Public reporting on drone warfare — aggregated daily.
On-board acoustic measurements of three commercial multirotors show that propeller design—not just sound pressure level—drives perceived noise annoyance. Low-noise blade configurations produced higher roughness and tonality artifacts, meaning acoustic detection systems relying on overall sound levels alone may misclassify drone annoyance and surveillance risk.
SFC-DETR combines spatial and frequency domain processing to improve detection of small objects in drone imagery. The approach uses polarity-aware attention and adaptive frequency modulation to handle scale variation, occlusion, and complex backgrounds—key challenges for surveillance and inspection applications.
A lightweight cascade-front method combining thermal and motion masks from consecutive frames enables rapid target localization in large thermal frames before deep classification. The physics-based approach shifts detection away from shape matching toward thermal field dynamics, addressing real-time performance constraints in automated surveillance systems.
A sigmoid-based extended state observer with dynamic bandwidth tuning reduces disturbance rejection-noise trade-offs in quad-ducted fan UAVs flying low. The method handles coupled effects—thrust lag, ground effect, and wind—relevant for operators fielding small platforms in confined or cluttered airspace where attitude hold is critical.
CAF-YOLO combines progressive multi-scale attention, adaptive cross-scale feature fusion, and improved boundary localization to reduce false positives and false negatives when detecting small, distant objects in drone imagery. The model addresses scale variation, motion blur, and complex backgrounds—persistent sources of missed and spurious detections in airborne visual surveillance.
MSEV-DETR, a detection framework built on RT-DETR, improved average precision on aerial imagery by 3.7 percentage points over YOLOv8m through staged feature enhancement, cross-level fusion, and variance modulation. The gains on VisDrone2019 data demonstrate that selective refinement of multi-scale representations preserves small-object detail in complex UAV scenes.
MSFE-DETR addresses small-object detection challenges in aerial images through multi-scale feature enhancement and deformable attention mechanisms. For surveillance professionals, the integrated modules—CMFE, C2f, and MPCA—reduce missed detections and background interference, improving target discrimination in complex UAV scenes.
Researchers implemented virtual-array methods in 4×4 UAV-integrated sensing and communication transceivers to overcome physical aperture limits. Virtual-7 and Virtual-10 configurations increased effective aperture size and improved multi-target detection performance, directly addressing the angular-resolution constraints that limit UAV-based radar and RF sensing deployments.
DAA-YOLOv8s, a modified object detector, retains fine-grained spatial information lost during standard downsampling to improve detection of small vehicles in UAV scenes. The framework achieves 52.70% mAP@50 on VisDrone2019-DET, enabling more reliable localization under scale variation and occlusion—direct gains for aerial surveillance operators deploying resource-constrained platforms.
SW-DEIM couples local convolution, multi-scale context, and learnable channel residuals to preserve detection-relevant features under adverse weather without restoration overhead. Two synthetic benchmarks—Drone-SW and SW-UAVDT—enable evaluation of small and occluded target detection across rain, fog, and snow conditions using end-to-end architecture scaling that prior weather-specific detectors have not addressed.
A modified YOLO detector using stride-4 prediction and bidirectional feature fusion lifts small-object average precision from 8.7% to 12.8% on COCO validation data while reducing parameters. The configuration trades inference cost for detection accuracy, relevant to operators balancing detection performance constraints in aerial surveillance systems.
Radar micro-Doppler signatures of a hovering DJI Mini 2 reveal that axial wind produces spectral splitting independent of payload mass, enabling wind speed estimation before payload correction. At 7 GHz, payload shifts the dominant peak ~0.75 Hz per gram; wind separation scales at 16 Hz/(m/s), allowing platform-specific decoupling of environmental and mass effects.
CAS-YOLO combines directional strip convolution and circulant attention to detect infrared targets from UAV platforms under computational constraints. The detector addresses thermal blur and extreme aspect ratios—common in airborne surveillance—while reducing edge processing load below standard real-time detectors.
A review of 50+ studies on adversarial attacks against airborne and UAV object detectors reveals that defenses treat viewpoint variation only through object scale, neglecting explicit geometric modeling. For surveillance and delivery systems relying on onboard deep learning, this gap leaves realistic threat vectors unaddressed.
