Semantics First, Pixels Later: Mission-Driven Event Streams for Drone Video over Bandwidth-Starved Links
Organizations: Independent Researcher
Abstract
A surveillance drone on a bandwidth-starved link cannot send the video its operator would need, yet the operator's questions are often not about pixels: did three people converge, when, and where? We study what should cross a link of 1 Mbps down to 1 kbps instead, and in what order. BPS, a behavior-primitive stream, sends semantics first and pixels later: it runs a detector and a tracker onboard and sends short records (25 bytes on average) of multi-frame events selected by a mission card of 16-125 bytes, ahead of positions and pictures. We measure it by what a ground-side language model answers from it, under a preregistered protocol with a held-out set, against H.265 read by the vision fine-tune of the same model, H.265 on which the ground runs our own perception, a stream of quantized tracks, and onboard captions. On the development set (84% of questions about convergence, mostly one location) the event stream answers 74.3% of questions at 0.13 kbps. A reader that receives nothing answers 57.4%, and the video reader 59.6% at 100 kbps and 64.7% on the source frames. When the ground instead runs our perception on the video, video matches the stream at 100 kbps, about 800 times the stream's rate, and falls behind at 64 kbps and below. A track stream matches it at 96 times the bytes and loses 0.03 at 17 times. On held-out data the result is weaker: with events taken from annotated tracks, an oracle, the stream beats H.265 at 100 kbps (+0.13), but with onboard perception it does not (+0.01 [-0.04, +0.06]) and is no better than receiving nothing. Sent alone through emulated links the event records answer 0.739 at 1 Mbps and still 0.724 at 0.3 kbps; an every-frame track stream holds down to 1 kbps and video with ground-side perception down to about 100 kbps. On a Jetson Orin Nano the unchanged pipeline runs at 4.5 fps and misses most events.
Figures & tables
| primitive | statistic in the window (canonical pixels, 30 fps) |
|---|---|
| dwell | tracks with speed px/frame and net displacement px, among moving neighbors |
| density change | subject-class tracks present in the region |
| birth after quiet | new tracks after frames at or below the 25th-percentile occupancy |
| flow deviation | tracks moving px/frame at to the baseline flow direction, when that flow is coherent (mean resultant length ) |
| approach | largest group of tracks whose mutual distances close by and px, ending within 120 px |
| persistent change | an object that appears and stays (synthetic insertions; no onboard detector) |
| row | kbps | accuracy [95% CI] | event | nothing | bold row row |
|---|---|---|---|---|---|
| no-stream reader (question only) | – | 0.574 [0.546, 0.602] | 0.349 | 0.826 | +0.169 [+0.134, +0.204] |
| video reader on | |||||
| source frames (no codec) | source | 0.647 [0.606, 0.687] | 0.597 | 0.702 | +0.096 [+0.049, +0.145] |
| H.265 t1000 | 998.0 | 0.645 [0.608, 0.685] | 0.558 | 0.743 | +0.098 [+0.050, +0.146] |
| H.265 t100 | 102.9 | 0.596 [0.558, 0.635] | 0.592 | 0.601 | +0.147 [+0.100, +0.195] |
| H.265 t10 | 10.8 | 0.583 [0.548, 0.616] | 0.596 | 0.568 | +0.160 [+0.118, +0.207] |
| row | accuracy [95% CI] | event † | nothing † | row t100 |
|---|---|---|---|---|
| no-stream reader | 0.626 [0.574, 0.679] | 0.367 | 0.923 | – |
| H.265 t1000 | 0.647 [0.585, 0.707] | 0.610 | 0.688 | – |
| H.265 t100 | 0.603 [0.540, 0.668] | 0.619 | 0.585 | – |
| H.265 t10 | 0.531 [0.473, 0.598] | 0.572 | 0.485 | – |
| region-preserving t100 | 0.566 [0.494, 0.646] | 0.604 | 0.523 | -0.037 [-0.068, -0.003] |
| captions, generic | 0.435 [0.383, 0.492] | 0.435 | 0.436 | -0.168 [-0.223, -0.111] |
