P2
Preserve fine features
A high-resolution P2 prediction head retains shallow spatial detail needed for very small targets.
YOLO-P2 · ADAPTIVE SAHI · PYTORCH CPU
Small objects.
Wide-area intelligence.
Deployable evidence.
An end-to-end aerial image and video analytics platform engineered to recover small targets, adapt computation to scene complexity, accelerate long recordings and produce evidence that can withstand technical scrutiny.
Detail-preservingP2 head + spatial slicing
Compute-awareAdaptive image and frame policies
Experiment-readyJSON, CSV, heatmaps and reports
AERIAL FRAME · ADAPTIVE ANALYSISPIPELINE READY
ADAPTIVE ROUTECONTENT-AWARE SLICING
GLOBAL FUSIONCLASS-WISE NMS
EVIDENCE OUTPUTREPORT + JSON + CSV
THE CORE PROBLEM
Why ordinary full-frame detection fails in wide-area scenes
Resizing a large aerial frame compresses distant people, vehicles and platforms into a handful of pixels. The evidence disappears before the detector sees it. EdgeSAHI changes both the feature scale and the inference strategy.
ROI
Resolve the right regions
Adaptive overlapping slices magnify local evidence without treating every scene with the same compute budget.
LTF
Control temporal cost
Target-FPS sampling and scene-change override prevent wasteful inference on every video frame.
END-TO-END ARCHITECTURE
From raw media to a defensible analytical product
Each stage exposes its decision, timing and output through the operations console. The objective is not merely to draw boxes, but to create a reproducible inference trail.
THE IMPROVISATION
More than a standard object-detection deployment
The research contribution lies in the orchestration: spatial adaptation, temporal adaptation, quantitative comparison and operational reporting work together around the trained detector.
Conventional deployment
FIXED PIPELINE×One global resize for every image
×Inference on every video frame
×No cross-tile duplicate analysis
×Box visualisation without experimental evidence
EdgeSAHI deployment
ADAPTIVE PIPELINE✓Complexity-driven full-frame or sliced inference
✓Target-FPS policy with scene-change override
✓Coordinate fusion with measurable NMS suppression
✓A/B comparison, analytics and reproducibility artefacts
DUAL-USE VALUE
One capability, two mission environments
The platform provides authorised human-in-the-loop decision support wherever objects are small, sparse, distant or distributed across a large field of view.
DEFENCE & SECURITY
Wide-area situational awareness
Support observation, prioritisation and review without automating operational judgement.
Perimeter monitoringPeople and vehicles across large protected areas
Low-altitude awarenessSmall aerial-object observation support
Coastal surveillanceSmall craft and shoreline activity
Convoy observationVehicle presence and trajectory evidence
Border analyticsWide-area visual change and object cues
Search and rescueHuman and vehicle location assistance
CIVILIAN & PUBLIC SAFETY
Infrastructure-scale visual intelligence
Convert aerial and fixed-camera media into measurable operational information.
Traffic analyticsVehicle density and class distribution
Disaster responseSearch areas and access-route assessment
Industrial safetyPersonnel and vehicle-zone monitoring
Wildlife surveysSmall-object review over large habitats
Infrastructure inspectionDistributed object and anomaly evidence
Crowd safetyWide-view occupancy decision support
MEASURABLE OUTPUT
Built for technical review, not visual theatre
The system explicitly separates inference diagnostics from ground-truth accuracy. mAP, precision and recall remain evaluation metrics; deployment evidence reports what the pipeline actually measured.
Validity boundary: propagated boxes provide display continuity and are never counted as fresh inference. Estimated tracks are not ground-truth identities. Operational decisions remain human-authorised.
Move from detection output to deployable evidence.
Start with one image, inspect the adaptive decisions, then execute the long-video policy.