches include POMCP search with functioning sensors, language-conditioned observation models, and distance-based observability for UAV search and rescue.Source: §III · pp. 3–4
- Slicing and coarse-to-fine detection can suffer resolution downscaling and uncalibrated confidence, while ensembles are costly and require sufficient features at a specific scale.The paper proposes acquiring richer features with PTZ at fixed resolution and selectively applying ensembles when belief is concentrated.Source: §III · p. 4
- Existing PTZ control often targets tracking and assumes complete observability, whereas this work uses online POMCP and models observation uncertainty by target scale within the FOV.The paper frames these assumptions as insufficient for sparse high-altitude ground-target search.Source: §III · p. 4
IV. PROBLEM FORMULATION
The problem is discretized into grid windows, with a static target and scale-conditioned observation reliability. Rewards trade off scan cost, detections, and false positives while favoring earlier detection and declaration.
IV. PROBLEM FORMULATION
- The search space is discretized into m × m grids, with states, actions, and observations instantiated as windows that lie fully within the grid.Each window is parameterized by its top-left coordinates and side length.Source: §IV · p. 4
- The target is assumed static, so the transition model is the identity; the observation model gives the probability of an observation for a true target state and scan action.Detection reliability depends on target presence in the scan window and apparent scale, with Beta-Bernoulli priors maintained by discretized scale bin for true detections and false alarms.Source: §IV · p. 4
- In-tree rewards combine a negative per-step cost with a detection bonus or false-positive penalty, encouraging fewer detector inferences and actions more likely to yield true detections.The terminal declaration reward is applied outside the POMCP tree.Source: §IV · p. 5
- Discounting the in-tree and terminal rewards encourages early detection-generating actions and rewards correct target declaration in fewer steps.The terminal reward is much larger for declaring the true target state than for an incorrect state.Source: §IV · p. 5
V. METHODOLOGY
The method plans scans with POMCP over a particle belief updated from detector observations, while a coarse scan initializes a saliency prior. A selective ensemble checks concentrated beliefs, and declaration occurs when belief concentration exceeds a threshold.
V. METHODOLOGY
- The belief is represented by K unweighted particles and updated by rejection sampling; a coarse full-area scan seeds the search with a saliency prior.The MAP estimate and confidence are computed from the particle histogram.Source: §V · p. 5
- POMCP simulations sample states from the root belief, traverse the tree with PO-UCT, and balance action value against exploration before rollout and incremental Q-value backup.The rollout policy selects scans uniformly at random.Source: §V · p. 5
- When belief concentration exceeds τe and the planned scan covers the MAP at scale ratio ρ = 1, five-model ensemble detection replaces the default scan; a detection is recorded if any model detects.Otherwise the observation is null, followed by the regular belief update.Source: §V · p. 5
- After each belief update, the agent declares the MAP state when concentration exceeds τ; this termination policy is external to the POMCP tree.Figure 3 shows an ensemble rejection collapsing confidence at a false-positive MAP, followed by a true-target detection that concentrates belief and leads to declaration.Source: §V · p. 5, Figure 3 · p. 6
VI. EXPERIMENTS AND RESULTS
The experiments compare POMCP-PTZ with scan, crop, and coarse-to-fine baselines across altitudes, resolutions, and simulated environments. POMCP-PTZ performs best at the highest tested altitude, while its inference-call and search-time costs matter for mission choice.
VI. EXPERIMENTS AND RESULTS
- Evaluation uses three CARLA environments, fixed 640 × 640 resolution, 100–350 m altitudes, and ten trials per method across three seeds.Six YOLO models are trained, including one fine-tuned for 0–100 m; POMCP-PTZ uses 50k simulations and 11.5k particles.Source: §VI · p. 6
- Baselines include a single full-FOV scan, three-scale overlapping crops, and coarse-to-fine detection with crop or PTZ enhancement and an ensemble variant.The coarse-to-fine methods form candidate clusters from an initial low-threshold detection before a fine pass.Source: §VI · p. 6
- At 350 m, POMCP-PTZ significantly outperforms all methods across scenarios; at 300 m, it wins in Town 05 and Town 07 and matches C2F-PTZ-Ensemble.At 100 m, baselines perform as well as or better than POMCP-based methods; at 350 m, other methods fail through non-detection or wrong detection.Source: §VI · p. 6, Figure 4 · p. 7
- POMCP-PTZ uses significantly more YOLO calls than Fixed Scan and C2F methods, but not significantly more than Exhaustive Crop.Fixed Scan uses one call, Exhaustive Crop 35, C2F-Crop two, and C2F-PTZ-Ensemble six; each ensemble action counts as five calls.Source: §VI · pp. 6–7
- As resolution rises, Exhaustive Crop improves, while POMCP-PTZ changes only marginally and remains significantly better than baselines despite slight degradation from train-test resolution mismatch.POMCP-PTZ is not affected by the slicing-related resolution downscaling described for crop methods.Source: §VI · p. 7
- Fixed Scan, Exhaustive Crop, and coarse-to-fine ensemble can be reliable in known environments and altitude ranges; POMCP-based methods are more reliable as conditions change, but take planning and PTZ actuation time.The choice depends on mission requirements and the operational environment.Source: §VI · p. 7
VII. CONCLUSION
The paper presents a belief-space POMDP planning algorithm for high-altitude sparse-target search and reports advantages over several detection baselines at higher altitudes and across camera resolutions. Its current method and validation remain bounded by discrete planning spaces and simulator-only testing.
VII. CONCLUSION
- The belief-space POMDP algorithm incorporates perception uncertainty and outperforms single-scan, slicing-aided, and coarse-to-fine methods at higher altitudes beyond fine-tuning and across camera resolutions.Source: §VII · p. 7
- The method discretizes state, action, and observation spaces even though these variables are continuous in reality.The authors suggest extending the approach to continuous-space POMDP methods such as POMCPOW.Source: §VII · p. 7
- Its scale-conditioned observation model relies on conservative domain knowledge and is identified as needing online adaptation from search feedback.Source: §VII · p. 7
- The authors identify planning efficiency, time-consuming physical PTZ actuation, and the single-helipad scope as areas for further development.Source: §VII · p. 7
- The method was developed and tested only in simulation and still requires real-world prototype validation.Source: §VII · p. 8