Drones

Lightweight UAVs increasingly operate in constrained environments, where even minor collisions can damage exposed propellers, arms, sensors, or embedded components. Designing protective cages for such platforms requires more than adding an external shell: the structure must preserve propeller clearance, fit the chassis, remain lightweight, avoid obstructing sensors and cameras, and remain feasible to prototype. This paper introduces PICLIN (Pipeline for Incremental Constraint-Linked InnovatioN), a graph-based design-to-prototyping workflow for UAV protective cages. PICLIN starts from abstract topological representations and progressively constrains them through geometric deployment, curved-surface embedding, numerical refinement, drone-integrated verification, and physical realization. The approach is demonstrated through a UAV case study leading from graph exploration to assembled protective prototypes. Results show that PICLIN generates coherent cage configurations while preserving early-stage design freedom and reducing the solution space toward manufacturable structures. For the representative Optimal 9 prototype, the cage mass was 151 g, excluding the current 28 g aluminium UAV interfaces, with an overall diameter of 485 mm and a design clearance of 50 mm between the propellers and the protective cage. A preliminary 1 m drop test of the 1.186 kg cage–chassis assembly, corresponding to an impact energy of approximately 11.6 J, produced a peak acceleration of approximately 35 g over an impact interval of about 15 ms. These measurements are reported as proof-of-feasibility results rather than as a complete mechanical validation. The contribution is not a single optimized cage geometry, but a reusable workflow linking abstract graph modeling, constraint integration, UAV-specific verification, and prototype-oriented realization. Full article

UAV pursuit–evasion in multi-obstacle environments constitutes a safety-critical decision-making problem in which the pursuer must intercept a maneuvering evader while satisfying obstacle-avoidance constraints. Existing safety-constrained methods typically optimize expected safety costs or apply immediate interventions, but they do not explicitly determine when recovery should begin once finite-horizon risk emerges. To address this issue, we propose the ReSwitch framework, which separates task-oriented pursuit from safety-oriented recovery. When flight risk emerges, a value-preserving switch-time planner evaluates candidate switching times through opponent-conditioned hybrid rollouts. The feasible switch time with the largest pursuit value after recovery is selected for execution, allowing the controller to prioritize safety while preserving pursuit effectiveness whenever possible. Experiments demonstrate that ReSwitch achieves a mean success rate of 89.02% and a physical safety rate of 91.67%, showing favorable performance compared with the baselines. These results indicate that ReSwitch provides a favorable pursuit–safety trade-off under the tested conditions. Full article

Accurate medium-horizon trajectory prediction, spanning roughly one to three seconds ahead, is essential for the safety and autonomy of uncrewed aerial vehicle (UAV) systems. Trajectory prediction assists UAVs in performing collision avoidance, path planning, and cooperative airspace coordination. Deep sequence models are now widely used for trajectory prediction. The gated recurrent unit (GRU) and long short-term memory (LSTM) models are among the most widely used architectures. However, published comparisons of GRU and LSTM encoder–decoder models rarely use identical data, splits, and hyperparameter budgets. This makes it difficult to attribute reported accuracy differences to the recurrent cell itself. We present a controlled comparison in which two otherwise identical velocity-based encoder–decoder networks jointly predict three future 3D velocity waypoints at +10, +20, and +30 steps ahead. Both networks share an identical 5-fold cross-validation protocol, held-out test split, fixed seed, and 81-point hyperparameter grid, for 405 runs per architecture and 810 runs total. At each architecture’s best configuration, the LSTM model reaches 7.2% lower validation Mean Squared Error (MSE) than GRU. On the held-out test set, LSTM achieves 6.5% lower test MSE under each architecture’s independently selected best configuration (best-practice comparison); a complementary matched-configuration comparison, in which each architecture is retrained under the other’s configuration, shows this advantage is concentrated in robustness to hyperparameter choice rather than in the recurrent cell alone. We also report per-waypoint, per-dimension, and Monte Carlo dropout (MC-dropout) epistemic-uncertainty metrics for both architectures. These metrics are pooled over test windows dominated by synthetic, near-planar flight data (~74%) and should not be read as general conclusions for real, free-form UAV flights. Vertical-velocity error is consistently higher than horizontal-velocity error for both models, reflecting limitations in how well the training data represent vertical movement. The MC-dropout epistemic-uncertainty intervals show 34–36% empirical coverage, substantially lower than the nominal Gaussian-reference target of 68.3% at 1σ. On our 8 × H200 GPU cluster, LSTM’s training time is 51% longer than GRU’s per run, a training-side cost that should not be read as a proxy for embedded inference cost. These results show that LSTM’s main advantage over GRU is its greater robustness to suboptimal hyperparameter settings, rather than substantially better performance when both architectures are well tuned. They also show that the uncertainty estimates from both architectures require post hoc calibration before they can be reliably used as safety bounds. Full article

