Communication signals double as radar: multistatic terahertz echoes identify object shape, holding at three-quarters occlusion

Indoor object shape can be read from the echoes of the communication link itself, with no sensing hardware beyond that link. Most terahertz sensing pursues reconstruction, recovering a spatially resolved image before reading geometry and material; here the observation is instead an opportunistic multistatic echo of ordinary communication traffic. Under a protocol whose training and test partitions occupy disjoint pose bins, accuracy rises with transmitter count from a low single-transmitter level to a higher multistatic range, so the gain cannot come from training set size or from memorizing orientation.

Shape identifiability rests on angular diversity rather than on spectral slope. In the physical-optics mechanism hierarchy, six physically distinct mechanisms occupy only three frequency exponents, so a spectral slope alone cannot determine which mechanisms a target presents. Cube, cone and cylinder produce clearly distinct bistatic patterns despite overlapping spectral slopes, and a single transmitter samples one direction while a multistatic constellation samples many.

The learned model wins clearly in two regimes, while the physics-derived baseline matches it on metal under quasi-uniform placement. The physics-informed RBF-SVM uses magnitude-spectrum summary features such as the log-log slope, aggregated across links in a permutation-invariant way. Under random placement the SVM plateaus from about four transmitters onward while the CNN keeps improving and opens a gap at higher transmitter counts; on metal the two end within four points, whereas on plastic and wood the gap widens with each added transmitter.

Results cover shape and material classification of a cube, cone and cylinder in an indoor multistatic terahertz setting, with training and test poses in disjoint bins so that reported accuracy measures generalization to unobserved orientations. The occlusion study reports two training conditions, matched and trained only on unoccluded targets, the latter being the deployment-relevant case. Multi-object scenes hold up to three objects and pose identification as three independent binary presence tests that share one input tensor. The authors propose next steps: a magnitude-only CNN ablation to isolate the phase contribution, a bandwidth sweep as the direct test of the degeneracy thesis, experimental validation at the stated band, and extension to shapes that share a dominant scattering mechanism.

All results derive from a single forward model that has not been checked against full-wave simulation, ray tracing or measurement. The framework assumes direct-path suppression, multistatic phase coherence and coarse object localization. Surfaces are treated as optically smooth, whereas at millimeter wavelengths a roughness factor would attenuate the coherent return most strongly for the lossy materials that are already hardest here. Part of the gap between the CNN and the SVM reflects the CNN's access to complex phase that the SVM never sees, and the authors leave separating those two effects open.