Vision-Based Perch Site Detector for UAVs

• The paper integrates dual-marker detection with deep learning segmentation to achieve sub-centimeter pose estimation for UAV perching applications.
• It employs scale-adaptive expert switching and multi-marker fusion to address challenges like occlusion, variable scale, and environmental constraints.
• Real-time processing using RGB-D sensors and optimized pipelines demonstrates enhanced landing precision and robust control integration for energy-efficient maneuvers.

A vision-based perch site detector is a specialized perception module enabling unmanned aerial vehicles (UAVs) to autonomously identify, localize, and align with suitable perching targets using onboard imaging sensors. Such detectors facilitate precise energy-conserving maneuvers—stabilizing the craft for long-duration monitoring, surface sampling, or safety-critical landing—by parsing complex visual scenes and robustly estimating perch-site geometry under variable scale, occlusion, and environmental constraints. Recent system-level research spans natural perches (tree trunks) and artificial targets (fiducial markers, helipads), demonstrating the integration of deep-learning segmentation, geometric heuristics, multi-scale marker fusion, and dual-expert switching to achieve closed-loop, real-time operation with sub-centimeter accuracy.

1. Hardware Configurations and Sensing Modalities

Vision-based perch site detectors utilize high-frequency RGB and depth sensors rigidly mounted on UAV platforms, typically aligned with the vehicle’s principal axis of motion. For tree-trunk perching, SLAP employs an Intel RealSense D435 RGB-D camera (0.3–3 m, 30 Hz) (Di et al., 1 Jan 2026), with no extraneous illumination or optics introduced. Nano-UAV approaches favor monocular camera modules, yielding 640×480 RGB streams at similar frame rates (Do et al., 2023, Do et al., 2023). The critical design consideration is field coverage—the detector’s efficacy at various altitudes/distances—which motivates multi-scale strategies (embedded dual-marker or dual-expert scale-specialized CNNs) to avoid “target loss” near touchdown (Tasnim et al., 16 Dec 2025). Mounting is constrained to guarantee a transformation from camera to body frame, preserving calibration integrity for pose back-projection.

2. Perch Site Detection Methodologies

Detection methodologies bifurcate along two axes: feature-based approaches using artificial fiducials and data-driven segmentation pipelines for natural perches.

• Marker-Based Detection: Dual-square ArUco markers (DICTIONARY 4X4_100, 150 mm; DICT_ARUCO_ORIGINAL, 25 mm) are embedded concentrically, enabling sustained visibility across approach ranges (15–115 cm). Detection proceeds via grayscale conversion, adaptive thresholding, morphological opening, contour extraction, quadrilateral filtering, and ArUco decoding (Do et al., 2023, Do et al., 2023). When both markers are detected, multi-stage fusion algorithms merge pose signals for optimal estimation (see Table 1).
• Deep Segmentation for Natural Perches: SLAP leverages “PercepTree,” a pre-trained U-Net-style forest tree segmenter. The pipeline involves static occlusion-masking, CNN-based pixelwise segmentation (yielding binary tree trunk masks), morphological cleaning, contour-connected component analysis, and geometric filtering (trunk diameter, depth-profile consistency) (Di et al., 1 Jan 2026).
• Scale-Adaptive Expert Switching: For large targets undergoing rapid scale transitions (helipads, perch sites), dual YOLOv8 experts are trained independently on far/near-scale datasets. Each expert processes images at specialized resolutions (832×832 for distant, 512×512 for close range), with a geometric gating mechanism selecting the detection hypothesis closest to image center (Tasnim et al., 16 Dec 2025). Noise is suppressed via moving average filtering over the last N detections.

Approach	Sensing Modality	Detection Principle
Dual-marker	Monocular RGB	ArUco contour/ID fusion
Tree segmentation	RGB-D	Pretrained U-Net CNN
Dual-expert YOLO	Monocular RGB	Bounding-box, gating, scale

3. Feature Extraction, Fusion, and Pose Estimation

Feature extraction is dictated by the target environment. Natural perch detectors expose the following heuristics implicitly:

• Binary mask $M_\text{tree}(u,v)$ demarcates likely trunk pixels post-segmentation.
• Bounding box width $w_i$ estimates trunk diameter (converted to metric space via median depth and known camera intrinsics).
• Depth statistics (mean, median, variance) screen non-planarity and reject overhangs or irregular bark (though explicit texture descriptors are unimplemented).

Fiducial approaches use explicit corner correspondences. The PnP pose estimation routine computes the SE(3) transform from the marker frame to the camera frame:

$x = K [R \mid t] X,$

where $K$ is the intrinsic calibration matrix, $R$ the rotation, $t$ the translation (Do et al., 2023, Do et al., 2023). Pose fusion exploits learned weighting (LMS) to combine coarse/precise estimates during overlapping visibility. Kalman filters further stabilize pose trajectories against intermittent detection outages—with velocity decay applied during missing measurements—ensuring reliable servoing inputs for perching maneuvers.

