Detection Pipeline Latency Analyzer

Interactive animated pipeline showing every stage from camera frame capture through YOLOv8 inference, MGRS projection, and MANET transmission to COP display update.

Interactive: Detection Pipeline Latency Analyzer

End-to-End Detection Latency from Sensor to Decision Maker

The detection pipeline analyzer measures every millisecond from frame acquisition through target symbology rendering on the operator display on the commander's COP display. The total pipeline latency determines how quickly the brigade can respond to emerging threats. With zero MANET relay hops (direct link), the complete pipeline takes 35 milliseconds — below the threshold of human perceptual latency. With three relay hops (typical brigade-wide deployment), total latency increases to 59 milliseconds. Even at maximum relay depth (7 hops across a 30-kilometer mesh), total latency remains under 100 milliseconds — still effectively real-time for tactical decision-making.

Stage-by-Stage Pipeline Breakdown

The pipeline consists of eight sequential stages, each with measured latency. Camera frame capture from the IMX477 sensor takes 5 milliseconds. YOLOv8 inference on the Jetson Orin Nano (FP16 TensorRT optimized) takes 15 milliseconds — this is the processing bottleneck consuming the greatest computational resources, running the neural network forward pass on a 640×640 pixel input to detect and classify vehicles, personnel, and drones. Pixel-to-MGRS ground projection takes 2 milliseconds — converting the bounding box center pixel coordinates to real-world military grid coordinates using the drone's attitude (roll, pitch, yaw from EKF3) and barometric altitude. Dempster-Shafer fusion with existing observations takes 3 milliseconds. Cursor on Target XML serialization takes 1 millisecond. AES-256 encryption takes 1 millisecond. MANET transmission takes 8 milliseconds per hop. Lisa 26 PostGIS database insertion takes 5 milliseconds. WebSocket push to COP displays takes 3 milliseconds.

MANET Hop Count and Relay Architecture Impact

The configurable MANET hop slider demonstrates the linear relationship between network depth and transmission latency. Each relay hop adds 8 milliseconds as the Silvus StreamCaster receives, processes, and retransmits the packet. At zero hops (drone directly linked to COP), only the local processing stages contribute to latency. At seven hops (maximum realistic brigade deployment spanning 30+ kilometers), MANET transmission accounts for 56 milliseconds — more than all other stages combined. This is why Fischer 26 relay positioning matters: placing the relay drone optimally can reduce hop count from seven to three, cutting total latency by 32 milliseconds. The pipeline animation runs in real-time, with each stage bar expanding sequentially to show the exact timing of each processing step.

Read the full technical documentation →

YOLOv8 Model Selection and Nordic Fine-Tuning

The 15-millisecond inference time assumes YOLOv8n (nano) exported to TensorRT FP16 format running on the Jetson Orin Nano. The nano model was chosen over the larger small (s) and medium (m) variants because the Orin Nano has limited GPU memory (8GB shared) and the inference time for YOLOv8s would exceed 30 milliseconds, doubling the AI stage latency. To compensate for the smaller model size, FSG-A fine-tunes YOLOv8n on a Nordic-specific dataset containing Swedish military vehicles (Strv 122, Strf 90, BV 206), Russian military vehicles (T-72B3, BMP-2, BTR-82A), and Nordic terrain features (snow-covered vehicles, vehicles under spruce canopy, vehicles at -30°C with minimal thermal signature). Fine-tuning on 5,000 annotated images improves detection accuracy from 45 percent (COCO pretrained) to 87 percent on Nordic targets.

Pixel-to-MGRS Coordinate Projection Without GPS

Converting a bounding box center pixel coordinate to a real-world MGRS grid reference requires four inputs: the pixel position (u, v) in the camera frame, the camera intrinsic parameters (focal length, principal point, distortion coefficients), the drone attitude from the EKF3 AHRS (roll, pitch, yaw — available without GPS), and the drone altitude from the BMP390 barometer (±0.5m relative accuracy, immune to electronic warfare). The projection constructs a ray from the camera through the pixel, rotates it by the IMU attitude to world coordinates, and intersects it with a horizontal ground plane at the terrain elevation beneath the drone. Without GPS, the absolute position has 50-200 meter uncertainty depending on visual navigation drift, but the relative positions of detections within a single flight are accurate to 5-10 meters because the IMU drift is consistent across all projections from the same flight.

Implementation

# Detection Pipeline Latency — Stage Breakdown
STAGES = [
    ("IMX477 capture",      5),   # ms
    ("YOLOv8 FP16",        15),   # Jetson Orin Nano
    ("Pixel→MGRS",          2),   # coordinate projection
    ("DS fusion",           3),   # Dempster-Shafer
    ("CoT XML",             1),   # serialization
    ("AES-256",             1),   # encryption
    ("MANET TX",            8),   # per hop
    ("PostGIS insert",      5),   # database
    ("WebSocket push",      3),   # COP update
]

def pipeline_latency(manet_hops=3):
    """Total latency from camera to COP display."""
    total = 0
    for name, ms in STAGES:
        actual = ms * manet_hops if "MANET" in name else ms
        total += actual
        print(f"  {name:20s} {actual:4d}ms  (cumulative: {total}ms)")
    return total

for hops in [0, 3, 7]:
    print(f"\n=== {hops} MANET hops ===")
    t = pipeline_latency(hops)
    print(f"  TOTAL: {t}ms — {'REAL-TIME' if t < 100 else 'NEAR RT'}")

Sources

• ArduPilot — ardupilot.org
• FOI — FOI publications catalogue (foi.se/publications)