The growing demand for precise unmanned aerial vehicle (UAV) operations in dynamic environments is often compromised by unmodeled wind disturbances, calling for robust and adaptive control strategies to ensure accurate trajectory tracking. This paper presents a meta-learning augmented model predictive control (ML-MPC) framework for quadrotor trajectory tracking under strong and horizontal wind disturbances with different nominal wind-speed settings. The framework uses a meta-learned basis function to capture shared nonlinear features of aerodynamic disturbances across different wind-speed conditions, while an online adaptation mechanism continuously estimates the corresponding coefficients from flight data. Their combination provides a real-time estimate of the residual aerodynamic force, which is incorporated into the MPC prediction model to compensate for wind disturbances. Extensive flight experiments validate the effectiveness of the ML-MPC framework, showing consistent gains in tracking accuracy across multiple trajectory types and under nominal wind-speed settings of up to $16\mathrm{m}/\mathrm{s}$, defined by measurements $1\mathrm{m}$ downstream of the fan array. Compared to a baseline GP-MPC controller, the approach achieves average performance improvements of 73.5% in simulation and 50.6% in real-world flight tests, with maximum reductions of 86.5% and 59.7%, respectively. Full article

Cooperative task planning for amphibious firefighting UAVs deployed across multiple bases is challenging because of complex resource coupling, evolving fire conditions, and real-time decision requirements. This study proposes an H-MAPPO-LSTM method with a three-level hierarchical multi-agent structure: a Command-Center Agent performs global situation assessment and macroscopic task allocation, Base Agents manage UAV resources and decompose regional tasks, and UAV Agents execute water collection, flight, water-dropping, and return operations. LSTM networks are incorporated into the policy networks and combined with PPO and centralized training with decentralized execution to enable multilevel cooperative learning and online dynamic replanning. During deployment, network parameters remain fixed, while online policy inference, real-time decision-making, and dynamic replanning respond to fire-scene changes. The model jointly considers UAV bases, amphibious firefighting UAVs, natural water sources, dynamic fire hotspots, and their constraints. Compared with Flat MAPPO in the static benchmark, H-MAPPO-LSTM reduces mission completion time by 20.2%, average response time by 14.7%, final burned area by 29.7%, and comprehensive cost by 26.0%. Fire-containment rates reach 88% in highly dynamic scenarios and 81.96% in large-scale scenarios. Ablation results indicate that the hierarchical architecture provides the most pronounced contribution under the tested settings, while LSTM, the level-specific centralized critics, and hierarchical rewards provide additional performance contributions under the tested settings. Full article

