Aerial Person Detection under Edge Constraints: Held-Out Evaluation and Edge-Ground Verification
Organizations: Numberz.ai Inc., Huntsville, AL, USA.
Abstract
A search drone sees a missing person as a few pixels, carries a small computer, and shares a narrow radio link. We study person detection under those limits; the primary evaluations were declared before they ran, and later analyses are marked exploratory. In an exploratory comparison, holding out recording blocks inferred from thermal frame identifiers cut recall at a fixed false-alarm target from 32.7% to 12.3%. With nothing tuned on test blocks, a 46,000-parameter detector gated by a contrast cue finds 10.4% of people against 5.7% for the same network applied everywhere. A standard detector, YOLO11n, trained under the same folds finds 54.5% at a similar false-alarm rate for about eleven times the arithmetic, and on our 4K field footage at 117 m has a walking person among its top 16 detections in up to 92% of frames against 74% for our board. Run alone, YOLO11n's thresholds set at one location overshot their false-alarm target 1.7 to 3.4 times at new locations, and its best configuration at altitude cost 4 to 24 standard passes per frame. Re-scoring its 16 best crops (154 KB per frame) with an 842,000-parameter ground model, thresholds frozen on a validation location, raised recall from 44.9% to 53.7% against the same 16 crops with YOLO's own scores; in a post hoc paired analysis over 15 test locations the gain was +8.8 points [+2.4, +12.5]. Pooled false detections fell from 0.86 to 0.64 per frame, but that paired change was not distinguishable from zero. Coverage of a still person in earth tones was very low for both detectors in the clips we flew.
Figures & tables
| Work | Setting | Reported result |
|---|---|---|
| Schedl et al. [ 16 ] | Thermal AOS, about 35 m, 1 m/s, Pi 4 on board | 38 of 42 hidden people found; one false alert; about 6 ha over 8 flights |
| Manzini and Murphy [ 10 ] | Colour video from a real search, offline | Benchmark-strong detector gave no useful lead in the field |
| Nathan et al. [ 13 ] | Colour, about 300 m, 4 cm/px, offline | Anomaly detectors below 3.5% precision; YOLOv12 failed |
| This paper | Colour, 9 to 117 m and public sets, Pi 5 | Held-out and frozen-threshold evaluation; ground verification +8.8 points recall over the same candidates; still person in earth tones 0 to 7.5% coverage |
| Component | Data that informed it |
|---|---|
| Thermal cue settings | Nine settings fixed in advance; one chosen per fold from that fold’s training and validation blocks only. |
| Thermal networks, YOLO11n | Training blocks or locations of each fold. |
| Thresholds | Validation blocks or location of each fold, then frozen. |
| Memory settings | Validation blocks (memory was not re-run under the per-fold cue). |
| Colour cue (HERIDAL, field) | Fixed during work on a different dataset [ 9 ] , before any HERIDAL frame was used. |
| HERIDAL test-curve read-out | Chosen after the first run; reported as secondary. |
| Split | Gated | Dense |
|---|---|---|
| Published split (blocks shared) | 32.7% (0.11) | 25.2% (0.12) |
| Inferred blocks held out, five folds | 12.3% (0.18) | 6.5% (0.17) |
| Shuffled-label control, held out | 0.0% (0.07) | 0.0% (0.02) |
| Arm | Recall [95%] | False alarms/frame | GMAC/frame |
|---|---|---|---|
| Cue alone | 1.4% | 0.16 | 0 |
| Dense network | 5.7% [3.1, 9.6] | 0.17 | 1.43 |
| Gated, cap 128 | 10.4% [6.6, 15.1] | 0.19 | 0.28 |
| Gated + exploration (deployed) | 10.5% | 0.20 | 0.43 |
| YOLO11n, same folds | 54.5% [49.9, 58.9] | 0.14 | 3.2 |
| Arm | Target 0.5 | Target 1.0 |
|---|---|---|
| Our board, all candidates | 50.9% [45.9, 58.2] at 2.24 | 59.8% [53.9, 65.5] at 3.61 |
| Our board, top 16, board scores (exploratory) | 47.5% at 1.92 | 53.4% at 2.79 |
| Our board, same 16, ground scores | 46.7% [40.4, 61.3] at 0.80 | 51.7% [45.1, 66.8] at 1.32 |
| YOLO11n, all detections | 45.1% [38.0, 61.0] at 0.87 | 53.1% [46.7, 67.0] at 1.68 |
| YOLO11n, top 16, YOLO scores | 44.9% [37.7, 61.0] at 0.86 | 52.5% [45.9, 67.0] at 1.63 |
| YOLO11n, same 16, ground scores | 53.7% [46.2, 66.3] at 0.64 | 61.6% [55.4, 74.4] at 1.15 |
| Mode | Above idle (W) | Whole board per frame (J) | Frames/s at full speed |
|---|---|---|---|
| Cue only | 0.05 | 1.16 | 50.7 |
| Dense network | 0.43 [0.40, 0.50] | 1.36 | 15.0 |
| Gated | 0.28 [0.25, 0.30] | 1.28 | 16.4 |
| Height ( ) | Our board | W640 | W1280 | Tiles T | Board-chosen G | COCO, best |
|---|---|---|---|---|---|---|
| 117.5 m (109) | 74% | 1.8% | 89.0% | 91.7% | 79.8% | 15.6% |
| 89.2 m (8) | 50% | 37.5% | 50.0% | 87.5% | 66.7% | 12.5% |
| 60.3 m (26) | 40% | 23.1% | 46.2% | 50.0% | 53.8% | 17.9% |
| 9.3 m (52) | 92% | 69.2% | 17.3% | 3.8% | 28.2% | 88.5% |
| Passes per frame | n/a | 1 | 4 | 24 | 9–11 |
| Measured | Implies |
|---|---|
| 154 KB of crops per frame | About 2.46 Mbit/s per drone before protocol overhead; about 9.9 Mbit/s for four drones. |
| 0.64 false detections per frame | About 77 frame-level false detections per minute per drone. One background object can repeat across frames, so this is not a count of distinct operator alerts. |
| 16 crops at 0.43 GMAC | About 6.9 GMAC per frame on the ground computer. |
| Top 16 instead of all detections | 10 points less candidate coverage (81% to 71%). |
Object Detection
Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.
See all 673 papers