Vehicle detection and classification using passive multi-modal sensing and TinyML

This research evaluated TinyML for the classifi- cation and localization of military vehicles. A measurement platform with seismic, acoustic, and magnetic capabilities was used for data collection. A multi-modal ML model showed an 87% classi- fication accuracy at 100m with a model size of 527kB. Using quantization, the model was further reduced to 141kB, with 82% classification accu- racy. A microphone array was used for wideband acoustic beamforming, which was interpreted us- ing an ML model, resulting in a mean average er- ror of < 20o with a range of 50m. The a model size of 154kB was reduced to 48kB after quantization, with < 1% increase in error.