The ability for submersible vehicles to accurately maintain position and orientation is a necessity. The process of maintaining a vehicle's predefined pose in the presence of disturbances (undersea currents and reaction forces from manipulators attached to vehicle) is known as station keeping. Station keeping can be used for many different underwater applications such as repair of underwater structures and near seabed data collection. Station keeping using a vision system has the advantage of being able to use natural rather than manmade beacons for motion detection while inherently having a high resolution and update rate. The camera is setup in a similar fashion to that of image mosaicking and the methods for motion estimation overlap greatly between the two applications (see Fig. 2). The general method for visual station keeping is to maintain a reference image acquired from the station and compare live incoming frames with this image to estimate and correct for vehicle drift.
Stanford/MBARI researchers proposed a method of measuring vehicle drift using a texture tracking strategy (Marks et al. 1994). The method of motion estimation is the same process as described in video mosaicking (Marks, et al., 1995). Firstly the spatial intensity gradient of the images is filtered to highlight zero crossings using a Laplacian of Gaussian filter. The incoming images are then correlated with the reference image in order to measure movement of features. Filtering in this case is an attempt to highlight image textures and reduce the effect of noise and non-uniform illumination. Tests were performed in a test tank while the vehicle was on the surface but no external measurements were taken in order to thoroughly evaluate the performance of the system. The result consisted of the plots of commanded control effort to counteract the disturbances in order to hold station. Such a method depends on having a highly textured image in order to find regions of correlation. Correlation-based methods' inability to deal with changes in the image due to rotations will inhibit accurate motion estimation.
Negahdaripour et al. proposed a method of station keeping by directly measuring motion from spatio-temporal image gradient information (optical flow) (Negahdaripour et al. 1998; Negahdaripour et al. 1999). This method allows for the estimation of 3D motion directly using the spatio-temporal derivatives of incremental images captured over a short period of time (Negahdaripour & Horn 1987). A generalized dynamic image motion model was later developed (Negahdaripour 1998) to account for variations in the scene radiance due to lighting and medium conditions underwater. This is of particular importance when using flow-based methods in underwater imagery due to the artificial light source motion. A technique for calculating both instantaneous velocity and absolute position is implemented to increase the limit of inter-frame motion. The position calculated by integrating the velocity over time is used for course correction before the absolute position is used for finer adjustment. This method is susceptible to sporadic miscalculations in velocity, which, accumulated over time, can result in inaccurate position estimations.
Cufi et al. (Cufi et al. 2002) make use of a technique previously developed for a mosaicking application (Garcia et al. 2001b). The acquired images are convolved with high pass filters in both the x and y direction in order to find small windows with the highest spatial gradient (interest points). These windows are then compared to the reference image using two methods. Firstly a correlation based strategy is used to find candidate matches for each interest point. Then a texture characterisation method is performed on each point to select the best correspondence using different configurations of the energy filters (Garcia et al. 2001a). As stated above the correlation method is incapable of dealing with large rotations in images due to yaw motion of the vehicle. This problem is overcome in this case by simultaneously creating an image mosaic. The mosaic creation method is based on previous work completed by Garcia et al. and is discussed further in section 6 (Garcia et al. 2001b). The implementation of the image mosaic also allows for greater inter-frame motion. No overlap between image iterations is needed as the mosaic can be referenced for motion estimation. This method improves on previous correlation based approaches but could again suffer from a lack of distinct textures in the subsea environment while the execution of the mosaicking system may be too computationally expensive to be performed in a real-time on board computer.
Other methods implement a combination of methods to achieve station keeping. Van Der Zwaan et al. use a technique of integrating both optic flow information with template matching in order to estimate motion (van der Zwaan et al. 2002). The station keeping system tracks an automatically selected naturally textured landmark in the image plane whose temporal deformations are then used to recover image motion. A prediction of the location of the landmark is made by utilizing optical flow information. This estimate is then refined by matching the image with the selected reference frame. This system performed in real-time and showed robust results even in the case of limited image textures however, experiments were performed on poor resolution images thus decreasing accuracy and improving algorithm speed.
Station keeping, much like mosaicking, has many methods for tracking motion from vision: correlation based, feature based, optical flow based etc and selection of the most appropriate method is by no means a trivial task. Many factors have to be considered to obtain accurate results with the final goal of creating an autonomous real-time station keeping system. The methods discussed are hard to compare due to differing test setups and vehicle dynamics, however, none of the methods mentioned appears fully capable of overcoming the difficulties of station keeping faced in underwater environments, at least in a real-time on board system in an unstructured environment. While improved hardware will allow for the analysis of higher resolution images and thus superior accuracy, there still remains room for algorithm advances and sensor fusion research in order to reproduce the results gained in controlled pool trials and simulations in actual real ocean environments.
Was this article helpful?