Aerial Change Detection & Military Object Recognition

01 Executive summary

A single drone sortie produces tens of minutes to hours of high-resolution video. The bottleneck is no longer flying — it is the analyst who must watch every frame to find the one thing that matters. Drone Intelligence automates that work with two complementary capabilities running on a shared imaging pipeline:

• Change Detection — compare two flights over the same route, flown days, weeks or months apart, and surface what changed: new construction, demolitions, encampments, damaged structures, new vessels.
• Military Object Detection — inventory every object of interest in a single clip — personnel, vehicles, vessels, aircraft — with per-object tracking so each is counted once.

Both modules are deployed and operational today on a password-protected HTTPS console, GPU-accelerated, with an automatic CPU fallback. The system is designed to be installed and run entirely on the unit's own machines — no data leaves the building.

02 The operational problem

Finding "what changed" or "what to watch" in raw footage still depends on an experienced operator comparing imagery by eye. In practice this means:

The opportunity is to convert "watch the whole tape" into "review a short, ranked list of what the AI found" — moving the analyst from searcher to decision-maker.

03 System overview

Both capabilities sit on one shared, five-stage imaging pipeline, so investment in one improves the other.

04 Change detection

Feed two clips of the same route at different times; the system aligns them and reports the meaningful differences, ranked by confidence and tagged with GPS coordinates pulled from telemetry.

Capabilities

Object-level change attribution

Beyond where something changed, the system identifies what changed. Every change region is cross-examined with the aerial object detector (Section 05) on both the before and after frame and classified:

The final report leads with a tally of added / removed / modified / unknown changes and the identity of each changed object (e.g. "2× added car, 1× removed van"), so an analyst sees the operational meaning of a flight-to-flight comparison at a glance.

Robustness — engineered to suppress false alarms

• Histogram matching (CIELAB) neutralises seasonal and lighting differences between flights.
• Shadow masking ignores dark low-information regions that otherwise trip naïve diffs.
• Robust two-flight alignment — SIFT features fit a stable similarity model (not an over-fitting homography); pairs are kept only if the alignment is geometrically valid, overlaps ≥70%, and the imagery actually correlates, so mis-paired frames from two separate real flights are rejected instead of producing false changes (typical residual 1–2 px).
• Area & morphology filtering plus a dynamic floor scaled by alignment error removes speckle.
• Per-frame detection cap + probability-map blur keep noisy frames from flooding the report.

Detection models

The fallback design means the platform always produces a result; quality scales up the moment a deep model's weights are dropped in — no code change.

Deliverables

05 Object detection

Point the system at a single clip and it inventories every object of interest, frame by frame, then tracks each object across frames so it is counted exactly once.

How it works

• Aerial-trained detector: the default model is trained on VisDrone (low-altitude drone imagery), so it recognises top-down vehicles, vans, buses and people that generic ground-level (COCO) and satellite (DOTA) models miss entirely. The detector is a drop-in — a fine-tuned military model (tanks, naval classes, etc.) loads via a single environment variable with no other change.
• Tiled (sliced) inference: each frame is split into overlapping tiles that are each analysed at near-native resolution, so a car that is only ~80 px in a 4K frame becomes large enough to detect reliably. This is the single biggest driver of small-object recall in nadir footage.
• Per-category confidence floors: the vehicles/vessels/aircraft that matter most are kept at a sensitive threshold, while the noisiest class (aerial "pedestrian" over textured ground) must clear a high bar — turning a noisy raw output into a clean inventory.
• Tracking: a persistent ID is assigned to each object across frames, so "9 detections of one truck across 9 frames" is reported as one truck, not nine; objects seen only once are discarded as flicker.
• Taxonomy: raw classes are mapped into a military taxonomy — personnel · vehicle · vessel · aircraft — see Section 06.
• Compute: runs on GPU when available, falls back to CPU automatically.

Deliverables

06 Supported object types

The shipped aerial detector (VisDrone-trained) recognises the following drone-view classes and maps them to military categories. Colours match the annotated video output.

Extending to dedicated military classes

The taxonomy layer already understands military terminology, so a specialised model trained on military classes is bucketed automatically. Recognised keywords include:

Drop a fine-tuned .pt model in and set DRONE_OBJDET_MODEL; the inventory, tracking, annotation and API all work unchanged. iApp's synthetic-data generator (Section 11) produces the labelled aerial training data to build exactly these models for the Thai context.

07 Accuracy & how we evaluate

We measure against ground truth, not impressions. A 19-scenario benchmark spanning naval harbours, the Thai–Cambodia border, deep-south towns and riverine Bangkok — 42 ground-truth changes in all — ships with the system, each with a machine-authored solution.json. In addition, three real DJI passes over one route (real_footage/) validate the pipeline on genuine flight imagery.