| link rate (kbps) | 1000 | 100 | 10 | 1 | 0.3 |
|---|---|---|---|---|---|
| BPS event records | 0.739 | 0.739 | 0.741 | 0.726 | 0.724 |
| track stream, every frame | 0.728 | 0.729 | 0.724 | 0.712 | 0.638 |
| track stream, every 6th frame | 0.722 | 0.722 | 0.720 | 0.695 | 0.638 |
| H.265, ground-side perception | 0.726 | 0.726 | 0.651 | – | – |
| H.265, video reader | 0.645 | 0.596 | 0.587 | – | – |
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
| family | items | primitive | items |
|---|---|---|---|
| nothing happened | 771 | approach | 1,377 |
| localization | 372 | birth after quiet | 93 |
| presence | 180 | dwell | 66 |
| counting | 150 | persistent change | 51 |
| timing | 150 | density change | 36 |
| ordering | 15 | access-rate change | 15 |
| family ( ) | no stream | source frames | H.265 t100 | region-pres. t100 | t100 ground | t1000 ground | captions gen. | captions miss. | tracks | BPS causal | BPS real | BPS perfect |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| presence (180) | 0.30 | 0.63 | 0.66 | 0.74 | 0.73 | 0.72 | 0.61 | 0.71 | 0.74 | 0.72 | 0.73 | 0.84 |
| localization (372) | 0.44 | 0.67 | 0.65 | 0.67 | 0.68 | 0.72 | 0.42 | 0.46 | 0.74 | 0.74 | 0.75 | 0.85 |
| counting (150) | 0.30 | 0.40 | 0.43 | 0.44 | 0.36 | 0.35 | 0.25 | 0.20 | 0.36 | 0.35 | 0.37 | 0.65 |
| timing (150) | 0.21 | 0.56 | 0.49 | 0.50 | 0.61 | 0.64 | 0.31 | 0.47 | 0.66 | 0.61 | 0.66 | 0.83 |
| ordering (15) | 0.73 | 0.80 | 0.80 | 0.80 | 0.73 | 0.87 | 0.73 | 0.80 | 0.80 | 0.93 | 0.80 | 0.73 |
| nothing happened (771) | 0.83 | 0.70 | 0.60 | 0.53 | 0.84 | 0.82 | 0.56 | 0.43 | 0.81 | 0.84 | 0.83 | 0.79 |
| row | accuracy [95% CI] | row t100 | byte reduction |
|---|---|---|---|
| no-stream reader | 0.439 [0.353, 0.600] | – | – |
| H.265 t1000 | 0.711 [0.600, 0.794] | – | – |
| H.265 t100 | 0.702 [0.600, 0.735] | – | – |
| H.265 t10 | 0.500 [0.235, 0.614] | – | – |
| BPS, perfect perception | 0.649 [0.500, 0.900] | -0.053 [-0.235, +0.300] | 119 |
| BPS, real perception | 0.658 [0.529, 0.900] | -0.044 [-0.206, +0.300] | 137 |
| scope | seq. | items | perfect | real | causal | t1000 | t100 | no stream | real t100 | real perfect |
|---|---|---|---|---|---|---|---|---|---|---|
| held-out templates (primary) | 27 | 832 | 0.737 | 0.614 | 0.588 | 0.647 | 0.603 | 0.626 | +0.011 [ 0.038, +0.063] | 0.123 [ 0.174, 0.076] |
| development-style templates | 27 | 1248 | 0.751 | 0.664 | 0.640 | 0.639 | 0.586 | 0.533 | +0.079 [+0.028, +0.134] | 0.087 [ 0.124, 0.052] |
| held-out, without the 20-px rule | 27 | 826 | 0.718 | 0.619 | 0.594 | 0.625 | 0.574 | 0.611 | +0.045 [ 0.005, +0.093] | 0.099 [ 0.148, 0.053] |
| held-out, pooled with round 1 | 30 | 946 | 0.726 | 0.619 | 0.605 | 0.654 | 0.615 | 0.604 | +0.004 [ 0.042, +0.056] | 0.107 [ 0.158, 0.062] |
| development-style, pooled with round 1 | 30 | 1419 | 0.738 | 0.662 | 0.652 | 0.646 | 0.598 | 0.532 | +0.063 [+0.011, +0.119] | 0.076 [ 0.112, 0.043] |
| card | seq. | items | no stream | t100 | t1000 | perfect | real | causal |
|---|---|---|---|---|---|---|---|---|
| round 1, card 1 | 1 | 28 | 0.54 | 0.96 | 0.89 | 0.75 | 0.75 | 0.79 |
| round 1, card 2 | 3 | 86 | 0.41 | 0.62 | 0.65 | 0.62 | 0.63 | 0.71 |
| round 1, card 3 (derived, excluded) | 2 | 34 | 0.59 | 0.41 | 0.53 | 0.91 | 0.71 | 0.71 |
| round 2, card 1 | 19 | 322 | 0.59 | 0.60 | 0.70 | 0.68 | 0.54 | 0.53 |
| round 2, card 2 | 5 | 122 | 0.61 | 0.70 | 0.74 | 0.80 | 0.61 | 0.60 |
| round 2, card 3 | 5 | 58 | 0.60 | 0.53 | 0.57 | 0.81 | 0.72 | 0.67 |
| family | items | no stream | t100 | t1000 | perfect | real | real no stream | real t100 |
|---|---|---|---|---|---|---|---|---|
| counting | 80 | 0.175 | 0.375 | 0.362 | 0.775 | 0.275 | +0.100 [-0.026, +0.223] | -0.100 [-0.267, +0.054] |
