Obstacle-aware Adaptive Informative Path Planning for UAV-based Target Search

Over the past decade, UAVs have transitioned from remote-controlled devices to autonomous agents capable of complex tasks such as terrain mapping, target detection, and environmental monitoring. Early approaches relied heavily on pre-defined coverage patterns, like lawnmower or spiral paths, which worked well in open, obstacle-free environments. However, these methods lacked adaptability in cluttered urban or natural terrains, where obstacles and dynamic elements pose significant challenges.

Recent advances introduced probabilistic mapping techniques, notably Gaussian Processes (GP), enabling UAVs to build continuous, uncertainty-aware models of target occupancy and terrain features. Simultaneously, obstacle-aware path planning algorithms, such as RRT* and CHOMP, improved safety and efficiency but often operated on discrete grids or simplified models, limiting scalability and resolution.

Despite these developments, existing methods generally treat exploration and obstacle avoidance as separate modules, leading to suboptimal performance in resource-constrained scenarios. The integration of Bayesian optimization with continuous path planning remains underexplored, especially in the context of real-time, adaptive target search. This gap motivates the development of unified frameworks that can dynamically balance information gain, safety, and resource management, particularly for urban search and rescue missions where environment complexity and uncertainty are high.

The core challenge addressed in this research is enabling UAVs to perform efficient, safe, and adaptive target search in cluttered environments with limited flight time and uncertain sensor data. Traditional methods either focus on systematic coverage, risking inefficiency, or on reactive obstacle avoidance, risking missed targets. The fundamental bottleneck is the lack of a unified, probabilistic framework that can simultaneously model environment obstacles, target occupancy, sensor uncertainties, and optimize the UAV's trajectory in real-time.

Specifically, the problem involves: (1) constructing an accurate, probabilistic map of the environment and target locations; (2) planning smooth, collision-free paths that maximize information gain; (3) dynamically re-planning based on new sensor data; and (4) balancing exploration of unknown regions with exploitation of promising areas. Achieving this within strict resource constraints, such as battery life and computational limits, makes the problem particularly challenging. Addressing these issues requires integrating environment modeling, sensor performance considerations, and multi-objective optimization into a cohesive, real-time capable framework.

The primary innovations of this work include:

• �� A continuous-space, obstacle-aware path planning framework that leverages Euclidean Signed Distance Fields (ESDF) for real-time obstacle detection and avoidance.
• �� Probabilistic target occupancy mapping using Gaussian Processes (GP) with Matérn kernels, capturing spatial correlations and uncertainties, enabling informed exploration.
• �� A layered optimization strategy that combines Bayesian Optimization (BO) with Upper Confidence Bound (UCB) acquisition functions for adaptive exploration-exploitation balancing, considering sensor performance variability.
• �� Multi-objective path optimization integrating information gain, collision risk, and flight time, solved via CMA-ES, ensuring smooth, safe, and efficient trajectories.
• �� A dynamic re-planning mechanism that updates paths based on real-time sensor measurements, supporting target re-observation and robustness against false positives.
• �� Extensive simulation validation demonstrating superior performance over existing methods across various environmental complexities, including urban scenarios with high obstacle density.

• �� Environment modeling: Use ESDF to represent obstacles, enabling fast collision checks. Obstacles are simplified as geometric shapes, with a signed distance function providing safety margins.
• �� Target mapping: Employ Gaussian Processes with Matérn 3/2 kernel to model target occupancy over the terrain. Hyperparameters are learned via maximum likelihood, and the map is updated sequentially with sensor measurements.
• �� Sensor modeling: Simulate altitude-dependent detection performance of YOLO, incorporating detection probability curves into the planning process.
• �� Path parameterization: Use polynomial splines connecting control waypoints, ensuring smooth, dynamically feasible trajectories.
• �� Initial path generation: Apply greedy search (COARSEGREEDYSEARCH) to select promising waypoints based on current map and sensor data.
• �� Global optimization: Use CMA-ES to refine the initial path, optimizing for maximum information gain while avoiding obstacles.
• �� Replanning: During flight, continuously update the environment map with new measurements, and re-optimize the path if resource limits permit, enabling target re-observation and adaptive focus.

• �� Setup: Simulate urban search and rescue scenarios in RotorS, with environment dimensions of 30×30×26 meters, including obstacles and 7 human targets.
• �� Sensor configuration: Downward-facing camera with YOLO Tiny 2.0 detector, detection threshold 0.05, frequency 0.15Hz, FoV (45°, 60°).
• �� Baselines: Compare against lawnmower coverage and random waypoint strategies, each with 25 trials, flight time budget of 150 seconds.
• �� Metrics: Root Square Error (RSE) for map accuracy, detection success rate, and GP posterior covariance trace for uncertainty.
• �� Hyperparameters: Path speed 5m/s, acceleration 3m/s², Bayesian optimization parameters tuned for sensor performance.
• �� Variations: Test environments with different obstacle densities and heights to evaluate robustness and scalability.

• �� OA-IPP reduced RSE by approximately 30% compared to baseline methods, achieving an average error of 0.15 within 150 seconds, successfully detecting all targets.
• �� The method maintained stable performance across environments with high obstacle density, with errors below 0.2 and consistent uncertainty reduction.
• �� Adaptive re-planning enabled the UAV to focus on promising regions, improving target re-observation and detection accuracy, even in the presence of false positives.
• �� In urban simulations, the generated paths effectively avoided obstacles, with trajectories validated in complex scenarios, demonstrating practical applicability.

• �� Search and rescue: Rapid localization of victims in urban or forest environments, especially where obstacles and uncertainties are prevalent.
• �� Urban monitoring: Autonomous patrols for infrastructure inspection, crowd monitoring, or security surveillance.
• �� Border security: Continuous monitoring of sensitive regions with obstacle-rich terrains.
• �� The framework can be extended to multi-UAV systems for large-scale, coordinated search missions, significantly enhancing operational efficiency.

• �� Assumes environment is known and static; dynamic obstacles or environment changes are not yet supported.
• �� Sensor models are based on empirical data, which may not generalize across different sensors or environmental conditions.
• �� Computational demands are high, especially in large or highly cluttered environments, necessitating further optimization for real-time deployment.