Remote monitoring of drones is important for airspace management and security, including assessment of unauthorized payloads. Prior radar studies have established that payload state can affect drone signatures, but wind can alter rotor speeds and confound a payload estimate. Here, we examine a hovering DJI Mini 2 under controlled anechoic-chamber and wind-tunnel conditions. A balanced, top-mounted payload shifts the dominant blade-related micro-Doppler peak, whereas axial wind produces two spectral branches. We interpret the splitting as consistent with controller-mediated body tilt and thrust redistribution; rotor speed and attitude were not independently measured. At a 7 GHz carrier, increasing payload from 0 to 120 g shifts the dominant peak from approximately 340 to 430 Hz, giving a descriptive sensitivity of about 0.75 Hz/g for this platform. In the wind tunnel, the branch separation follows Δf = (16 Hz/(m/s))v, is the axial wind speed in m/s, and reaches approximately 52.8 Hz at 3.3 m/s while remaining nearly independent of the tested payload. The separation can therefore provide a model-specific wind estimate before payload correction. The resulting relations are presented as a platform-specific empirical calibration under controlled hovering. The specific contribution is the observation and empirical modeling of wind-induced branch splitting, which complements prior payload-classification methods. Full article

Reliable UAV command-and-control (C2) links require spectrum adaptation that responds to detected incumbent activity while keeping the ground station and aircraft synchronized. This study evaluates an incumbent-aware handoff workflow integrating FFT energy sensing, three-state temporal stabilization, pair-specific spectral admissibility filtering, bidirectional link-margin ranking, and a PROPOSE–ACK–EXECUTE transaction with retries and ordered backups. With a clean 32-window history, the corrected $K=768$ detector produced empirical $P_{\mathrm{fa}}=0.0997$ [0.0980, 0.1014] and $P_{\mathrm{d}}=0.9012$ [0.8995, 0.9029] at $-10$ dB. The point detection target was attained, while the 95% lower confidence bound was 0.8995. A representative logical detector-to-protocol composition at $-10$ dB and 10% independently imposed symmetric IID signaling loss gave a binary detection and recovery probability of 0.844 for the clean-history corrected $K=768$ profile, compared with 0.258 for corrected $K=51$; 95% occupied history reduced the $K=768$ value to 0.121. This composition does not exercise the mobility-coupled packet pathway or a common RF realization across sensing and signaling. In matched protocol trials, the full policy achieved $P_H=0.9321$ [0.9270, 0.9369] while removing EXECUTE; retries, backups, or margin-based ranking reduced reliability and changed the signaling–latency trade-off. Three idempotent EXECUTE transmissions reduced unresolved desynchronization to 0.008, and adding rendezvous recovery increased the eventual-recovery probability to 0.9582 at the cost of a 341-ms 95th-percentile recovery time. At 10% IID loss, the deadline-constrained success probability under the 250-ms internal interruption benchmark was 0.894. Bursty loss produced the largest degradation among the evaluated stress conditions. The results quantify how detector history quality, commit recovery, and protocol redundancy jointly affect reliability and latency. All reported quantitative validation is simulation-based or analytical/software verification; the fifteen-band by five-environment sweep is a static configuration-dependent operating envelope rather than field, hardware, or mobility-coupled validation. The reported detector–protocol integration is logical rather than fully RF-coupled, and no incumbent receiver or secondary-to-incumbent interference path is modeled. Full article

Air-to-ground (A2G) channel modeling supports low-altitude wireless networks for unmanned aerial vehicle (UAV) services in 5G/6G systems. Existing data-driven models are often benchmarked against weak empirical baselines, decoupled from the underlying measurement campaigns, and prone to overfitting on small UAV datasets. This paper presents PACE-A2G, a calibrated deep-tree ensemble on physics-aware features, together with a stricter evaluation methodology. A quadrotor-mounted software-defined-radio measurement system produces five datasets covering 433, 740, 1500, and 3300 MHz in overwater, suburban, and mountainous environments. The deep tabular backbone combines periodic numerical embeddings, feature-group attention, and a parameter-efficient ensembling head; propagation physics enters through the features rather than through the architecture. At inference, the deep estimate is blended with a tree ensemble, and a split-conformal quantile head provides calibrated prediction intervals. Evaluation under random, chronological, and spatial-block protocols shows that random partitions overestimate accuracy: the chronological-split root-mean-square error (RMSE) of the strongest tree-ensemble baseline is 1.9 to 8 times its random-split value. In distribution, the hybrid predictor is statistically on par with the strongest tree ensembles. Across nine cross-scenario and cross-frequency protocols, it transfers more robustly than the tree ensemble on balance, with a mean paired RMSE reduction of 0.34 dB, although the fitted log-distance baseline remains the strongest single predictor under shift. Split-conformal calibration reaches near-nominal 90% coverage in distribution but collapses under domain shift; a feature-space out-of-distribution gate detects the shift on every source–target pair, and ten labelled records of the new domain restore near-nominal coverage. Full article

Existing single-modality object-tracking methods for unmanned aerial vehicles (UAVs) suffer from template mismatch under illumination variation, target occlusion, and fast motion, leading to tracking failure. We therefore propose an algorithm based on the one-stream Transformer tracking framework OSTrack. Our method retains the joint feature learning and candidate target pre-elimination mechanisms while introducing dynamic template updating and quality scoring judgment modules. Dynamic template updating balances the strengths of static and dynamic templates by utilizing both initial and historical target state information. This approach prevents tracking drift from target state changes and improves target appearance. Additionally, a gating strategy based on quality evaluation screens candidate update templates to suppress contamination caused by occlusion, modal mutation, and false detections. Comprehensive experiments on the public CMOTB dataset and our self-built UAV RGB-NIR dataset validated the method’s effectiveness. For CMOTB, a 68.7% success rate (SR) and 66.5% precision (P) was achieved, outperforming vanilla OSTrack by 6.4 and 5.8 percentage points, respectively. For UAV RGB-NIR, it yielded a 74.6% SR and 76.8% P, surpassing the baseline OSTrack under complex real scenarios. The proposed method enhanced the robustness and adaptability of UAV RGB-NIR object tracking while maintaining favorable real-time performance, indicating a strong potential for practical application. Full article

Unmanned aerial vehicle (UAV) imagery presents small-object footprints, dense layouts, occlusion, and weak appearance cues, requiring both local spatial detail and broader context under constrained computation. Existing high-resolution prediction paths and attention mechanisms can strengthen these representations, but their extensive use increases feature-map, attention, and fusion costs. We propose Partial Attention–YOLO (PA-YOLO); its primary methodological contribution is a partial attention module with multi-group query attention, which applies grouped contextual modeling to selected channels while retaining a direct bypass for locally derived convolutional detail. The architecture further integrates established RepVGG-style re-parameterized downsampling with high-resolution four-scale feature fusion and detection. Fixed-recipe component ablation on VisDrone indicates complementary contributions from partial-channel attention and the integrated multi-scale design. On VisDrone, PA-YOLO-n and PA-YOLO-s report AP₅₀/AP_50:95 point estimates of 39.4/23.8% and 45.0/27.2%, respectively; under a secondary, non-official 2000-frame UAVDT-derived protocol, the corresponding values are 96.6/69.3% and 97.3/72.0%. These accuracies are single-run point estimates. Controlled forward-only profiling on one RTX 3090 platform gives mean eager-FP32 latencies of 13.795 and 14.197 ms/image, showing that compact parameterization does not necessarily yield lower inference latency than convolution-centric baselines. These findings motivate deployment-oriented optimization and onboard evaluation. Full article

Urban Air Mobility (UAM) is a dynamic field, and interest has grown rapidly as aircraft technologies and automation advance and as new applications are identified for UAM to support urban activities, economic development, and the movement of people and goods in cities [...] Full article

