Monocular Drone Data Collection

• Monocular drone-based data collection is a method that uses a single-camera UAV paired with deep learning and sensor fusion to enable 3D mapping and localization.
• Workflows integrate real-time perception, synchronized sensor logging, and scale anchoring to overcome challenges like scale ambiguity and computational constraints.
• Applications include SLAM, photogrammetry, object detection, and human-drone interaction, offering practical insights for resource-constrained aerial robotics.

Monocular drone-based data collection refers to the systematic acquisition and processing of visual data from unmanned aerial vehicles (UAVs) that are equipped solely with single (monocular) cameras for tasks including 3D mapping, localization, scene understanding, and downstream learning or perception applications. Monocular sensing leverages lightweight, low-power, and passive sensors—making it attractive for deployment on small and micro aerial vehicles—which, due to constraints in payload and bandwidth, cannot afford multi-sensor suites. Through deep learning, model-based vision, visual odometry, and algorithmic fusion with inertial or pose data, monocular drone systems now robustly support applications in photogrammetry, 3D object detection, real-time SLAM, inspection, and human-drone interaction, even under resource constraints or in GPS-denied environments.

1. Workflow Architectures for Monocular UAV Data Collection

Monocular UAV data collection workflows comprise hardware, data acquisition protocols, real-time processing, and post-processing to yield actionable 3D data or labeled datasets. Representative system designs include:

• Sensing and synchronization: Typical platforms utilize commodity RGB sensors (e.g., PlayStation Eye, DJI Phantom-series cameras), synchronized with IMU and GPS for pose tracking; onboard and offboard compute hardware range from embedded ARM CPUs (e.g., 1.6 GHz, quad-core, running Ubuntu + ROS) to Wi-Fi-linked remote stations (Dey et al., 2014, Rosario et al., 2017).
• Real-time perception-control pipelines: Incoming monocular streams are processed for patch-level or dense feature extraction under CPU budget constraints, with depth predicted via learned regressors or deep networks (e.g., FastDepth, MiDaS, Depth Anything) (Dey et al., 2014, Zhong et al., 6 Mar 2025, Danial et al., 18 Nov 2025).
• Data logging and synchronization: Raw RGB, estimated depth, pose information, cost maps, and auxiliary sensor data are timestamped and logged at each frame, supporting subsequent dataset construction or learning (Dey et al., 2014, Rosario et al., 2017).
• Active coverage planning: Deliberative waypoint tiling, randomized yaw perturbations, and “hover-and-pan” for multi-view coverage maximize datadiversity (Dey et al., 2014).
• Sensor fusion pipelines: Visual-inertial extended Kalman filtering (EKF) fuses monocular visual odometry (VO) with IMU or SDK-provided velocity, crucial for resolving scale ambiguities in single-camera setups (Stumberg et al., 2016, Danial et al., 18 Nov 2025).

2. Monocular Depth Estimation and Metric Recovery

Monocular images lack inherent metric scale. Robust depth estimation is achieved by:

• Learned depth estimation: Neural depth-predictors such as MiDaS v3.1 (BEiT transformer backbone) and Depth Anything v2 (DINOv2 large) infer scene-relative (“pseudo”) depth for each pixel (Zhong et al., 6 Mar 2025).
• Scale anchoring: Sparse, accurately triangulated tie points from aerial SfM triangulation (e.g., using POS and bundle adjustment/extracted 3D points) are used to fit affine or rational functions mapping relative monocular depths to metric values:

$$D_{ij} \approx s_i \cdot d_i(p_{ij}) + t_i \quad \text{or} \quad D_{ij} = \frac{a_i d_{ij} + b_i}{c_i d_{ij} + d_i}$$

where $D_{ij}$ is the survey-accurate depth for tie-point $j$ in image $i$ (Zhong et al., 6 Mar 2025).

• Dense reconstruction: Scaled depth maps are back-projected using known camera extrinsics to generate dense, metrically accurate 3D clouds and digital surface models, even in regions with minimal or single-view overlap (Zhong et al., 6 Mar 2025).

3. 3D Perception, SLAM, and Visual Odometry

Monocular UAVs have advanced from sparse keypoint-based methods to architectures integrating deep learning and edge-aware mapping:

• Keypoint-based and semi-dense SLAM: Systems such as LSD-SLAM and ORB-based pipelines enable keyframe graph construction and semi-dense depth estimation by direct photometric minimization or feature matching, selectively reconstructing high-gradient regions (Stumberg et al., 2016, Danial et al., 18 Nov 2025).
• Visual-inertial fusion: Extended Kalman filters anchor the scale and correct drift by jointly tracking pose from monocular VO and metric velocity from inertial sensors (Danial et al., 18 Nov 2025, Stumberg et al., 2016).
• Deep learning-based visual odometry: Sequence models, such as SelfAttentionVO with CNN + Bi-LSTM + multi-head attention, regress SE(3) pose directly from stacked image frames, outperforming RNN baselines (22% reduction in mean translational drift, 12% lower absolute trajectory error) and demonstrating robustness under noise (Dufour et al., 2024).
• Self-supervised model identification: Advanced frameworks enable onboard training of pose and dynamics models using monocular video, IMU, and motor signals, with improved occlusion handling via multi-frame masking/minimum loss, yielding up to 15% lower odometry RMSE and robust scale adaptation without MoCap or GPS (Bahnam et al., 30 Apr 2025).
• Real-time constraints: Edge-aware semi-dense mapping and local bundle adjustment support >10 Hz end-to-end operation on embedded hardware (Jetson TX2, Raspberry Pi), enabling dense navigation and obstacle avoidance in resource-constrained micro UAVs (Danial et al., 18 Nov 2025).

