Robust Tracking and Classification of Multiple Objects in Sector-Scan Sonar Image Sequences
   The fast update rate and good performance of new generation electronic sector scanning sonars is not allowing practicable use of temporal information for signal processing tasks such as object classification and motion estimation. Problems remain however, as objects change appearance, merge, manoeuvre, move in and out of the field of view, and split due to poor segmentation. This paper presents an approach to the segmentation, 2D motion estimation and subsequent tracking of multiple objects in sequences of sector scan sonar images. Applications such as ROV obstacle avoidance, visual servoing and underwater surveillance are relevant. Initially, static and moving objects are distinguished in the sonar image sequence using frequency domain filtering. Optical flow calculations are then performed on moving objects with significant size, to obtain magnitude and direction motion estimates. Matches on these motion estimates, and the future positions they predict, and then used as a basis for identifying corresponding objects in adjacent scans. To enhance robustness, a tracking tree is constructed storing multiple possible correspondences and cumulative confidence values obtained from successive compatibility measures. Deferred decision making is then employed to enable best estimates of object tracks to be updated as subsequent scans produce new information. The method is shown to work well, with good tracking performance when objects merge, split and change shape. The optical flow is demonstrated to give position prediction errors of between 10 and 50cm (1 to 5% of scan range), with no violation of smoothness assumptions using sample rates between 4 and 1 frame per second.
   This paper will also present an investigation of the robustness of an inter-frame feature measure classifier for underwater sector scan sonar image sequences. In the initial stages the images are of either divers or ROVs. The inter-frame feature measures are derived from sequences of sonar scans to characterize the behaviour of the objects over time. The classifier has been shown to produce error rates of 0 to 2% using real non-noisy images. The investigation looked at the robustness of the classifier with increased noise conditions and changes in the filtering of the images. It also identified a set of features that are less susceptible to increased noise conditions and changes in the image filters. These features are the mean, variance and the variance of the rate of change in time of the intra-frame feature measures area, perimeter, compactness, maximum dimension and the first and second invariant moments of the objects. It is shown how the performance of the classifier can be improved. Success rates of up to 100% were obtained for a classifier trained under normal noise conditions, SNR around 9.5 dB, and a noisy test sequence of SNR 7.6 dB.

Navigational Data Fusion in the Ocean Explorer Autonomous Underwater Vehicles
   One of the primary difficulties associated with command and control of an autonomous underwater vehicle is timely incorporation of accurate navigational information into the positional control algorithms. This task is complicated by the fact that both geodetic, and relative velocity positional information from a suite of navigation sensors may be available simultaneously, or from any combination of individual sensors at arbitrary time intervals. Such a suite of sensors may include GPS, acoustic baseline tracking, and acoustic beacon tracking as geodetic referenced sensors; along with compass/water speed sensors, accelerometers, and acoustic doppler velocity instruments as relative position data sources. It is then appropriate to implement a navigational data-fusion which can synergistically combine the available positional information in order to achieve the most accurate and robust input to the navigational control structure. This paper details the navigational data-fusion algorithms employed in the Ocean Explorer series AUVs. The algorithm meets the above requirements for continuous updating of positional information from arbitrary subsets of the onboard navigation sensors, which may or may not be available at any point in time. The data-fusion process is explicitly detailed along with a strategy for emergency navigation in the case of primary sensor failure. Data from AUV missions is presented which illustrates the algorithms utility.

Estimating the Sea Current and Finding the Optimal Heading of an AUV Through Neural Network Motion Control
   For an Autonomous Underwater Vehicle (AUV), sea current is an unknown varying parameter which can not be given before operation and which may impose significant influence on the performance of an AUV. Further more, it is difficult to measure the sea current from a moving or suspending AUV. Conventional motion controller, such as PID or optimal controller based on simplified hydrodynamics model of an AUV can hardly cope with the complicated environment of unknown varying sea current, which may induce unfavorable oscillation. That is likely to degrade the performance of sensors such as visual image. In order for an AUV to improve the ability of autonomy, adaptability and performance, its control system should possess self-adaptability and learning ability. It is especially important for an AUV to have the capability of automatically estimating the sea current. This paper presents a technique for automatically estimating the sea current and finding the optimal heading through neural network motion control. The AUV model described here is mainly based on the test-bed "Intelligent Underwater Vehicle II" for evaluating new concepts and ideas from artificial intelligence and intelligent control, which is developed at Harbin Engineering University, Harbin, PR China. The main body of the vehicle has the shape of a cylinder with the length much greater than the diameter. The vehicle also has a spherical head and a cone tail. Therefore, the lateral resistance is much greater than the longitudinal one. The optimal heading where the lateral resistance is minimal for the vehicle to travel long distance or to keep position may vary with the motion states and the sea current condition. By using a model reference neural network controller, the velocity vector of the vehicle can be oriented to point to the goal position. Meanwhile the heading of the vehicle turns to the angle where the lateral resistance approaches to zero. When the vehicle reaches the goal point, the heading is approximately opposite to the direction of the current. Thus, the AUV can estimate the direction of the current with little error. Considered in a sense of hydrodynamics, the vehicle moving with constant velocity in calm water is equivalent to the vehicle keeping stationary in sea current. Thereby, the AUV can estimate the speed of the current too. The model reference neural network motion controller has been proved working satisfactorily through large amount of experiments in ship model tank and in natural lake environment with surface wave, but with no current. Constrained by experimental conditions, the sea current estimation and optimal heading searching experiments were carried out in computer simulation environment, with good results.