Researchers developed an intrusion detection system that reconstructs UAVCAN transfers and extracts payload and transfer features to identify message injection, replay, and flooding attacks. Transfer-rule features proved particularly effective, achieving 99.89% accuracy on a public dataset. The approach addresses detection gaps in drone networks where automotive CAN systems fall short.
SAFLR-Net combines multi-scale feature enhancement, high-resolution fusion, and adaptive localization to detect weakly textured, occluded small objects from UAV platforms. The method addresses density, texture weakness, and boundary ambiguity—critical challenges for aerial surveillance systems relying on computer vision processing.
RASM-Net coordinates cross-modal fusion with structural and multiscale modeling to handle varying reliability between RGB and thermal imagery across scenes and spatial regions. The approach addresses missed detections and localization errors for small objects and weak-boundary targets in low-light and occluded environments critical for all-weather aerial surveillance.
A partial-channel attention module paired with multi-scale fusion improves detection of small, densely packed targets in UAV frames while reducing computational overhead. On VisDrone, PA-YOLO-s achieved 45.0% AP50 and 27.2% AP50:95, addressing the occlusion and weak appearance cues that challenge aerial surveillance systems.
RViTCANet combines radar cross section imaging with temporal modeling and cross-modal attention to identify low-altitude drones from 26–40 GHz millimeter wave data. The framework addresses fixed receptive fields and rigid temporal modeling in prior RCS methods, enabling more robust recognition of attitude-dependent scattering patterns across multi-scale dynamics.
A lightweight tracking pipeline applied to airborne video reveals that detector coverage, not identity preservation, limits multi-object performance on resource-constrained platforms. Test-set MOTA of 27.2% reflects 86% false negatives from incomplete detection rather than tracking failures—a distinction critical for engineers optimizing aerial surveillance systems under computational budgets.
A systematic review of small-object detection methods in aerial imagery identifies multiscale representation, feature enhancement, and scenario constraints as key technical challenges. For sensor and surveillance engineers, the framework clarifies deployment trade-offs across military, agricultural, and disaster-response applications where object scale and imaging conditions vary.
MSF-DETR, a Transformer-based object detector, achieves 86.3% mAP50 on the DIOR benchmark by combining adaptive cross-scale feature interaction, frequency-domain attention, and spatial-detail preservation. The framework addresses a key limitation in aerial surveillance: reliably detecting small and densely packed objects in UAV imagery where scale variation and complex backgrounds degrade performance.
A new convolutional network suppresses background clutter in frequency space to boost detection of small, low-contrast UAVs in aerial imagery. The approach combines cross-scale feature fusion with side-channel refinement, addressing missed detections and false alarms that plague conventional methods on small airborne targets.
A dual-pathway approach isolates the detection failures specific to small objects in drone imagery: border-biased labeling and feature loss from downsampling. CPD-FCOS adds a real stride-4 pyramid level with center-constrained assignment, recovering 2.98 mAP on VisDrone where median targets span 26 pixels. The gain is scale-specific, relevant when objects fall below four feature cells.
Researchers demonstrated that a Soft Actor-Critic algorithm reduces penetration time for low-altitude fixed-wing UAVs navigating radar threats and terrain. Hardware-in-loop tests across five- and two-radar scenarios confirm onboard execution feasibility, helping air-surveillance teams evaluate trade-offs between speed, detected exposure, and computational load.
YOLO object-detection models lose accuracy when UAV images are compressed with JPEG before transmission to ground stations. Engineers must balance data reduction against detection performance, requiring quality-factor analysis to preserve critical target localization and classification in compressed aerial imagery.
Researchers propose EWH-YOLO, a deep learning method for detecting small unmanned aerial vehicles in vision-based systems with weighted bidirectional feature fusion and hybrid bounding box regression. The approach reduces model complexity while improving detection accuracy over existing methods, advancing aerial collision avoidance and surveillance capabilities for sensor engineers.
FLAD addresses pixel-level detection of small UAVs in visible-light imagery by stacking edge enhancement, attention-based feature fusion, and a lightweight decoupled head. The architecture isolates and corrects distinct failure modes—contour degradation, clutter ambiguity, and localization instability—relevant to operators deploying optical detection systems in complex backgrounds.