Precision agriculture increasingly relies on UAV-based multispectral sensors, but cross-platform data fusion remains limited by differences in spectral response, image geometry, and structural reconstruction accuracy. This challenge is particularly important in dense crop canopies, where vegetation indices and canopy height estimates may vary substantially between sensor platforms. This study compared the multi-lens DJI Phantom 4 Multispectral (P4M) and the dual-lens Sentera Double 4K in a maize hybrid trial (DKC 4596) in Debrecen, Hungary, using aerial surveys conducted at 35.5 m to assess DSM, DTM, NDVI, and NDRE outputs across 15 synchronized sampling plots processed in WebODM and RStudio. Agreement between platforms was evaluated using Bland–Altman analysis and Pearson correlation. The results showed a substantial systematic bias for NDVI, with Sentera producing 36% higher mean values than P4M and showing almost no linear association between platforms (r = 0.07). In contrast, NDRE demonstrated stronger and more manageable cross-platform agreement, with a significant positive correlation (r = 0.66, p = 0.007). Dense canopy closure reduced the reliability of absolute canopy height estimation for both platforms, but the P4M performed particularly poorly because of lens parallax and lower effective pixel density, while the higher-resolution Sentera imagery provided a more representative structural trend. Overall, NDRE is more suitable than NDVI for cross-platform vegetation index fusion between P4M and Sentera, while reliable structural assessment in tall, dense maize requires higher-resolution imagery, rigorous ground control, or complementary technologies such as LiDAR. Full article

Active signal-source seeking by swarms of unmanned aerial vehicles (UAVs) requires reliable waypoint decisions under measurement corruption, misleading response peaks, and geometric constraints. We propose MI-PSO-Adaptive, a high-level waypoint planning framework based on particle swarm optimization (PSO). The framework combines repeated sampling with robust aggregation, adaptive exploration with archive-based social target selection, candidate waypoint guidance, and geometric constraint repair. We evaluated the framework in a three-dimensional simulation using a five-sample measurement budget, waypoint limits, obstacles, and reproducible random seed protocols. In the default unimodal Gaussian field, MI-PSO-Adaptive achieved a mean localization error of 35.36 m and a success rate of 94%; model-matched Gaussian estimation baselines remained more accurate. In the fixed Gaussian mixture field, it achieved a mean localization error of 54.03 m and a success rate of 66%. In contrast, the other evaluated optimization methods and unimodal-model estimation baselines failed to reach the target region. Under 5% outlier contamination, robust aggregation achieved a mean localization error of 157.80 m and a success rate of 52%, outperforming arithmetic mean MI-PSO, although the error distribution remained heavy tailed. Functional ablation showed that adaptive exploration with social target selection helped the swarm escape misleading interference peaks, whereas candidate waypoint guidance further reduced localization errors after effective exploration. The method is therefore intended for discrete high-level waypoint planning under the tested model-mismatch conditions rather than serving as a general replacement for model-based estimation or a guarantee of low-level flight feasibility. Full article

Cooperative multi-UAV path planning must produce coupled trajectories that respect inter-UAV separation constraints through environments with moving obstacles, wind, GPS jamming, and communication loss. Population-based metaheuristics handle the resulting nonconvex, dynamic objectives well, and reinforcement-learning (RL) guidance can adapt their search behavior online; however, existing learned controllers apply one decision homogeneously to the whole swarm, even when part of the fleet is threading a cluttered corridor while the rest cruises in open space. This paper introduces group-level heterogeneous phase control for RL-guided Jellyfish Search Optimization (RL-JSO): UAVs are partitioned into risk-based or random groups, and a Dueling Double DQN selects a joint-phase assignment from an enumerated group-level action space of size $3^g$, so different groups execute different search phases within the same optimizer iteration. Across three distinct stochastic training replicates and five evaluation campaigns of increasing adversity, group-level control reduces mean mission cost by 21–71% relative to the global RL-JSO controller (Global 3A), raises curriculum mastery from 4.0 to 7.3–8.7 of nine stages, and achieves 95–100% hard-collision-free rates across campaigns under the adopted evaluator, while soft separation margins are frequently violated under the hardest conditions. A risk-versus-random ablation indicates that the specific risk heuristic is not the sole driver of the gains, which are consistent with a substantial role for heterogeneous phase diversity and group-membership dynamics; in deployment-time sensitivity analyses the fixed risk-based policy showed local robustness to moderate risk-weight and group-fraction perturbations, and the overall method ordering was preserved across objective-adaptation strengths. The results support group-level learned optimizer control as a robustness mechanism for cooperative UAV mission planning under the adopted simulation framework. Full article

Uncrewed Aerial Vehicle (UAV) swarms offer a significant paradigm for scalable and resilient mission execution, yet the integrated literature examining the intersection of individual dynamics and collective intelligence remains limited. This study conducts a structured mapping of 116 articles published between January 2021 and August 2025, providing a comprehensive overview of the current state of the art concerning the components of a UAV swarm, progressing from individual vehicle dynamics to coordinated collectives. Quantitative results reveal a dominant shift towards decentralized architectures, with connectivity maintenance and collision avoidance identified as primary functional domains. Methodologically, while control theory and optimization models form the field’s “deterministic spine”, data-driven approaches and artificial intelligence are emerging as significant research trends. Despite the widespread application of UAV swarms in simulation, the analysis highlights a disparity in experimental implementation, with only 23.3% of the reviewed research indicating hardware implementation beyond numerical simulations, predominantly restricted to controlled indoor environments, underlining critical research gaps in Sim-to-Real transfer and environmental robustness. Based on a multi-perspective analysis, the authors propose a multidimensional taxonomy categorized through three lenses—coordination mechanisms, algorithmic nature, and collective behaviours—aiming to provide a structured overview of the swarming paradigms identified in the recent literature. Full article

Existing contract-net-based UAV scheduling methods commonly use static or event-switched objective weights and provide limited evidence on how state-responsive scoring behaves under failures and communication impairment. This paper proposes a Dynamic Weight Multi-Objective Algorithm (DW-MOA) within a manager-mediated contract-net framework. Five candidate-level objectives are converted into normalized benefit utilities and combined using non-negative state-dependent weights. An independent urgency-gated spatial incentive supports cross-region dispatch without introducing negative objective weights. The method is evaluated against Traditional-CNP, Static-MOA, and HCNP-2022-adapted through matched dynamic scenarios, together with ablation, sensitivity, normalization-stability, communication-impairment, and scalability analyses. DW-MOA improves temporal performance and completed-load balance relative to Traditional-CNP, while comparisons with the stronger baselines reveal trade-offs among response time, normalized energy expenditure, flight distance, and load balance rather than uniform superiority. The mechanism remains computationally tractable at the largest tested scale and shows measurable degradation under modeled auction delay and candidate-bid loss. These findings support DW-MOA as an interpretable state-responsive scheduling method for normalized multi-UAV simulations, while physical energy calibration, continuous wireless-channel modeling, and hardware validation remain subjects for future work. Full article

We consider the technical feasibility of a wirelessly distributed antenna array for C-UAS using directional microwave frequency energy. Importantly, the large possible size of the distributed antenna permits engagement of UASs and UAS swarms within the near-field region, enabling highly localized effects, ~5 m at a range of 1 km when operating at 1 GHz. In addition to near-field operation, combining N coherent emitters gives an N² enhancement of irradiance over that of a single emitter, enabling extension of system range by a factor N or a reduction in total system power requirement of N⁻¹ to achieve an equivalent effect. Systems operating at 1 GHz require precise synchronization of the order of 100 picoseconds to ensure a high degree (>90%) of coherence at target and therefore would require semiconductor power amplifiers rather than vacuum electron devices which are typically used as high-power microwave sources. The authors develop a mathematical model implemented in software to explore the potential performance of a wirelessly distributed antenna for C-UAS and suggest some approaches for realization of such a technology. Full article