Obstacle Avoidance Using Fuzzy Neural Networks
   If an underwater vehicle is to be completely autonomous, it must have the ability to avoid obstacles to safely operate. Many solutions to the problem of obstacle avoidance of an autonomous underwater vehicle (AUV)
   around obstacles have been posed in resent years, but all have limitations of one form or another. This paper describes a three-dimensional obstacle avoidance controller for AUV using a fuzzy neural network. The fuzzy neural network controller architecture is one of model reference including a fuzzy logic controller and a neural network controller. The fuzzy logic controller receives the information of the distance and change of its between the AUV and the obstacles, then outputs the thrust commands to control the motion of obstacle avoidance. The neural network controller receives the information of disturbance (such as wind, wave, current,
   etc.) of ocean environment, then its control commands modify the outputs of the fuzzy logic controller to adapt to the varied oceanic environment. Since the disturbance of ocean environment cannot be measured, the state errors of thrusters adding those of motion of the AUV are inputs to the neural network. And the errors between the expected values and the real ones are used to change the weights of neural network on-line. Results of simulation studies using a five degrees of freedom nonlinear ship maneuvering mathematical model show that the proposed method can be efficiently applied to obstacle avoidance of the AUV in complex and unknown oceanic environment.

AUV Obstacle Avoidance and Navigation Using Image Sequences of a Scanning Sonar
    In this work, we use the continuous image sequences generated by an electronic scanning sonar to achieve the aim of obstacle avoidance and visual navigation for AUV (Autonomous Underwater Vehicles). Using sonar systems for sensing of unknown underwater environments is the bet selection in practice. However, the critical demand for real-time signal processing and the uncertainties of AUV's dynamics make online detection of obstacle a challenging task. Track-Before-Detection (TBD) algorithm uses information contained in image sequences to estimate the dynamics of AUV, then apply Dynamic Programming (DP) algorithm to solve for the problem of detection. This method reduces the computational cost to meet the real-time demand of AUV system. An application for obstacle avoidance and navigation system based on this method is proposed.

Asynchronous Multi Sensor Fusion for Discrete AUV Navigation
    This work develops a nonlinear filter of discrete system that will fuse measurements from Doppler sonar giving autonomous underwater vehicle speed over water and speed over ground, inertial measurements from gyros and a compass-each with biases-and differential GPS true position (with error) in surfaced transit. While underwater, LBL or SBL acoustic data would replace the DGPS. Navigational position and rate is obtained in spite of the arrival times of the different sensory inputs being non uniform and arbitrary.
    One of the problems with data of this type is that inputs are received at non uniform intervals and the filter has to deal with local body referenced data as well as globally referenced data at the same time. In conventional discrete linear filtering with uniformity of data reception the filter gains may be pre comupation of optimal gains in terms of assumed known statistics of the input data and the discrete system model.
    This problem is formulated as a discrete time extended Kalman filter. The base update rate is chosen such that data arrival times are arbitrary integer multiples of the base. Optimal filter update gains are obtained at data arrival times such that estimation errors remain bounded although increasing between updates. Convergence properties are proven are results from at sea experiments with the Florida Atlantic University Ocean Voyager Autonomous Underwater Vehicle are given. Real time execution characteristics as well as some computational issues are discussed. The similarity to control with quantized output data is discussed.

Robust Filtering and Its Application to SINS Alignment
    The H infinite optimization method is applied to the initial alignment of strap-down inertial navigation system(SINS). The robust filter is used to estimate the angle misalignment for a moving base during the initial alignment process. It improves the standard Kalman filter performance when the system noises are not known completely. The H infinite filter is more robust and also improves the accuracy. A simulation example is given which demonstrates the availability of the robust filter.
    The strap-down inertial navigation system based on numerical computer has many good features such as high reliability, strong functions, and low cost etc.. It is a very promising navigation system. The conflict between the robustness and swiftness and accuracy in the alignment process is a problem that is to be resolved in the SINS. The alignment method of transmission from master SINS ( SINS on the naval vessels or submarines ) to slave SINS ( SINS on the missiles ) is swift. In this method the speed ( acceleration and position ) error between the master SINS and slave SINS is fed to a Kalman filter to estimate the angle misalignment. The angle misalignment is compensated in the SINS and the accuracy is raised. But the standard Kalman filter is sensitive to the uncertainty of the noise property in the system. It is very difficult to get high accuracy with a standard Kalman filter and sometimes it could even be divergent. On the other hand, the existence of uncertainty of the noise property is very common in the initial alignment process of a moving base. It is very important to improve the robustness of the filter.
    To cope with the uncertainty which exists commonly in practice. A number of methods have been developed such as virtual noise compensation method, attenuated memory method, and min-max design method etc.. But in some extent, all these methods sacrifice optimum to robustness. In recent years, method based on H infinite norm is introduced to design the robust filter. But the algorithm is often too complicated to be used. U. Shaked has presented a relatively simple structure of linear observer and proved that L2 estimator is a special case of H infinite filter when gama tends to become infinite. What is needed in this design method is to find the solution of a revised Riccati equation. Based on the work of U. Shaked, Bernstein and others, a robust filter is designed in this paper which can be used in the transmission alignment process of a SINS.

A New High Accuracy Super Short Base Line (SSBL) System
   This paper describes an acoustic Super Short Baseline (SSBL) tracking system capable of tracking up to twenty targets simultaneously, accurately and at long ranges. The design of the hydrophone array is based on an
   expansion of a conventional SSBL array configuration, an equilateral triangle of multiple elements with greatly increased inter-element spacing to provide high tracking accuracy. With this design the bearing and elevation of targets at ranges greater than 10 km can be tracked with an accuracy better than 0.1 degree.
   The application of SSBL to target tracking is not new and there are several SSBL products on the market. Several recent applications require tracking accuracy that exceeds those of existing SSBL systems. In these applications a significant improvement in the bearing and range accuracy of the SSBL system can substantially reduce operational cost. These applications include full ocean depth SSBL systems for ROV navigation and tracking, towbody tracking, integrated LBL and SSBL navigation, and long seismic streamer tracking. The best currently available SSBL systems advertise a bearing accuracy of 0.25 % of the slant range.
    This paper discusses the patented design and performance of the integrated SSBL/LBL system that provides 0.1% slant range accuracy and range that are not achievable using currently available SSBL tracking

Automatic Installation of Underwater Elastic Structures under Unknown Currents
    Target water depth of ocean development is getting deeper, and several automatic installation techniques have been proposed for installation of underwater equipment. Considering the trend, the authors propose a new automatic technique, learning Tracking Controller (LTC), to install structures whose flexibility can not be ignored any more in unknown current. This method also control elastic response of the equipment to be installed and applicable much larger structures. The proposed LTC can install underwater structures not only in excellent accuracy in any water depth but also control the elastic responses of structures simultaneously.
    In the underwater installation operations, the steady nonlinear hydrodynamics due to ocean current, tidal current and hydrodynamic interaction between structures are regarded as main disturbances. However, the information about these disturbances, which is generally important and necessary for the conventional active controller, is difficult and costly to be collected before operation.
    The present LTC, which consists of both feedback controller and feedforward controller, can be operated without any information about current. By learning unknown disturbances translated from the errors between the realized trajectory and objective trajectory, LTC improves its feedforward control force and makes the structure track the objective trajectory accurately in the final stage. The convergence condition and robustness of LTC is shown.
    Limited number of elastic modes, such as rigid body modes and significant lower order elastic modes are controlled, because the number of sensors and actuators which can be used for the structure are limited. But the feedback controller is formulated so as to augment the damping of the elastic responses for the residual higher order modes.
    In order to conform the capability and effectiveness of LTC, basin tests are carried out under unknown current. Two types of experimental models, a plane flexible structure and a cubic rigid structure, are designed. In the experiments, both of two models were successfully made track the given trajectory and docked to their targets, by their own LTC respectively. To obtain docking accuracy of +-5 mm the models tracked the trajectory six to seven times for learning. The maximum speed of unknown steady current is 0.07m/s. No elastic responses of the flexible model were stimulated.

