This work follows previous works on tracking of targets on Forward Looking Sonar images. A Kalman filter based on a process model of the vehicle was implemented considering two strong assumptions: firstly, the obstacle is fixed in relation to the world reference frame and secondly, it lies proud on the seabed. Consequently, Kalman filtering leads to a biased estimation of successive positions of an obstacle floating in the water column. Starting with this observation a new algorithm has been developed to allow a deferred estimation of the z-coordinate (along the absolute z-axis) of the obstacle related to the vehicle. This is performed offline by minimizing at each step of the sequence the root mean squared deviation (RMSD) between measured sonar positions and predicted positions, i.e. by minimizing the innovation values of the Kalman filtering. Results are given on real data recorded in March 2009 and April 2010 during sea trials organized by GESMA involving the Rapid Environment Assessment (REA) "Daurade" AUV with its BLUEVIEW P450 obstacle avoidance sonar. Tethered mines and other floating obstacles like plastic chains were laid.