In sudden disaster scenarios, the rapid strategy-driven task-chain construction of rescue task chains in UAV-assisted heterogeneous emergency rescue systems is critical for an effective emergency response. Such systems usually involve multiple types of rescue entities, such as reconnaissance UAVs, delivery UAVs, decision centers, and other emergency resources, which provide four-dimensional capabilities including reconnaissance, intelligence analysis, command and control, and rescue disposal. However, existing manual planning and heuristic optimization methods often suffer from low online decision efficiency in large-scale disaster scenarios and rarely examine how different task-chain construction strategies affect task-chain construction. To address this problem, this paper studies a strategy-aware rescue task-chain construction problem for UAV-assisted heterogeneous emergency rescue systems, in which heterogeneous rescue entities are mapped to reconnaissance, intelligence-analysis, command, and disposal nodes according to their four-dimensional capabilities under capability, distance, cost, availability, and strategy constraints. The problem is formulated as a constrained combinatorial optimization problem, and eight task-chain construction strategies are designed from three dimensions: command mode, information-sharing scope, and disposal mode. To solve this problem, a deep reinforcement learning framework integrating a pointer network and entropy-regularized advantage actor–critic (A2C) is proposed. The pointer network generates variable-length entity-selection sequences for task chains, while entropy-regularized advantage actor–critic (A2C) optimizes the selection policy using an entropy-regularized objective that considers success probability, execution cost, and resource utilization. Experiments across 72 scenarios show that the proposed method outperforms the genetic algorithm in terms of timeliness and overall multi-indicator performance, especially in large-scale rescue scenarios. The results further indicate that distributed command strategies achieve a more balanced and robust performance, providing practical guidance for strategy selection and resource scheduling in UAV-assisted heterogeneous emergency rescue. Specifically, compared with the genetic algorithm under the full-scale 72-scenario evaluation, the proposed method improves the cost-control indicator E2 by 51.2% and the topological indicators E4, E5, E6, and E7 by 86.8%, 346.5%, 375.7%, and 421.5%, respectively, all while reducing the online decision time from 36.5 s to 3.0 s (a 91.8% speedup). The genetic algorithm still outperforms the proposed method on the success-rate indicator E1 and entity-utilization indicator E3, which is discussed as a limitation of the proposed method in the discussion of applicability. Full article

Human detection in drone imagery is an important capability for applications that require information on the location and number of people, including search and rescue, urban pedestrian-flow analysis, crowd-density assessment, and public-space safety management. However, in oblique drone imagery, acquiring and annotating sufficient real data remains difficult because of crowd density, privacy constraints, background clutter, and position-dependent scale variation. This study proposes a copy-and-paste-based synthetic data augmentation procedure for imagery acquired using a DJI Phantom 4 RTK under a 30° camera pitch angle and a 30 m flight altitude, following DJI search-and-rescue observation guidelines. The proposed procedure combines human-free drone background images with a public full-body human dataset and incorporates perspective-based geometry-aware scaling, context-aware placement, and density-matched chip selection based on real person-count distributions. Three YOLO-family detectors (YOLOv8x, YOLO11x, and YOLO26x) and the Transformer-based RT-DETR-X were evaluated at three real-data proportions (10%, 25%, and 100%). Across these twelve combinations, four augmentation strategies (random, geometry-, context-, and combined geometry- and context-aware augmentation) were compared under the same 50% synthetic ratio. For the YOLO family, geometry-aware augmentation improved mAP50-95 in eight of the nine combinations and achieved a mean gain of +0.64 percentage points over the real-only baseline, with the largest gain of +1.99 percentage points under the YOLO11x 25% real-data condition. For RT-DETR-X, the corresponding changes were +0.23, −1.83, and −0.15 percentage points at 10%, 25%, and 100% real data, respectively. These results suggest that, under the acquisition conditions evaluated in this study, aligning pasted person size with the image-position-dependent perspective scale can improve the effectiveness of synthetic augmentation when the chip-level person-count distribution is controlled to match that of the real training data. Full article