Multibeam Surveys and Submersible Observations of a Catastrophic Summit Collapse on Loihi Submarine Volcano, Hawaii
   Extensive destruction of the summit of Loihi submarine volcano accompanied by earthquakes located beneath the summit took place beneath Loihi during August 1996. Bathymetric resurveys of the Loihi edifice using SeaBeam, Simrad and deep-towed Reson-Focus multibeam systems showed the presence of new north-south striking fissures, normal faults, a new pit crater with high-temperature hydrothermal venting of 198íC and polymetallic sulfide formations located at the southern edge of the 1,000-meter deep summit. Comparisons between the pre-1996 SeaBeam data and the new bathymetry show a net loss of about 100-million cubic meters of rock volume into the new 1,300-meter-wide, 300-meter-deep pit crater. Submersible observations with Pisces V show the northern edge of Pele's Pit to be bound by a normal fault located north of the site of the former Pele's Vents cone that has disappeared into the new pit crater. The southern edge of the pit is represented by irregular north-south striking ridges and pinnacles extending several hundred meters above the floor of the pit. Submersible observations of these new vents during September 1996 showed vigorous venting at the site with extensive white-colored bacterial mats and polymetallic sulfides hanging off the rock face above the vents. Follow-up mapping and sampling of the vent sites during August 1997 showed vent temperatures of up to 198íC, clear venting from orifices up to 10 centimeters in diameter, dense populations of vent shrimps around the vents, and a centimeters-thick veneer of fine-grained polymetallic sulfide mineralization (including pyrite, sphalerite, bornite, and anhydrite) along the rock walls above and around individual vents. The new crater appears to have been formed by the sudden withdrawal of near-summit magma back into the magmatic plumbing system of Loihi. Long-term monitoring of the summit of Loihi with robotic observatories over the past five years showed the first indications of a possible summit magma withdrawal during December 1991 when an intense earthquake swarm was accompanied by a 40-centimeter deflation of the summit and temperature instabilities in the summit hydrothermal vents.

Monitoring Seafloor Crustal Movements: Progress Report of Relative Positioning on the Seafloor
   1. Introduction
    We are developing systems for monitoring seafloor crustal movements by measuring horizontal and vertical relative positions on the seafloor. Horizontal positioning utilizes precise acoustic ranging. Positioning relative to a land reference site is also pursued by using sea surface Differential GPS as well as the acoustic ranging. We are examing ocean bottom pressure gauges and an ocean bottom gravimeter to monitor vertical crustal movement on the seafloor. Preliminary results will be presented.
   2. Horizontal positioning
    Precise acoustic seafloor ranging with a resolution of 1 cm is a straightforward method to detect a deformation across a fault zone. Although the maximum range of a baseline is limited to be less than about 10 km, this method has an advantage that it can be applied to any site in the open ocean irrespective of the distance from land. We carried out a horizontal seafloor ranging experiment at a baseline of 1350 m. The system uses a m-sequence signal for linear pulse compression.The results show that each measurement has an rms error of 0.9 cm. We have also deployed three units of acoustic ranging system across a spreading axis of the super-fast spreading Southern East Pacific Rise, where the average full spreading rate is about 15 cm/y. We plan to recover the instruments in September 1998 after measurements of about 1 year.
    The most critical problem is the effect of temperature variation. The effect is estimated to be corrected for measurements on the deep seafloor, where temperature variation is relatively small.
    We have carried out the first experiment in Sagami Bay aiming at cm-order seafloor positioning by DGPS/acoustic measurements jointly with a group in the Hydrographic Department of Japan. However, only a little part of data were logged on real-time kinematic GPS positioning, and we are now processing the data of acoustic ranging to estimate the resolution of the acoustic positioning at sea surface. We estimate that major problems remain in precise kinematic GPS positioning at sea surface at a baseline longer than 100 km.
   3. Detection of vertical movement
    Ocean bottom pressure measurements can contribute to monitoring vertical movements of the seafloor in two ways. Pressure and gravity measurements can distinguish between a movement of the seafloor and one in the sea water. An array of ocean bottom pressure gauges is interpreted as a monitoring system of vertical movements if the effects of water currents are weak. Results of pressure measurements at two points about 100 m apart from each other on the seafloor of Sagami Bay show a resolution of differential pressure measurement comparable to seafloor vertical movement by 1 cm.
    We developed an ocean bottom gravimeter (OBG) for use by a submersible and for stand alone measurements. Free gimbals suspensions with oil damper roughly keep a sensor package of Scintrex CG-3M/SB vertical and the effect of the remaining tilt is numerically corrected by the gravimeter. The gravimeter is in a pressure-tight spherical housing made of titanium alloy rated to 7200 m water depth. Several tests we carried out on land show repeatability of gravity measurements comparable to a LaCoste & Romberg land gravimeter. Results of sea trials show that the OBG has a resolution better than several microgals on the calm seafloor.