AI drone imagery analysis: object detection and counting

AI Analysis of Drone Aerial Images

Engineers working with aerial images often face a problem: standard computer vision models on orthophoto maps yield mAP 20–30% lower than on ground-level data. The reason is unaccounted GSD, lack of georeferencing, and cross-scale drift. We solve this comprehensively—from preprocessing to deployment on an onboard NVIDIA Jetson. With 15+ projects in agriculture, energy, and construction, we've built pipelines that guarantee detection accuracy of 90%+ on GSD from 1 to 10 cm.

Why Standard YOLO Doesn't Work on Orthophotos

Analysis of UAV data differs fundamentally from ordinary photos: large orthophoto maps (5–20 GPx), non-standard GSD, multispectral and thermal channels, and the need to georeference results in WGS-84 or local CRS. The standard YOLOv8 → inference → results pipeline fails without proper tiling considering ground resolution. Our methodology includes three key stages: GSD-based tiling, sliced inference via SAHI, and coordinate transformation to GeoJSON.

Tiling and Georeferencing—The Key Step

A typical mistake: tiling orthophotos in pixels without accounting for GSD. At GSD 2 cm, a 640×640 pixel tile = 12.8×12.8 m ground area. At GSD 8 cm, the same tile is 51×51 m. A model trained on one scale will deliver mAP 15–25% lower on another.

import rasterio
from rasterio.windows import Window
from pathlib import Path

def tile_ortho_by_ground_size(
    ortho_path: str,
    tile_ground_m: float = 50.0,
    overlap_ground_m: float = 10.0
) -> list[dict]:
    """
    Tiling orthophoto by ground tile size.
    Guarantees constant scale regardless of GSD.
    """
    with rasterio.open(ortho_path) as src:
        gsd = abs(src.transform.a)          # meters/pixel
        tile_px = int(tile_ground_m / gsd)
        overlap_px = int(overlap_ground_m / gsd)
        stride = tile_px - overlap_px

        tiles = []
        for row in range(0, src.height - tile_px + 1, stride):
            for col in range(0, src.width - tile_px + 1, stride):
                win = Window(col, row, tile_px, tile_px)
                data = src.read(window=win)          # (C, H, W)
                bounds = rasterio.windows.bounds(win, src.transform)
                tiles.append({
                    'data': data,
                    'bounds': bounds,
                    'gsd_m': gsd,
                    'window': win
                })
    return tiles

Overlap overlap_ground_m=10 is critical: objects on tile edges are detected in both, then NMS by IoU removes duplicates. Without overlap, ~12% of objects at seams are lost. We use 20% overlap as standard—it provides the best balance between recall and computational load.

SAHI for Small Objects—Comparison with Naive Approach

When detecting people or cars on orthophotos with GSD 3–5 cm, the object occupies 30–80 pixels—much smaller than the receptive field of YOLO optimized for 640px. YOLOv8 with SAHI outperforms full-scale inference by 25–35% for small object detection. SAHI solves this by slicing with overlap and NMS across all predictions:

from sahi import AutoDetectionModel
from sahi.predict import get_sliced_prediction

model = AutoDetectionModel.from_pretrained(
    model_type='yolov8',
    model_path='drone_people_v2.pt',
    confidence_threshold=0.35,
    device='cuda:0'
)

result = get_sliced_prediction(
    image=tile_array,           # np.ndarray (H, W, 3)
    detection_model=model,
    slice_height=640,
    slice_width=640,
    overlap_height_ratio=0.2,
    overlap_width_ratio=0.2
)
# result.object_prediction_list → coordinates in tile pixels

On a construction site people counting dataset (5000 annotated images, GSD 4 cm): without SAHI [email protected] = 0.61, with SAHI—0.84. The difference is fundamental.

How We Process Multispectral and Thermal Images?

