A structured world-state (entities, relations, context, and predictive cues) is designed to preserve prediction-critical content when perception degrades, but it presumes observations to populate it; when the primary visual modality is occluded or degraded, those observations may be missing. We address how to sustain the world model from a complementary modality by treating the absence of expected co-evidence as evidence of a hidden cause. The abductive framework is modality-agnostic; this article instantiates it acoustically. A microphone-array front-end estimates the bearing of engine and tire sources and extracts approach-rate evidence (Doppler when a stable tone exists, a broadband looming readout otherwise); the event "signature present, visual co-evidence absent" then triggers abductive inference of a hidden road user, emitting a calibrated risk advisory rather than a control command. Recoverability of the hidden state is analyzed as an identifiability question separating shared from modality-unique information, and cueing is cast as Neyman-Pearson detection under an explicit false-alarm budget. On real occluded-approach recordings at blind junctions, the method warns a mean 1.7 seconds before line-of-sight entry, matches the sustained-window variant of the published acoustic baseline's detection rate with 42% fewer false alarms, localizes to 3.4 degrees median once in view, is well calibrated (expected calibration error 0.034), and keeps hazard awareness above 0.87 under staged vision degradation that collapses a vision-only channel to 0.03. We also measure the method's limits: calibration transfers to an unseen junction almost losslessly, the signature classifier does not, and moving-ego noise is the binding deployment constraint.
Figures & tables
Fig. 1: Overview. (a) At a blind junction, an approaching road user is hidden from every line-of-sight sensor by the building, while its engine/tire noise diffracts around the corner and reaches the roof-mounted microphone array seconds before line-of-sight entry. (b) The acoustic front-end estimates bearing, approach rate, and spectral signature; the visual channel is queried for the co-evidence that the signature class predicts, with its miss rate βc(θ,d) measured under degradation. The joint pattern “signature present, co-evidence absent” drives abductive inference among hidden-source, visible-source, and no-source explanations, and a Neyman–Pearson-calibrated risk advisory r updates the structured world-state.
Fig. 2: Calibrated cueing performance. (a) Window-level ROC of the hazard statistic (occluded-approach windows vs. hazard-free windows) at three pre- T0 horizons: discriminability grows as the hazard nears (AUROC 0.70 over all occluded windows, 0.92 in the last second), the audibility-with-distance structure predicted by Sec. VI . (b) Reliability of the calibrated risk r (isotonic map, recording-level cross-validation; ECE 0.034). (c) Recording-level detection vs. achieved false-alarm rate; the NP-calibrated operating point matches the sustained baseline’s detection at 42% fewer false alarms. (d) Post-line-of-sight hazard-awareness coverage vs. the measured visual miss rate βc(d) under staged degradation (blur/dark/noise): a vision-only channel decays as 1−β by construction, while the fused advisory stays above 0.87 even where vision collapses to 0.03.
Fig. 3: One occluded approach end to end (location SA1, vehicle from the right). Top: camera frames with the vehicle still hidden behind the corner (4.5 s, street empty), emerging shortly after line-of-sight entry T0=7.0 s, and at the pass; dashed boxes mark the vehicle. Bottom: hazard statistic s(t) , NP threshold γ(α=0.02) , and the emitted hidden-hazard advisory (green), which fires 2.3 s before T0 ; the visual detector (red) responds only after T0 . After the pass the statistic drops as the classifier reassigns the now-visible vehicle to the front class.
Method
Lead (s)
Det. (occl.)
FA/h
Bearing ( ∘ )
Vision-only †
0
0
0
(ref.)
Radar FoV model ‡
0
0
—
n/a
OC-SORT recovery ‡
0
0
—
n/a
Acoustic [ 2 ]
2.29
1.00
350
n/a
+ sustained rule
1.88
0.88
185
n/a
Proposed ( α=0.02 )
1.69
0.88
107
3.4
TABLE I: Warning performance on the OVAD static test set (83 recordings: 42 occluded approaches, 41 hazard-free). Lead time is mean seconds before line-of-sight entry ( T0 ); FA/h measured on hazard-free recordings (0.103 h total, so one alarm onset ≈ 9.7 FA/h). Bearing error is the median against the visual reference during line-of-sight (Sec. IX ); occluded-phase bearing is direction-coarse by Corollary 1 .
Peter L. Reichertz Institute for Medical Informatics, Hannover Medical School, Germany · Lower Saxony Center for Artificial Intelligence and Causal Methods in Medicine (CAIMed), Hannover, Germany
Department of Engineering and Architecture, University of Parma, Parma, Italy · Fraunhofer Institute for Digital Media Technology IDMT, Oldenburg, Germany