Modern ISR (Intelligence, Surveillance, Reconnaissance) systems generate data volumes that fundamentally exceed human processing capacity. A single medium-altitude UAV operating a full-motion video payload generates approximately 2–4 TB of raw video per day at standard resolution, plus associated sensor logs and metadata. A deployed SIGINT collection system may produce terabytes of IQ data per day across its monitored spectrum. The bottleneck in modern ISR is not collection — it is processing and analysis.

The traditional response to this bottleneck is bandwidth: transmit raw data to a ground station and apply analyst labor there. This approach faces three structural constraints in modern operational environments. First, the link budget — satellite and tactical radio links simply cannot carry full-resolution full-motion video from a large UAV fleet continuously. Second, the analyst shortage — there are not enough qualified image analysts to review all collected footage frame-by-frame. Third, the time value of intelligence — by the time raw video reaches a ground station, is queued, and receives analyst attention, the actionable window for time-sensitive targets may have closed.

AI-assisted triage at the edge addresses all three constraints simultaneously. The AI pipeline runs on the collection platform — the UAV itself or the sensor node — and automatically filters the data stream, retaining and transmitting only the portions that contain objects of interest, while discarding or heavily compressing the background of empty terrain, sky, and water that constitutes the majority of raw ISR collection.

The ISR Data Overload Problem

The scale of the data overload problem requires precise framing. Consider a deployed reconnaissance UAV operating an EO/IR dual-sensor payload at 1080p resolution, 30fps, for 16 hours per day. At standard H.264 compression, this generates approximately 50 GB of video per flight. If only 3% of the collected footage contains objects of interest (a generous estimate for wide-area coverage missions), then 97% of the bandwidth and storage budget is consumed by data that will never be actionable. Edge AI triage changes the ratio fundamentally: by detecting and flagging only frames containing detections, the transmission bandwidth requirement drops from 50 GB to approximately 1.5 GB per flight day — within the range of a satellite uplink operating at modest data rates.

SIGINT collection faces an analogous problem. A broadband SDR collection system monitoring a 200 MHz spectrum slice generates several hundred gigabytes of IQ data per hour. Only a small fraction of the monitored spectrum is active at any time, and only a fraction of active signals are of analytical interest. Automated spectrum scanning and signal classification at the edge reduces the downstream processing burden from the full collected bandwidth to only the classified signals of interest — a reduction of two to three orders of magnitude.

Edge Triage Pipeline: Raw Sensor Input to Priority Scoring

The edge triage pipeline for UAV video processing proceeds through four stages:

1. Raw sensor input. Video frames from the EO and/or IR sensor are received at the edge compute hardware. For a real-time processing requirement at 30fps, the compute pipeline must complete one full inference cycle — preprocessing, detection model inference, post-processing, and metadata generation — within 33ms.

2. Object detection. Each frame is processed through a lightweight object detection model (YOLOv8-nano or YOLOv8-small, quantized to INT8) that identifies the presence and location of objects of interest — vehicles, persons, structures, or sensor-specific targets. The detection output is a set of bounding boxes with class labels and confidence scores.

3. Classification and context enrichment. Frames containing detections above a confidence threshold are passed to a secondary classification stage. This stage applies more compute-intensive analysis to the detected objects: vehicle type classification (wheeled vs tracked, civilian vs military profile), activity classification (stationary, moving, grouped), and geospatial annotation (GPS coordinates of detected objects using gimbal and sensor geometry). For multi-object detections, a clustering step identifies whether detected objects form groups consistent with convoy formations or dispersed patrol patterns.

4. Priority scoring. Each annotated detection event is scored for operational priority. Scoring factors include: object class and type (a military vehicle scores higher than a civilian vehicle); confidence score; proximity to previously identified locations of interest (a detection near a previously flagged compound scores higher than a first detection); activity indicators (moving targets typically score higher than stationary); and temporal density (multiple detections of the same object type in a 10-minute window increases priority). The priority score determines whether the event is transmitted immediately, queued for batch transmission, or archived without transmission.