For multispectral data (Micasense, Parrot Sequoia), we apply index transformations (NDVI, NDRE) and feed them as additional channels to the model. The thermal channel (FLIR) requires a separate pipeline:

import numpy as np

def detect_pv_hotspots(
    thermal_kelvin: np.ndarray,    # (H, W), values in 0.01K
    panel_mask: np.ndarray,        # binary mask of panels
    delta_threshold: float = 10.0  # °C above panel median
) -> list:
    """
    Hotspot detection on solar panels.
    IEC 62446-3: defect when ΔT > 10°C from reference.
    """
    temp_celsius = thermal_kelvin * 0.01 - 273.15
    panel_temps = temp_celsius[panel_mask > 0]
    reference_temp = float(np.median(panel_temps))

    hot_mask = (
        (temp_celsius > reference_temp + delta_threshold) &
        (panel_mask > 0)
    ).astype(np.uint8)

    import cv2
    contours, _ = cv2.findContours(
        hot_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE
    )
    hotspots = []
    for c in contours:
        x, y, w, h = cv2.boundingRect(c)
        roi = temp_celsius[y:y+h, x:x+w]
        delta = float(roi.max() - reference_temp)
        hotspots.append({
            'bbox': [x, y, x+w, y+h],
            'max_temp_c': round(float(roi.max()), 1),
            'delta_t':    round(delta, 1),
            'severity':   'critical' if delta > 25 else 'warning'
        })
    return sorted(hotspots, key=lambda h: h['delta_t'], reverse=True)

On a real solar farm inspection project (142 panels, 3 flights), the system identified 17 defective panels with ΔT > 15°C that manual visual inspection missed. ROI paid off in one season. We guarantee recall not lower than 95% for hot-spot detection.

Coordinate Transformation and GeoJSON Output

Analysis results must be in geographic coordinates—otherwise they're just pictures, not integrable into GIS systems.

import pyproj
from shapely.geometry import box, mapping
import json

def detections_to_geojson(
    detections: list,
    tile_bounds: tuple,           # (left, bottom, right, top) in CRS
    tile_px_size: tuple,          # (width, height) in pixels
    src_crs: str = 'EPSG:32637'   # UTM zone for project
) -> dict:
    transformer = pyproj.Transformer.from_crs(
        src_crs, 'EPSG:4326', always_xy=True
    )
    left, bottom, right, top = tile_bounds
    px_w, px_h = tile_px_size
    scale_x = (right - left) / px_w
    scale_y = (top - bottom) / px_h

    features = []
    for det in detections:
        x1, y1, x2, y2 = det['bbox']
        # Pixels to projection coordinates
        geo_left   = left + x1 * scale_x
        geo_right  = left + x2 * scale_x
        geo_top    = top  - y1 * scale_y
        geo_bottom = top  - y2 * scale_y
        # Projection to WGS-84
        lon1, lat1 = transformer.transform(geo_left, geo_top)
        lon2, lat2 = transformer.transform(geo_right, geo_bottom)

        features.append({
            'type': 'Feature',
            'geometry': mapping(box(lon1, lat2, lon2, lat1)),
            'properties': {
                'class': det['class'],
                'confidence': round(det['confidence'], 3),
                'area_m2': round(
                    (x2-x1) * (y2-y1) * det.get('gsd_m', 0.05)**2, 2
                )
            }
        })

    return {'type': 'FeatureCollection', 'features': features}

This module is included in every project—it generates GeoJSON files ready for QGIS or ArcGIS.

Metrics by Task Type

What's Included in the Work

We provide:

• Pipeline documentation (architecture, deployment instructions).
• Trained model with MLflow logs and validation metrics.
• Source code for all modules (Rasterio, SAHI, transformation, web interface).
• Integration with GIS systems via GeoJSON / Shapefile.
• Training of your team (2–3 days remote or on-site).
• Model warranty for 6 months (adjustments for data drift).

Timelines

Order AI system development for your UAVs—we'll evaluate the project in 2 business days. Contact us to discuss your task and provide sample images.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.