A feature-level fusion method combining spatial compression and cross-attention enhancement improved UAV detection on the Anti-UAV300 benchmark from 76.2% to 83.6% mAP, while controlling computational load. The advance addresses modality mismatch in electro-optical and thermal imaging for airborne surveillance.
A refined YOLO architecture addresses scale mismatch and receptive-field gaps in bird's-eye detection tasks, reaching 40.1% mAP on VisDrone—5 percentage points above baseline. The three-part method reweights receptive fields, fuses cross-scale features, and dynamically calibrates predictions, improving recall for tiny instances critical to surveillance and inspection workflows.
SiamDC adds spatial and channel fusion modules to Siamese correlation tracking, reducing false locks on distractors and handling occlusion in aerial video. Tested on three UAV benchmarks, the method maintains real-time speed while outperforming the baseline, directly improving robustness for operational tracking systems.
Analysis of military surveillance radars reveals that single-band systems lack the scanning speed and algorithms to track hypersonic targets reliably. Operators require complementary radar networks across different frequency bands to maintain continuous detection of modern air threats including small UAVs and hypersonic missiles.
A fixed-wing UAV flight experiment revealed that wind-induced aircraft drift introduces angular errors into radar cross section (RCS) measurement diagrams during real-world testing. Researchers developed a correction procedure using flight-controller telemetry and coordinate smoothing to quantify and compensate for this distortion, improving the reliability of RCS testing methodologies for drone detection systems.
A new detection framework combines super-resolution with human-guided feature enhancement to improve small-object detection in aerial images. By integrating operator knowledge directly into the pipeline for categorical and regional refinement, the method achieves better target–background discrimination while maintaining a simple, detector-agnostic architecture.
Researchers added high-resolution detection heads and structural attention modules to YOLOv12n to recover small-scale pedestrians lost to network downsampling in drone footage. On VisDrone2019-DET, DGF-YOLO achieved 59.7% mAP50 and 26.9% mAP50-95, directly addressing detection and tracking challenges for surveillance platforms operating at altitude where target pixel area remains constrained.
A YOLOv12-based model combining RepViT backbone and attention-scale fusion improves detection of densely packed, small targets in aerial video while maintaining real-time performance on edge hardware. The 4.9 percentage-point gain in mAP@0.5 on VisDrone2019 addresses the core challenge of extracting small objects from complex aerial backgrounds without sacrificing speed—critical for operational drone surveillance systems.
A new detection model uses controlled feature flow to handle unequal reliability between thermal and visible sensors when tracking small UAVs at low altitude. The method achieves 49.53% mAP@0.5 on drone-person detection while running at 60 FPS, addressing a core challenge in multi-modal surveillance systems where one sensor often degrades faster than the other.
A new system integrates YOLOv8 object detection with nano instance segmentation and dueling double DQN for tracking multiple aerial targets. The approach addresses the challenge of maintaining object identity across video frames in real-time UAV surveillance, relevant to operators managing detection confidence and track continuity in dense airspace.
A new detection framework combining high-resolution feature layers with an adaptive bounding-box loss function achieves 94.8% mAP on small-target UAV detection—gains of 4.2 percentage points over baseline methods. The approach directly addresses scale variance and spatial detail loss that degrade performance in airport and perimeter surveillance.
Two commercial 3D ground-based radars detected moving targets and produced georeferenced tracks within tens of meters of actual drone positions during controlled flight tests simulating bird and bat movement. Detection performance degraded near radar coverage limits and during dynamic maneuvering, informing radar system selection for ecological monitoring and collision risk assessment.
Researchers built AQF-Net to address tiny-object detection in high-resolution drone imagery by combining fixed-query self-attention, multi-scale feature fusion, and density-aware query allocation. Tests on three datasets show the framework outperforms baseline methods, enabling better detection of small aerial targets in complex backgrounds and variable target densities.
A conditional diffusion framework fills gaps in fused radar and RF trajectories caused by occlusion, maneuvering, or sensor dropout. AeroDiff-TIR treats interrupted segments as missing time-series data and infers 3D paths between observed points, improving long-term surveillance reliability for ground-to-air monitoring systems.
