Does scheduling onboard perception by the validity of the current picture help a small vessel? An equal-work evaluation on real maritime recordings with measured processing cost

A small autonomous vessel cannot always afford to run its perception pipeline on every sensor frame. A natural idea, with support in simulation studies of event-triggered and self-triggered control, is to run perception only when the current picture of the surroundings is about to stop being valid. We test this idea on real recordings. Using six sequences of the Pohang Canal dataset (about 138,000 front-LiDAR frames), nineteen MUT harbour recordings and nine Autoferry scenarios, we measure the processing cost of fixed LiDAR pipelines, and then replay the recorded voyages with a single non-preemptive perception job whose duration is the measured cost multiplied by a stated slowness factor. The outcome is a detector-relative proxy: how often the held picture reports more than 2 m more clear water ahead than the same detector would report with no delay. On Pohang, the 99th-percentile cost of the 3-D detector used in the replay is 6.8–8.2 times its median, and cost rises near obstructions. The coupling is a slow regime effect, however, and the 99th-percentile cost of the simpler grid detector is 159–314 times shorter than the 5th-percentile time to the nearest obstruction. In a pre-specified comparison at the primary slowness factor, validity-based timing was worse than a fixed rate at matched processor use in all six sequences (median +43%). Allowing for job duration added +3.7% (worse in all six), and predicting job duration instead of using a constant allowance changed the score by +0.2% (better in three of six). Of four published remedies (a maximum gap, an unseen-arrival bound, a cheap novelty cue and short picture memory), the maximum gap, alone and with the cue, passed a frozen decision rule only because the matching rule scored it at 1.06–18 times the intended work; at equal work none beat a fixed rate. Four further fixes, tested with equal-work matching, also failed: retaining recent pictures traded fewer overestimates for many more underestimates, and the only lead, cheaper forward-corridor jobs (−10% without memory, −25 to −31% against full-frame jobs with the same memory), was not selected by the protocol's rule and is confounded with a processor change. We report the result as a bounded negative finding for these detectors, recordings and outcome measure, and we give practical guidance for evaluating perception schedulers at equal work.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24367386
Primary Topic
Maritime Navigation and Safety
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Does scheduling onboard perception by the validity of the current picture help a small vessel? An equal-work evaluation on real maritime recordings with measured processing cost

S. M. Abdulla Al Mamun
Zenodo (CERN European Organization for Nuclear Research)
Maritime Navigation and Safety
preprint

Does scheduling onboard perception by the validity of the current picture help a small vessel? An equal-work evaluation on real maritime recordings with measured processing cost

S. M. Abdulla Al Mamun
preprint en

Abstract

A small autonomous vessel cannot always afford to run its perception pipeline on every sensor frame. A natural idea, with support in simulation studies of event-triggered and self-triggered control, is to run perception only when the current picture of the surroundings is about to stop being valid. We test this idea on real recordings. Using six sequences of the Pohang Canal dataset (about 138,000 front-LiDAR frames), nineteen MUT harbour recordings and nine Autoferry scenarios, we measure the processing cost of fixed LiDAR pipelines, and then replay the recorded voyages with a single non-preemptive perception job whose duration is the measured cost multiplied by a stated slowness factor. The outcome is a detector-relative proxy: how often the held picture reports more than 2 m more clear water ahead than the same detector would report with no delay. On Pohang, the 99th-percentile cost of the 3-D detector used in the replay is 6.8–8.2 times its median, and cost rises near obstructions. The coupling is a slow regime effect, however, and the 99th-percentile cost of the simpler grid detector is 159–314 times shorter than the 5th-percentile time to the nearest obstruction. In a pre-specified comparison at the primary slowness factor, validity-based timing was worse than a fixed rate at matched processor use in all six sequences (median +43%). Allowing for job duration added +3.7% (worse in all six), and predicting job duration instead of using a constant allowance changed the score by +0.2% (better in three of six). Of four published remedies (a maximum gap, an unseen-arrival bound, a cheap novelty cue and short picture memory), the maximum gap, alone and with the cue, passed a frozen decision rule only because the matching rule scored it at 1.06–18 times the intended work; at equal work none beat a fixed rate. Four further fixes, tested with equal-work matching, also failed: retaining recent pictures traded fewer overestimates for many more underestimates, and the only lead, cheaper forward-corridor jobs (−10% without memory, −25 to −31% against full-frame jobs with the same memory), was not selected by the protocol's rule and is confounded with a processor change. We report the result as a bounded negative finding for these detectors, recordings and outcome measure, and we give practical guidance for evaluating perception schedulers at equal work.

Zenodo (CERN European Organization for Nuclear Research)
University of Manchester (GB)
Maritime Navigation and Safety
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon