Conference Agenda

Over the past years, Unmanned Aerial Vehicles (UAVs) have become essential tools for inspection and monitoring tasks in various sectors due to their versatility, low operational cost, and ability to carry a wide range of sensors with which they can be equipped. This paper presents an integrated methodology that combines two complementary approaches using UAV-based imagery and deep learning techniques: the inspection of electrical grid components and the monitoring of urban expansion near power line systems. For infrastructure inspection, object detection models such as YOLOv5 are used to automatically identify electrical assets, enhancing reliability and reducing manual inspection efforts. In parallel, semantic segmentation is applied to orthophotomosaics generated from drone flights to detect and quantify urban growth, allowing the identification of unregistered constructions and potential sources of unauthorized energy use. The proposed dual approach provides actionable geospatial intelligence to support energy utility companies in both maintaining infrastructure integrity and managing irregularities in urban development. Results demonstrate the potential of combining photogrammetry, computer vision, and deep learning for continuous, large-scale monitoring of energy systems and their surrounding environments.

Low-cost alternative solutions have advanced last-mile delivery, with Unmanned Aerial Vehicles (UAVs) emerging as a promising option for logistics tasks. However, as UAV operations increasingly occur in densely crowded urban areas, safety concerns - especially for people nearby - have intensified. To ensure safe deliveries, real-time UAV path planning is essential for avoiding no-fly zones defined around individuals detected along the route. This study addresses this challenge by evaluating human detection confidence in UAV imagery using the YOLOv7 model and a custom dataset. It also estimates individuals’ positions through the Monoplotting technique. The customized YOLOv7 model achieved an average precision of 53.8% and an inference time of 9.9 ms, supporting real-time deployment. Human detection confidence exceeded 85% for individuals located approximately 30 meters from the UAV when flying at altitudes up to 30 meters. However, detection accuracy declined with greater distances or higher altitudes. The Monoplotting method produced an average planimetric error of 4.296 meters, with errors increasing as the distance between the UAV and the detected person increased, mainly due to the UAV’s inertial system. These findings offer valuable insights for enhancing the safety and reliability of UAV-based last-mile delivery operations. Continued advancements in human detection are essential to support the scalable and responsible integration of UAVs into urban logistics systems.

Unmanned Aerial Vehicles (UAVs) are increasingly operating in complex urban environments, requiring robust path planning methods that balance safety and efficiency. This work presents a general-purpose methodology for generating safe UAV flight routes by transforming spatial risk heatmaps into geospatial graphs. Using superpixel segmentation, the heatmaps are discretised into risk-weighted regions that form graph nodes, with edges representing spatial proximity and weighted by Euclidean distance. Classical pathfinding algorithms — Dijkstra, A*, and Ant Colony Optimisation — are applied to compute optimal or near-optimal paths that balance both distance and risk exposure. Experiments conducted over a large urban area in São Paulo, Brazil, demonstrate that graph-based routing effectively incorporates environmental risk data, with Dijkstra and A* offering superior computational performance suitable for real-time applications. The approach is modular, scalable, and adaptable to various heatmap types, providing a flexible foundation for UAV traffic management and autonomous navigation in dense airspace.

To address the time-consuming and complex multiple agile Earth observation satellite scheduling problem (multiAEOSSP), we propose LLM-EC, an LLM-assisted evolutionary task dispatching heuristic design framework with three key innovations: 1) Customized prompt strategies for initialization, crossover, and mutation tailored for efficient order-dispatching heuristics. 2) The evolved heuristics decompose multi-AEOSSP into parallel single-satellite subproblems, significantly reducing solution complexity. 3) Experiments on realistic scenarios demonstrate that LLM-EC outperforms both standard LLMbased evolutionary algorithms and expert-designed heuristics in solution quality and convergence speed.

Earth Observation Satellites (EOS) are critical for acquiring space-based information, supporting diverse applications from environmental monitoring to urban planning. The increasing demand for satellite imaging services necessitates efficient planning for complex and heterogeneous observation tasks, including point, area, and especially challenging moving targets. Traditional mission planning approaches, often relying on "single objective, single model, single algorithm" paradigms with meta-tasks and time windows, struggle to integrate these diverse requirements and face scalability issues with increasing task complexity. To address these limitations, this paper proposes a novel Grid Heat-driven Imaging Satellite Complex Task Planning Method. We introduce a Grid Heat-based Requirements Observation Model (GHROM) that unifies point, area, and moving targets into a single, time-variant geospatial heatmap, where grid cell "heat" represents observation priority. For this grid-based representation, we design a Satellite Discrete State Task Planning Model that transforms the problem into finding the highest heat gain path within a Directed Acyclic Graph (DAG) of discrete satellite states, where attitude transition constraints are embedded as graph edges. Based on this model, we develop a Grid Heat-Driven Dynamic Programming Algorithm (GMDPA). Extensive simulation experiments demonstrate the effectiveness and significant advantages of GMDPA, particularly in large-scale scenarios, showcasing its superior performance in terms of solution quality and computational efficiency compared to traditional heuristic, metaheuristic, and genetic algorithms.