A systematic review of 59 event-based detection studies (2019–2025) reveals that evaluation inconsistencies—particularly inference frequency settings—create accuracy variations far exceeding typical algorithmic improvements. Event-native neural networks now achieve competitive automotive benchmarks while neuromorphic drone detectors operate at sub-10 mW power, but standardized testing protocols are essential for reliable sensor comparison.
Researchers deployed a traffic-window intrusion detection system within a MAVLink/PX4 proxy to identify attacks on drone command-and-control links. The Random Forest classifier, trained on 9,645 labeled traffic windows, enables real-time anomaly detection for operators securing autonomous vehicle telemetry networks.
A comparison of five machine learning models applied to acoustic drone-propeller data shows that each learner demands a distinct set of features to reach equivalent accuracy levels. Engineers cannot transfer feature selection across models; domain validation is required when automating feature choice for acoustic detection systems.
BMDnet, built on YOLOv11n, addresses scale variation and weak edge information in aerial imagery through bidirectional feature fusion and dynamic attention mechanisms. Performance gains on VisDrone2019 data demonstrate practical improvements for surveillance systems without prohibitive computational cost increases.
A simplified YOLOv11n architecture reduces model parameters and computational load for edge deployment while maintaining detection accuracy in low-visibility infrared. The approach uses depthwise separable convolution and dual-attention modules to preserve feature expression, enabling real-time pedestrian detection on resource-constrained airborne platforms.
Researchers combined convolutional and recurrent neural networks to classify small UAVs from radio frequency signals, reducing false positives across signal classes compared to traditional CNNs. The approach captures spatial and temporal dependencies in RF signatures but requires higher computational resources, posing trade-offs for deployment on processing-limited detection systems.
A heterogeneous detection network that fuses infrared and visible-light imaging improves UAV detection across varying illumination and weather conditions. The wavelet-guided fusion module separates structural and detail information in the frequency domain, enabling adaptive cross-modal encoding. This addresses a core vulnerability of single-modality detection systems in low-altitude surveillance.
Researchers fused independent radar and passive RF observations into a single causal state model, tested on known UAV flights from AERPAW public data. The approach substantially improved RF prediction accuracy across two validation flights while introducing minimal radar degradation, demonstrating feasibility before application to unidentified aerial phenomena.
A new dataset curation system called uavews integrates four heterogeneous sensor streams—takeoff indications, public warnings, mobile reports, and visual-acoustic records—into a unified event-centered model for validating small-UAV detection and tracking systems. The pipeline enforces ground-truth derivation, uncertainty quantification, and access controls, enabling reproducible evaluation of early-warning sensor fusion across multimodal sources.
Researchers released complete reproducibility materials for a thermal object detector trained on high-altitude UAV imagery, including 23 trained model checkpoints, 112 frozen prediction sets, and calibration methods. The archive enables validation engineers to audit detector performance and confidence calibration across thermal imagery conditions without retraining.
Source code release documents quantization, benchmarking, and hardware-cost trade-offs for deploying ground-based drone detection on edge platforms. The work addresses practical implementation constraints for surveillance systems operating without centralized compute, supporting deployment decisions across resource-limited surveillance hardware.
Researchers released training code for a lightweight fusion method that classifies UAVs using RGB, LiDAR, and radar while explicitly handling missing modalities. The approach uses a slot-rule mechanism to maintain classification performance when individual sensors fail or are unavailable—relevant for deployments where sensor reliability varies.
Researchers have released a 3D LiDAR point cloud dataset containing annotated frames of two DJI drone models and three bird species, recorded with a Livox AVIA sensor. The dataset supports development of LiDAR-based object detection systems for distinguishing aerial targets in surveillance and tracking applications.
A detector-augmented extension of the SAMURAI foundation model improves robustness in RGB-based drone tracking, particularly during long-duration surveillance and re-entry events. The approach mitigates temporal inconsistencies from detection dropouts and bounding-box sensitivity, addressing a gap in urban surveillance scenarios where conventional motion models fall short.
A technical review maps how different internal combustion engine types affect UAV performance across endurance, mass, and detectability. For detection and tracking professionals, the analysis clarifies how propulsion choices create distinct acoustic and thermal signatures—critical factors in sensor system design and target classification.