UAV Video Processing: Real-Time Object Detection at 30fps

Achieving sustained 30fps object detection on an embedded GPU requires careful pipeline engineering beyond simply deploying a fast model. The video input must be efficiently decoded and transferred to GPU memory; for H.264/H.265 encoded video streams from gimbal cameras, hardware-accelerated decoding (using the Jetson's NVDEC hardware video decoder rather than software CPU decoding) is essential to avoid consuming the CPU budget needed for control and communications.

NVIDIA's DeepStream SDK provides a GStreamer-based pipeline framework optimized for Jetson that handles hardware-accelerated video decoding, multi-stream support, and efficient GPU memory management for detection model inference. A DeepStream pipeline running YOLOv8-small INT8 on Jetson Orin NX can process four simultaneous 1080p video streams at 30fps within a 15W power budget — enabling quad-sensor payload configurations on medium-class UAVs.

Temporal smoothing is a critical reliability component. A single-frame object detection model produces detections that may flicker — an object detected in frames 1 and 3 but not frame 2 due to confidence threshold variance. A track-based aggregation layer (using ByteTrack or similar) assigns persistent track IDs across frames and applies temporal filtering: only tracks that persist for a minimum number of frames (typically 3–5) and maintain a minimum average confidence score are elevated to triage events. This eliminates single-frame false positives from the triage output without introducing significant latency.

Human-in-the-Loop: AI Escalation Thresholds

The AI triage pipeline is not designed to replace analyst judgment — it is designed to focus analyst attention. The escalation architecture has three tiers:

Automatic transmission. Events scoring above the high-priority threshold (typically a confidence-adjusted combination of object type, activity, and temporal density) are transmitted immediately via the available downlink. The metadata packet — GPS coordinates, object class, confidence score, timestamp, and a representative thumbnail — is approximately 50 KB per event. A system generating 200 high-priority events per flight day requires approximately 10 MB of transmission bandwidth for metadata alone — well within typical satellite link capacity.

Analyst review queue. Events in the medium-priority tier are buffered on-board and transmitted in the next available high-bandwidth transmission window (satellite contact, return to base). The analyst review queue includes both the metadata and a video clip (typically 10–30 seconds around the detection event at reduced resolution) for contextual review.

Archive-only. Low-confidence and low-priority events are archived on the UAV's local storage. If a subsequent high-priority event in the same area triggers retrospective analysis, archived footage from the period before the high-priority event can be reviewed for preceding activity patterns.

Key insight: The bandwidth savings from edge AI triage are not just logistical — they are operationally enabling. A UAV that previously required a high-bandwidth satellite link to maintain continuous intelligence output can now operate effectively on a much narrower link, extending the number of platforms that can be sustained within a given communications architecture by an order of magnitude.

Bandwidth Savings: Transmitting Clips vs Full Video Streams

The quantified bandwidth reduction from edge triage depends on target density in the operational area and detection model sensitivity settings. In low-activity terrain (open desert, forest, ocean), where targets of interest appear in under 1% of frames, edge triage can achieve a 100:1 reduction in transmitted data. In high-activity urban or contested areas where vehicle movement is continuous, the reduction is smaller — perhaps 10:1 — but still significant for link budget management.

A thumbnail-plus-metadata transmission for a detected event averages approximately 50–100 KB. A 30-second video clip at reduced resolution (480p, H.265) averages approximately 5–10 MB. Compared to transmitting full-resolution full-motion video at approximately 2 Mbps (approximately 900 MB per hour), the bandwidth savings for a flight day with 200 triage events are: 200 metadata packets (20 MB) plus 50 medium-priority clips (500 MB) versus 14.4 GB of full video — a 20:1 reduction for this scenario, reducing the required satellite link bandwidth from approximately 2 Mbps continuous to approximately 200 kbps average.