Low-speed targets sonar detection using autoregressive models in reverberation; experimental performances for wideband signals

We generalize here a suboptimal method to detect any kind of wideband signals embedded in the reverberation noise they have created. As this noise is strongly colored, first of all the optimal detector consists in whitening the observation. This step is made very difficult because of the nonstationarity of the reverberation noise. Under local stationarity assumption of the reverberation, we cut the received signal into blocks; On each block, we consider that the noise can be modeled as an AR process. Thus we perform the whitening step and finally an appropriate matched filter is applied. We also discuss the validity of local stationarity for the reverberation using the spectral measure of Itakura. We finally present performances of the approach valuated on two different experimental data sets.