Researchers deployed a quantized YOLOv11n model on embedded STM32 processors for real-time object detection without GPU overhead. The approach achieves 74.6% model compression with 2.3% accuracy loss, enabling surveillance and tracking platforms to extend flight endurance and reduce operator workload during onboard processing.
SionnaWallNet provides simulated 10 GHz Doppler radar signatures for through-wall human detection from UAVs in urban search-and-rescue scenarios. The dataset maps how building materials and drone standoff distance affect detection feasibility, enabling sensor engineers to validate radar-based detection algorithms for SAR operations without live-fire testing.
Researchers compared YOLOv7, YOLOv8, and YOLOv11 on the public YOLO Drone Detection Dataset using matched settings and a shared metric implementation. Results show performance differences across models under standard and degraded conditions, providing surveillance engineers a controlled reference for real-time detection architecture selection.
YOLO-SWIFT combines wavelet decomposition with coordinate attention to recover fine details lost during feature compression, addressing scale variation and background interference in UAV detection tasks. The system achieves 38.7% mAP50 on VisDrone2019, relevant for operators deploying automated detection in complex aerial surveillance scenarios.
A new framework called TAE improves aerial tracking of small targets in low-light environments by focusing enhancement only on target regions rather than amplifying background noise. The authors released DarkSOT, a 268-sequence benchmark across nine target categories, to advance evaluation of nighttime UAV-based tracking systems.
Version 1.4.3 integrates environmental sensors with UAV imagery to classify airspace risk levels from Normal to Critical. The system aligns weather and video evidence through metadata matching and supports manual pairing when synchronized data are unavailable, enabling operators to validate findings before deployment.
WS-RTDETR combines frequency-domain filtering, feature pyramid enhancement, and wavelet-guided spatial refinement to detect small objects in degraded aerial imagery. On synthetic foggy data, the method outperforms baseline RT-DETR by 2.9 mAP points while maintaining computational efficiency, advancing detection capability for low-altitude autonomous operations.
A knowledge-distillation framework reduces parameters in oriented SAR ship detectors while preserving rotated localization accuracy. The approach uses learnable gates to prune convolutional and fully-connected layers, combined with dual-teacher guidance in a shared proposal space. Enables deployment of precise ship detection on resource-constrained platforms.
A new real-world dataset of 7,118 images from 32 airports provides 10,047 UAV and 7,995 bird annotations for training small-object detection systems. The collection addresses the operational need for reliable discrimination between unmanned aircraft and wildlife in low-altitude airport environments where existing labeled data is scarce.
A revised YOLO architecture using multi-scale frequency decomposition and dynamic feature fusion reduces false negatives when small drones or aircraft appear densely packed in the same frame. The approach preserves fine detail while suppressing redundant cross-scale information, improving performance for radar and EO-IR systems tracking multiple simultaneous targets.
Researchers published audited code, evaluation notebooks, and detailed protocols for passive RF-based drone classification under wireless interference. The release includes DroneDetect V2 validation, initialization controls across multiple pretrained models, and efficiency measurements on NVIDIA L4 hardware—enabling direct reproduction and benchmarking of closed-set passive RF detection pipelines.
Researchers released raw experimental datasets validating an FM sonar approach using thermophones for airborne ultrasonic sensing despite drone flight noise. The work addresses a detection channel—ultrasonic acoustic—that operates independently of radar and EO-IR systems, offering sensor fusion candidates for operators designing redundant drone surveillance networks.
A UAV-mounted magnetometer coupled with support vector machine classification achieved 95.86% macro-F1 score distinguishing buried metallic landmines from debris, using percentile-distance features derived from magnetic anomaly signals. The approach demonstrates potential for autonomous mine detection surveys, though validation across additional sites and target types remains necessary for operational deployment.
A curated dataset isolates IQ fragments of unmanned aircraft flight control signals from complex baseband RF recordings, enabling direct analysis of drone RF signatures. The dataset supports detection, classification, RF fingerprinting, and protocol analysis—foundational tasks for RF-based drone surveillance systems.