4. Monocular Photogrammetry, Object Detection, and Scene Semantics

Monocular drone imaging supports a broad range of mapping and perception tasks:

• Low-overlap photogrammetry: Monocular metric depth filling addressed spatial incompleteness of MVS in low-overlap (≤30%) situations, attaining mean-absolute DSM errors of 4.29 m at 200 m altitude, and providing contiguous coverage where traditional methods fail (Zhong et al., 6 Mar 2025).
• 3D object detection: Synthetic (CARLA Drone CDrone, AM3D-Sim) and real datasets with metric, instance-level 3D bounding boxes have been produced to benchmark and train detectors. Specialized geometric BEV (bird's-eye view) pipelines leverage geo-deformable transformations, GroundMix-augmented training, and domain adaptation to handle high-altitude, oblique UAV perspectives, with AP₃D@IoU=0.5 for cars improving from 1.12% (baseline) to 10.7% with augmentation (Meier et al., 2024, Hu et al., 2022).
• Semantic and material mapping: Integration of RGB or multispectral (VNIR) imagery with elevation models enables unsupervised clustering into material classes using reflectance, NDVI, and local geometric features, with cross-modality registration for fusing spectral and photogrammetric data (Rosario et al., 2017).

5. Human-in-the-Loop, Model-Based, and Skeleton-Based Data Collection

Monocular setups have extended into tasks requiring high-level semantic interpretation or user interaction:

• Model-based tracking: Industrial inspection uses 3D skeleton models (minimal point/line sets) registered to image outputs via CNN-inferred heatmaps, then fused with IMU/GPS in a pose-graph optimizer, achieving up to 80% translation and 75% rotation error reduction with minimal sensor setups (Moolan-Feroze et al., 2019).
• Skeleton-pose human-drone interaction: Systems estimate 2D body keypoints, monocularly quantify distance classes via fully connected networks, and recognize gestures for direct control—achieving 93.5% recognition across 11 common gestures, with mean absolute distance errors of 17.75 cm (Marinov et al., 2021).
• Automation and annotation: Combined, these approaches permit autonomous, semantically annotated dataset collection, as well as the direct actuation of UAVs by gestural command or target-aligned coverage planning (Marinov et al., 2021, Moolan-Feroze et al., 2019).

6. Data Management, Calibration, and Best Practices

Protocolized data management procedures underpin monocular UAV collection:

• Calibration: Intrinsic calibration via checkerboard or Zhang’s method, and precise extrinsic calibration of camera-to-IMU/gimbal, are foundational (Rosario et al., 2017, Hu et al., 2022).
• Radiometric and geometric correction: Empirical line correction and reflectance normalization ensure cross-modal comparability; bundle adjustment and geo-rectification (using GCPs and similarity transforms) provide high-fidelity geo-referenced models (Rosario et al., 2017).
• Flight planning: ≥70% forward overlap and 30–50% side overlap are recommended for dense 3D reconstruction, with additional waypoints and randomized angles for scene diversity; hyperspectral surveys should be flown slower (≤2 m/s) to minimize motion blur (Rosario et al., 2017, Dey et al., 2014).
• Logging and quality control: Synchronization (hardware or ROS-based), comprehensive time-stamping, and logging of raw images, depth, pose, and cost maps enable reproducibility and dataset curation (Dey et al., 2014).

7. Limitations, Challenges, and Future Directions

Monocular approaches face inherent and practical limitations, but several remedies and extensions have been identified:

• Scale ambiguity and drift: Resolved through IMU-fusion, pose graph optimization, or external tie-points; residual long-term drift still possible in challenging scenes (Stumberg et al., 2016, Danial et al., 18 Nov 2025, Zhong et al., 6 Mar 2025).
• FOV and texture limitations: Narrow FOV misses side obstacles; semi-dense methods lose depth in homogeneous or saturated regions; solutions include multi-camera rigs, high-dynamic-range sensors, or active exploration (lateral parallax to reduce unknown indentations) (Dey et al., 2014, Stumberg et al., 2016).
• Computational constraints: Budgeted, anytime feature selection and edge-aware mapping minimize on-board processing; as CPUs/embedded GPUs improve, full pipelines can transition fully onboard (Dey et al., 2014, Danial et al., 18 Nov 2025).
• Synthetic-to-real transfer and dataset biases: Domain gap mitigation strategies such as GroundMix data augmentation, multi-illumination simulation, and cross-dataset pretraining expand real-world applicability of benchmarks like CDrone and AM3D (Meier et al., 2024, Hu et al., 2022).
• Potential extensions: Integration of stereo, LiDAR, active illumination, semantic segmentation, or lifelong online calibration will further close operational gaps, particularly for high-speed, aggressive, or perception-limited environments (Bahnam et al., 30 Apr 2025, Dey et al., 2014, Meier et al., 2024).

Monocular drone-based data collection now encompasses robust pipelines from hardware to annotation, synthetic to real scenarios, embedded to large-scale deployments, and sets best practices for future scalable, model-driven, and autonomous 3D data harvesting in aerial robotics.