Evaluation method

Detections are reported in clip-frame coordinates (after oblique warp, camera shake and crop), while ground truth is in base-map coordinates. Our harness projects the ground-truth label map through the identical render functions used to synthesise each clip, so detections and truth land in one coordinate space — no hand-derived geometry that could flatter the score. Because each change's ground-truth box is the actual set of pixels that changed (from the AI-graft step), the labels are exact by construction. For every change we record the frames it is visible in, the frames it was detected in, and we tally false positives per frame.

Measured results — 19-scenario benchmark, baseline backend

Figures below are the always-available classical baseline (no deep weights on the box) on the high-oblique clips — the system's floor, not its ceiling — aggregated over all 42 changes.

Reference benchmarks of the underlying models

These are published, third-party-verifiable figures for the model architectures we deploy. Fine-tuning on the unit's own footage is expected to raise field accuracy further.

08 Security & deployment

The current reference deployment runs as a managed system service behind nginx with automatic TLS renewal, isolated from other services on the host. Uploaded footage and results stay in a per-job directory on the unit's machine.

09 API reference

Every capability is available over a documented HTTP API, so the platform can be driven from the unit's own tooling. All endpoints are under https://drone.iapp.co.th; authentication is a signed session cookie obtained from /api/login. Examples use curl.

curl -c cookies.txt -X POST \
  -F username=iapp -F password=•••• \
  https://drone.iapp.co.th/api/login
# → {"ok": true}   (cookie stored in cookies.txt)

curl -b cookies.txt https://drone.iapp.co.th/api/scenarios
# → {"scenarios":[{"name":"bangkok_chaophraya","variants":["flylow","flyhigh"], ...}]}

# A) on a bundled scenario
curl -b cookies.txt -X POST \
  -F scenario=new_construction -F variant=flylow \
  https://drone.iapp.co.th/api/run/change
# B) on your own footage (before/after)
curl -b cookies.txt -X POST \
  -F video_a=@flight_A.mp4 -F video_b=@flight_B.mp4 \
  https://drone.iapp.co.th/api/run/change
# → {"job_id":"f0787d067de3"}

curl -b cookies.txt -X POST \
  -F video=@patrol.mp4 \
  -F sample_fps=4 -F conf=0.30 -F include_other=false \
  https://drone.iapp.co.th/api/run/objects
# → {"job_id":"81c579396c8a"}
# fields: video | odclip=<id> | scenario=<name> (+variant, clip), sample_fps, conf, include_other

curl -b cookies.txt https://drone.iapp.co.th/api/jobs/81c579396c8a
# → {"status":"running","progress":0.61,"message":"Scanning 22.0s / 29s — 7 objects so far"}
# status: queued | running | done | error

curl -b cookies.txt https://drone.iapp.co.th/api/jobs/81c579396c8a/result
# objects → {"result":{"summary":{"n_objects":10,"by_category":{"vehicle":9,"vessel":1},
#            "by_class":{"car":6,"train":2,"truck":1,"boat":1},"device":"GPU"},
#            "tracks":[...], "annotated_video":"od/annotated.mp4"}}
# change  → {"result":{"n_changes":56,"changes":[{"bbox":[x,y,w,h],"confidence":0.97,
#            "polarity":"new","t_seconds_b":...}], "overlay_video":"cd/overlay.mp4"}}

curl -b cookies.txt -o annotated.mp4 \
  "https://drone.iapp.co.th/api/jobs/81c579396c8a/file?path=od/annotated.mp4"
# change-detection artefacts: cd/overlay.mp4 · cd/report.html · cd/overlays/pair_XXXX.jpg

Jobs run asynchronously: submit, then poll /api/jobs/{id} until status:"done", then read /result and pull artefacts via /file. The same flow powers the web console.

10 Use cases

11 Roadmap

12 Honest disclosure

We state status plainly — that is how trust is earned for real missions.

• The default change-detection backend on the current box is the classical baseline (no deep weights installed). It catches large structural changes reliably with near-zero false alarms, but misses small clustered objects. Installing ChangeFormer weights is a drop-in upgrade.
• The shipped object detector is a VisDrone-trained aerial model with tiled inference — it reliably finds top-down vehicles on real low/medium-altitude footage that generic (COCO) models miss. It has no naval/aircraft-specific classes yet and can over-trigger "pedestrian" on textured ground (held back by a high per-category confidence floor); a military-fine-tuned model (Section 06/11) is the intended production path.
• The benchmark scenarios are synthetic, generated from real geography for repeatable, perfectly-labelled evaluation. Field accuracy is confirmed by fine-tuning on the unit's real footage.

13 Contact

iApp Technology Co., Ltd.
Kobkrit Viriyayudhakorn — Founder & CEO
kobkrit@iapp.co.th

Prepared for the Royal Thai Navy Drone Unit. Confidential — for authorised partners and customers. © iApp Technology.