Drone Swarm Strategy for the Detection and Tracking of Occluded Targets in Complex Environments

Drone swarms offer collaborative capabilities exceeding those of individual drones, finding applications in surveillance, mapping, target detection, and infrastructure inspection. This research focuses on improving target detection and tracking in challenging environments like dense forests, where occlusion significantly hinders visibility. Existing approaches, employing either centralized or decentralized control and various algorithms (heuristics, evolutionary, learning-based), often struggle with occlusion. The authors introduce an innovative method that combines synthetic aperture (SA) sensing with a novel particle swarm optimization (PSO) strategy to overcome these challenges. SA sensing, a signal processing technique used in radar, radio telescopes, and other applications, computationally combines measurements from multiple sensors to mimic a larger aperture, effectively mitigating occlusion. The paper's central hypothesis is that an adaptive drone swarm, guided by PSO, can significantly improve target detection and tracking speed in complex, occluded environments compared to traditional blind sampling methods. The study's importance lies in its potential to revolutionize search and rescue operations, wildlife observation, and other applications in difficult-to-access areas.

The paper reviews existing literature on drone swarms, highlighting centralized and decentralized control systems and the application of various algorithms for swarm behaviors. It discusses the use of meta-heuristic algorithms, particularly PSO, for their computational efficiency and robustness in handling partial observations. The authors also survey the applications of SA sensing in various fields, emphasizing its potential for enhancing imaging through occlusion. Previous work by the authors on Airborne Optical Sectioning (AOS) is described, outlining earlier sequential and parallel sampling techniques. These previous methods, while demonstrating occlusion removal capabilities, lacked the adaptive capabilities necessary to efficiently address highly variable and unpredictable occlusion patterns in complex environments. The limitations of these earlier methods motivate the development of the PSO-based swarm approach.

The core of the methodology is a novel PSO algorithm designed specifically for the challenges presented by the non-constant, highly random, and non-differentiable objective function of AOS in a dynamic environment with moving targets and variable occlusion. The authors address these challenges by making the PSO more explorative than exploitative. The swarm's behavior is constrained to the current time instance, biasing exploration toward a temporal global leader (the best sample at the current time). A minimal distance constraint is enforced to define the SA properties. The objective function is conditionally integrated, combining parallel (current time instance) and sequential (previous time instances) samples. Instead of an inertia weight constant, a conditional bias toward a default scanning pattern is implemented. The algorithm includes a collision avoidance strategy using simple altitude offsets. The objective function, O(Ibest), measures target visibility by calculating the contour size of the largest connected pixel cluster (blob) from anomalies in color and thermal channels after applying a Reed-Xiaoli anomaly detector. The algorithm iterates, refining the swarm's positions based on the objective function, until a sufficient target visibility threshold is reached or the process is manually aborted. The simulation environment uses a procedural forest model generating aerial images (RGB and thermal) for different forest densities and sampling procedures (parallel, sequential, single drones, camera arrays, swarms). The experiments compare different sampling strategies: (1) a single sequentially sampling drone, (2) a parallel camera array, and (3) swarms of varying sizes (3, 5, and 10 drones). The impact of forest density and target movement is also investigated. Hyper-parameters of the PSO, such as cognitive and social coefficients, scanning speed, and minimal sampling distance, are directly related to SA properties. The simulation also includes a detailed model of spatial sampling loss related to altitude differences among drones and pose estimation error.

The key findings are presented through several experiments and visualizations. The paper demonstrates that the proposed adaptive PSO-based swarm strategy significantly outperforms blind sampling methods. In a simulated environment with a 300 trees/ha forest density, the swarm achieved a maximum target visibility (MTV) of 72% in 14 seconds, compared to 51% after 75 seconds for sequential brute-force sampling and only 19% after 3 seconds for parallel sampling. Experiments reveal that increasing the swarm size improves visibility and coverage, while increased forest density reduces visibility, as expected. The algorithm effectively tracks moving targets, with minor discrepancies between estimated and actual target positions, speed, and direction. The simulation results are also compared to real-world integral images to validate the simulation setup. The performance of the algorithm is analyzed in terms of visibility improvement over time for different swarm sizes and forest densities, illustrating the algorithm's superior performance in locating and tracking targets efficiently. The impact of different hyperparameters on the swarm's performance and the algorithm's robustness to noise and uncertainty are also explored.

The superior performance of the adaptive swarm strategy compared to blind sampling is attributed to the swarm's ability to autonomously adapt to local viewing conditions. The swarm prioritizes optimal viewing angles for maximizing target visibility by converging on areas with lower occlusion and oblique viewing angles. The combined use of sequential and parallel samples further enhances visibility. The results confirm the hypothesis that an adaptive swarm approach leads to faster and more reliable detection of strongly occluded targets. The limitations of the existing statistical visibility model are addressed by extending it to include parallel-sequential sampling with moving targets, showcasing the effectiveness of the proposed method even in scenarios with significant dynamism. The findings are highly relevant to the field of swarm robotics and SA sensing, with implications for search and rescue, wildlife monitoring, and other applications involving target detection in complex environments.

This study successfully demonstrates the feasibility and efficiency of using a drone swarm guided by a novel PSO algorithm for detecting and tracking occluded targets in challenging environments. The proposed approach significantly outperforms traditional blind sampling methods in terms of both speed and accuracy. Future work will involve real-world experiments with physical drone swarms, optimizing the algorithm, and exploring more sophisticated swarm control strategies. The development of methods for automatically determining the outlier removal threshold and investigating alternative collision avoidance techniques would further enhance the approach's robustness and efficiency.

The current study relies on simulations, and while real-world data validation is provided for integral image comparison, additional real-world flight experiments are necessary to fully validate the findings. The simulations do not account for all real-world factors, such as drone acceleration/deceleration, data transmission times, sensor errors, and uneven terrain. The procedural forest model is simplified, and automatic and adaptive threshold determination is not included in this version of the algorithm.