| localization | 184 | 0.386 | 0.565 | 0.592 | 0.821 | 0.549 | +0.163 [+0.058, +0.278] | -0.016 [-0.142, +0.144] |
| nothing happened | 388 | 0.923 | 0.585 | 0.688 | 0.619 | 0.662 | -0.260 [-0.349, -0.169] | +0.077 [-0.037, +0.185] |
| presence | 100 | 0.680 | 0.840 | 0.760 | 0.850 | 0.690 | +0.010 [-0.104, +0.130] | -0.150 [-0.275, -0.022] |
| timing | 80 | 0.125 | 0.713 | 0.713 | 0.938 | 0.775 | +0.650 [+0.500, +0.770] | +0.062 [-0.066, +0.200] |
| event questions | 444 | 0.367 | 0.619 | 0.610 | 0.840 | 0.572 | +0.205 [+0.142, +0.273] | -0.047 [-0.115, +0.024] |
| what crosses the link | kbps | events | recall | precision | accuracy [95% CI] | event questions |
|---|---|---|---|---|---|---|
| BPS event stream (onboard events) | 0.126 | 107 | 0.67 | 0.79 | 0.743 [0.717, 0.770] | 0.667 |
| H.265 t1000 decoded, ground-side perception | 998.0 | 113 | 0.68 | 0.75 | 0.726 [0.700, 0.755] | 0.644 |
| H.265 t100 decoded, ground-side perception | 102.9 | 98 | 0.58 | 0.74 | 0.727 [0.690, 0.767] | 0.625 |
| H.265 h60 decoded, ground-side perception | 63.6 | 80 | 0.42 | 0.66 | 0.692 [0.657, 0.731] | 0.539 |
| H.265 h30 decoded, ground-side perception | 32.6 | 48 | 0.25 | 0.65 | 0.668 [0.638, 0.701] | 0.451 |
| H.265 q30 decoded, ground-side perception | 33.3 | 57 | 0.26 | 0.58 | 0.680 [0.657, 0.706] | 0.469 |
| what the text reader gets | events | recall | precision | accuracy [95% CI] | event | nothing | onboard stream row | rule-based |
|---|---|---|---|---|---|---|---|---|
| no stream (question only) | – | – | – | 0.626 [0.574, 0.679] | 0.367 | 0.923 | -0.012 [-0.068, +0.041] | – |
| H.265 t100, video reader | – | – | – | 0.603 [0.540, 0.668] | 0.619 | 0.585 | +0.011 [-0.038, +0.063] | – |
| BPS event stream, onboard events | 70 | 0.58 | 0.76 | 0.614 [0.564, 0.667] | 0.572 | 0.662 | – | 0.704 |
| H.265 t1000 decoded, ground-side perception | 71 | 0.55 | 0.72 | 0.599 [0.551, 0.645] | 0.570 | 0.631 | +0.016 [-0.026, +0.061] | 0.683 |
| H.265 t100 decoded, ground-side perception | 35 | 0.24 | 0.63 | 0.546 [0.500, 0.596] | 0.468 | 0.634 | +0.069 [+0.018, +0.114] | 0.601 |
| H.265 t30 decoded, ground-side perception | 7 | 0.07 | 0.86 | 0.512 [0.465, 0.566] | 0.419 | 0.619 | +0.102 [+0.053, +0.153] | 0.531 |
| link | plain H.265 | low-latency H.265 | BPS event records | captions | |||
|---|---|---|---|---|---|---|---|
| age 1 s | median age | age 1 s | median age | 0.5 s | median | median | |
| 1 Mbps | 55% | 0.7 | 100% | 0.3 | 100% | 0.00 | 0.99 |
| 100 kbps | 39% | 1.3 | 100% | 0.4 | 100% | 0.00 | 1.00 |
| 10 kbps | 1% | 4.3 | 98% | 0.5 | 100% | 0.10 | 1.06 |
| 1 kbps | 0% | 20.0 | 0% | 5.9 | 27% | 0.66 | 1.73 |
| onboard rate | event recall | precision | accuracy [95% CI] |
|---|---|---|---|
| 30 fps (all frames) | 0.69 | 0.82 | 0.739 [0.712, 0.767] |
| 15 fps (every 2nd frame) | 0.58 | 0.78 | 0.719 [0.690, 0.750] |
| 10 fps (every 3rd frame) | 0.53 | 0.70 | 0.711 [0.679, 0.747] |
| 5 fps (every 6th frame) | 0.34 | 0.78 | 0.684 [0.652, 0.721] |
| 4.3 fps (every 7th frame; Orin Nano) | 0.28 | 0.86 | 0.675 [0.644, 0.708] |
| ours | low-latency H.265 alone | |||||
|---|---|---|---|---|---|---|
| link | visible / covered | update (s) | age (s), PSNR | visible | new frame (s) | age (s), PSNR |
| 1 Mbps | 0.67 / – | 0.033 | 0.07, 31.7 | 0.52 | 0.033 | 0.27, 27.3 |
| 100 kbps | 0.49 / – | 0.045 (0.038) | 0.07, 29.2 | 0.33 | 0.032 | 0.37, 25.6 |
| 10 kbps | 0.02 / 0.67 | 0.125 (0.064) | 0.27, 23.7 | 0.02 | 0.51 | 0.50, 23.5 |
| 1 kbps † | 0.00 / 0.10 | 2.30 (0.64) | 2.8, 20.0 | 0.00 | 1.26 | 6.0, 18.4 |
Semantic Communication
Semantic communication sends meaning- or task-relevant information so receivers can act without reconstructing every transmitted bit or signal.
See all 116 papers