Silent Object Localization on a Grid Using RIR-Based Acoustic Fingerprints and EfficientNet

This paper investigates acoustic fingerprinting for passive localization of silent objects, motivated by device-free indoor monitoring scenarios where tags, radios, or cameras are undesirable. While radio-frequency fingerprinting and deeplearning-based sound source localization have been widely studied, grid-based localization of non-emitting objects using acoustic fingerprints remains underexplored. We propose a featureextraction pipeline that treats swept-sine-based room impulse responses (RIRs) as an acoustic analogue of channel state: multichannel recordings from a four-microphone array deconvolved into linear RIRs, transformed by short-time Fourier analysis, and converted into stacked log-magnitude and phase tensors. The tensors then serve as inputs to a light EfficientNet-based convolutional neural network for grid classification. In a controlled single-room evaluation on a 4 × 4 grid over a 3.2 × 3.2 m area, the approach achieves up to 97.7% top-1 accuracy across 16 classes, with misclassifications predominantly between neighboring cells. This work is positioned as a baseline feasibility study establishing a reproducible measurement-to-learning pipeline. Although it offers a viable path toward practical silent-object localization, several limitations remain, serving as a foundation for future research.