A new analysis examines how Edge AI deployed directly on unmanned aerial vehicles and surveillance systems performs target recognition and battlefield awareness without cloud reliance. For sensor and surveillance professionals, on-platform processing reduces latency in detection workflows and supports autonomous decision-making in contested environments where communication links may be denied.
Researchers demonstrate a defense-aware edge intelligence system that identifies and mitigates adversarial attacks targeting LiDAR, radar, and GPS inputs on autonomous drones and vehicles. The system increased attack detection rates and reduced adversarial success rates while maintaining real-time performance—critical for surveillance operators relying on sensor fusion for reliable target tracking.
Researchers quantified adversarial attack success rates against edge AI in autonomous UAVs and connected vehicles, finding 97% UAV navigation degradation and 80% LiDAR spoofing rates. A three-layer defense framework is proposed, though validation remains pending—relevant for surveillance engineers assessing sensor robustness in autonomous tracking systems.
Researchers resolved a fundamental trade-off in multirotor actuator allocation: greedy optimization demands impossible motor rates while smooth allocation sacrifices control authority. A barrier-function approach enforces minimum wrench-rate authority through quadratic programming, with tracking error bounded and motor-parameter uncertainty handled in closed form—relevant for systems demanding sustained surveillance or precision maneuvers.
Researchers deployed graph convolutional networks with reinforcement learning to reduce airborne video transmission delay by selectively streaming only the pixel regions containing suspicious objects rather than full frames. For operators monitoring UAV feeds in real-time surveillance, lower latency improves response speed to detected aerial threats.
Researchers generated 55,187 annotated images across rain, snow, and fog conditions to address the scarcity of real-world adverse-weather training data for drone detection systems. The dataset enables sensor engineers to evaluate detection performance across seasonal and weather domain shifts without costly field data acquisition, improving operational readiness across deployment conditions.
A framework tunes integrated sensing-and-communication systems to vary detection and localization accuracy based on target proximity to regions of interest. Sensor engineers can now allocate resources differently for threats near restricted airspace versus sparse areas, improving efficiency in distributed surveillance networks.
Researchers propose a two-timescale framework allowing multiple operators to jointly track UAVs in ISAC systems while keeping sensing data and control variables locally distributed. Covariance intersection and motion-uncertainty calibration enable reliable track fusion across operators without sharing raw measurements or knowing cross-operator estimation error correlations.
A systematic review identifies backbone networks, multi-scale fusion, and attention modules as key improvements to YOLO-series algorithms for small-target detection in remote sensing. The analysis shows progress toward integrated electro-optical and radar detection systems, addressing the shift from single-modality to collaborative multi-sensor approaches in low-altitude defense and UAV monitoring.
ZenTrack adds template-perturbation regularization during training to stabilize thermal infrared small-target tracking without computational overhead at deployment. The method addresses error accumulation and background clutter endemic to TIR imaging, enabling more reliable tracking for anti-UAV and surveillance applications.
Researchers generated deceptive surface textures that cause visible-light deep neural network detectors to fail on UAVs across varied viewpoints and ranges. The method uses 3D mesh modeling and differentiable rendering to optimize textures that fool detection systems—a finding relevant to operators defending against spoofing attacks on surveillance platforms.
A multi-scale group target counting method integrating attention mechanisms into LSC-CNN architecture reduces mean absolute error by 8% on benchmark datasets. The approach enhances detection of personnel, vehicles, and drones across scales by selectively prioritizing target information while suppressing background interference—relevant for density-map-based surveillance systems requiring improved localization accuracy.
A purely CPU-driven method combines uncertainty-aware background subtraction with semantic-guided denoising to extract moving objects from UAV and vehicular sensors in adverse conditions without GPU overhead. The approach suppresses false alarms from noise and dynamic backgrounds while preserving edge detail on targets, enabling real-time foreground detection on edge payloads with strict computational limits.
DISTA-Net++ unmixes closely-spaced infrared small targets using continuous position predictions rather than fixed discrete grids, reducing false and missed detections. The work establishes the CSIST-100K benchmark and toolkit to address a previously under-explored gap in long-range thermal target resolution for sensor fusion systems.
Researchers introduce HARD, an ultra-high-resolution airborne dataset (12768×9564), and propose streaming-HOTA, a latency-aware tracking metric that exposes how end-to-end processing pipelines—not just detectors—shape accuracy and speed. Findings show association latency skews tracking performance asymmetrically, reshaping baseline expectations for wide-area aerial surveillance systems.
RAD-YOLO, a YOLOv8s variant, uses multi-scale feature modelling and refined localisation to detect small objects in dense aerial scenes. An edge-deployable version enables real-time inference on embedded platforms, addressing the dual challenge of accuracy and computational efficiency for surveillance operators.
LAI-YOLO integrates attention blocks and a custom loss function to detect small aerial targets on computationally constrained flight platforms. The framework reduces model parameters while maintaining detection precision, addressing a core bottleneck in onboard real-time detection and tracking systems.
ScopeMamba-YOLO decouples long-range contextual modeling from the convolutional stream using a Cascaded Global-Context Module positioned outside the main feature path, improving detection of small aerial targets in UAV and remote sensing imagery. The architecture adds stride-4 detection while preserving peripheral spatial support, gaining 0.67 mAP50 points over baseline without on-path selective scanning interference.
A geometric memory framework prevents catastrophic forgetting when aerial platforms adapt visual geo-localization models across successive missions with changing environmental conditions. Hybrid diversity-based replay selection maintains reference satellite imagery utility while retaining mission-specific airborne observations, enabling robust localization over long-term autonomous operations.
Researchers demonstrate that PGD adversarial perturbations can compromise acoustic-based drone localization, exposing a vulnerability in non-visual navigation systems. The work addresses a gap in adversarial robustness research by analyzing attack effects on acoustic sensing and proposing a recovery algorithm, relevant to operators deploying drones in GPS-denied or visually-constrained environments.
UFO-DETR, an end-to-end detector for small airborne targets, uses cross-space frequency enhancement and multi-scale attention to outperform RT-DETR-L on both accuracy and computational load. The approach matters for edge-deployed surveillance systems where real-time processing on UAVs is constrained by onboard compute.
STATrack demonstrates fully spiking neural networks for visual tracking on UAVs using only RGB input, avoiding reliance on specialized event cameras. The approach introduces Adaptive Mutual Information Maximization to preserve target-specific detail in sparse temporal representations, offering path toward low-power tracking on standard aerial platforms.
A new detection transformer explicitly models both frequency and spatial domains to preserve high-frequency information needed for small object detection. The approach uses adaptive fusion and bidirectional feature propagation, addressing a core limitation in surveillance systems where conventional detectors lose critical detail when processing small aerial targets.
Researchers developed solar-powered military surveillance drones combining artificial intelligence, computer vision, and RRT-based autonomous navigation to enable real-time target identification and adaptive decision-making with minimal human intervention. The integration addresses operational efficiency demands in surveillance systems requiring sustained field deployment under variable environmental conditions.
A quadrotor with dual-axis rotor tilting completed a 25 m square trajectory 39% faster than conventional designs while generating 5.6× greater thrust and 3.22× greater torque. For surveillance operators tracking agile aerial targets, improved rotor authority and control responsiveness could enable faster intercept of small unmanned systems in dynamic environments.
Researchers integrated optical flow into reinforcement learning to create an adaptive UAV routing policy that responds to dynamic threats in real time. The approach uses motion patterns detected by onboard sensors and game-theoretic modeling to guide surveillance missions, offering surveillance operators a method to maintain coverage despite shifting adversarial activity.
A coarse-to-fine framework with denoising-assisted training improves infrared small target detection by explicitly learning target-background context. ECFNet reduces computation in non-target regions while balancing detection accuracy and efficiency, addressing key challenges in UAV-based surveillance and ground monitoring systems.
BirdsEye shifts drone annotation from image space to world coordinates using RTK positioning, then projects observations to video frames via calibrated camera geometry. For surveillance teams deploying field systems, the authors derive closed-form sensor specifications that convert annotation tolerance into admissible pose-noise budgets, reducing guesswork in sensor design.
A new framework uses a reinforcement learning controller to dynamically adjust audio segment duration during acoustic localization of unmanned aircraft, reducing latency while maintaining 3D position accuracy. This adaptive approach addresses the fixed-window trade-off that constrains existing audio-based detection systems used in early-warning surveillance.