Ground Vehicle Acoustic Signal Processing Based on Biological Hearing Models
Baras, John S.
MetadataShow full item record
This thesis presents a prototype vehicle acoustic signal classification system with low classification error and short processing delay. <p>To analyze the spectrum of the vehicle acoustic signal, we adopt biologically motivated feature extraction models - cochlear filter and A1-cortical wavelet transform. The multi-resolution representation obtained from these two models is used in the later classification system. <p>Different VQ based clustering algorithms are implemented and tested for real world vehicle acoustic signals. Among them, Learning VQ achieves the optimal Bayes classification performance, but its long search and training time make it not suitable for real time implementation. TSVQ needs a logarithmic search time and its tree structure naturally imitates the aggressive hearing in biological hearing systems, but it has a higher classification error. Finally, a high performance parallel TSVQ (PTSVQ)is introduced, which has classification performance close to the optimal LVQ, while maintains logarithmic search time. <p>Experiments on ACIDS database show that both PTSVQ and LVQ achieve high classification rate. PTSVQ has additional advantages such as easy online training and insensitivity to initial conditions. All these features make PTSVQ the most promising candidate for practical system implementation.<P><P>Another problem investigated in this thesis is combined DOA and classification, which is motivated by the biological sound localization model developed by Professor S. Shamma: the Stereausis neural network. <p>This model is used to perform DOA estimation for multiple vehicle recordings. The angle estimation is further used to construct a spectral separation template. <p>Experiments with the separated spectrum show significant improvement in classification performance. The biologically inspired separation scheme is quite different from traditional beamforming. However, it integrates all three biological hearing models into a unified framework, and it shows great potential for multiple target DOA and ID systems in the future.