4. Real-Time Operation, Processing Pipeline, and Optimizations

These systems operate at full sensor bandwidth (≈30 Hz), with onboard acceleration (Jetson Nano/Orin GPU or embedded controller). The detection pipeline is heavily optimized:

• Static occlusion masks pre-filter input ambiguity before segmentation (Di et al., 1 Jan 2026).
• Single-scale or dual-scale processing avoids computationally intensive pyramids or multi-resolution fusion (Tasnim et al., 16 Dec 2025).
• Early candidate rejection (e.g., bounding-box width filters, convexity checks) limits expensive inference to plausible components.
• Multi-marker and multi-expert frameworks maintain detection continuity via geometric/scale gating and temporal smoothing, minimizing alignment jitter near scale crossovers.

No explicit reporting on CPU/GPU utilization was observed, but all learning (CNN segmentation, YOLO object detection) is encapsulated within pre-trained networks or standard OpenCV routines. Classical image-processing (HSV conversion, Canny edge/Hough lines, handcrafted features) is absent in these workflows.

5. Integration with UAV Control Systems

The final stage entails translating visual detection products into actionable setpoints for perching and landing planners. For tree perching, detected site centroids are back-projected from image to camera coordinates using calibration parameters $(u^*, v^*, d^*)$:

$$X_c = (u^* - c_x)d^*/f_x, \quad Y_c = (v^* - c_y)d^*/f_y, \quad Z_c = d^*$$

Extrinsic conversion aligns the perch-site pose into the drone’s reference frame, and the planning module computes the desired tip position and nominal surface normal. Trajectories are generated to produce controlled approach velocities, minimizing mechanical stress and maximizing grip reliability (Di et al., 1 Jan 2026). For marker-based schemes, the seven-phase planner sequentially drives the craft through alignment, ascent, engagement, and post-perch abort cycles, tightly integrating visual servoing with cascaded position/altitude/attitude controllers (PD/PI, as per Crazyflie firmware) (Do et al., 2023, Do et al., 2023).

6. Experimental Metrics and Performance Benchmarks

Performance is reported in terms of:

• Detection Ranges: Dual-marker setups synthesize detection for $z \in [15, 115]$ cm, with outer markers serving far-field acquisition and inner markers supporting close-range final alignment (Do et al., 2023, Do et al., 2023).
• Pose Estimation Accuracy: Position errors are consistently $<0.6$ cm peak-to-peak; heading (yaw) errors are typically within $3^\circ$ (Do et al., 2023).
• Perching Precision: The nano-UAV demonstrations achieve $<2$ cm landing precision; SLAP records a $75\%$ perch success rate on oak trunks across 20 flights, and $100\%$ failure recovery over 2 induced faults (Di et al., 1 Jan 2026).
• Dual-Expert Landing Outcomes: Under scale-adaptive gating, dual-expert YOLOv8 frameworks reach $100\%$ target detection rate across all relevant altitudes, with mean touchdown error at $2.53$ m ($\pm 1.03$ m) in simulation (Tasnim et al., 16 Dec 2025).
• Jitter and Stability: Alignment jitter under dual-expert fusion is $<2$ px average, eliminating the rapid oscillation near scale crossovers suffered by single-expert models (Tasnim et al., 16 Dec 2025).

No mention is made of explicit precision/recall curves or resource consumption; instead, practical benchmarks center on detection robustness, positional/rotational accuracy, and success/failure rates under operational constraints.

7. Generalizations, Limitations, and Future Directions

A key generalization is that scale-adaptive architectures—whether through embedded multi-marker layouts or dual-expert CNNs—consistently overcome the limitations imposed by fixed-scale detectors, providing robust tracking from initial approach through final perching/touchdown. This suggests further research into multi-expert gating, dataset stratification by object pixel size, and augmentation strategies tailored to thin or irregular perches could advance detector resilience and precision (Tasnim et al., 16 Dec 2025).

SLAP’s focus on natural perches sidesteps classical handcrafted features and exposes an implicit tradeoff: all “CV magic” (segmentation accuracy, species differentiation) is concentrated inside a single pre-trained model, limiting transparency for troubleshooting or environment adaptation (Di et al., 1 Jan 2026). Where marker-based methods achieve near-perfect laboratory performance, natural scene understanding remains sensitive to lighting, occlusion, and bark texture, with failure modes infrequently characterized.

A plausible implication is that future systems will integrate expert switching, multi-head segmentation/orientation, and distributed sensing for high-dimensional perching scenarios (e.g., urban infrastructure, curved surfaces). Research may further delineate joint detection-planning frameworks, enhancing closed-loop safety and operational efficiency.

---

Table 1. Pose Estimation and Fusion Regimes (Dual Marker System)

Regime	Detected Markers	Pose Fusion Output
Stage 1 (far)	Only M₁	$\mathbf{P_{M1}}$
Stage 2 (overlap)	M₁ and M₂	LMS-weighted combination
Stage 3 (close)	Only M₂	$\mathbf{P_{M2}}$

This table summarizes the pose selection logic for dual-marker approaches, as reported in (Do et al., 2023, Do et al., 2023).

---

Vision-based perch site detectors thus represent a confluence of robust visual perception, geometric reasoning, and tightly coupled control integration, forming a foundation for current and future autonomous UAV perching systems.