The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
740 
The motion analysis techniques were initially developed to 
identify candidates for counting and sizing fish in aquaculture 
(Harvey et al., 2004). Motion analysis is first used to identify 
sections of the image sequences that contain features of interest, 
effectively eliminating portions of the video that are devoid of 
features and not of direct interest to habitat mapping. This 
processing is effectively an image compression technique that 
dramatically reduces the amount of video sequences requiring 
inspection, and reduces digital video file sizes. The motion 
analysis is then used to estimate the percentage cover of 
selected regions within the video transects. The motion 
detector can be tuned to detect featureless versus feature-rich 
regions, or specific marine fauna or flora. 
The fundamental algorithm of the motion detector uses 
differences in intensity between consecutive frames. The most 
common approach recognises differences in colours based on 
thresholds and gains (Cheng et al., 2001; Ohta et al., 1980). A 
pixel is detected as a change if the difference between 
consecutive frames, multiplied by the gain, exceeds the 
threshold. Gains are used to amplify subtle differences and 
detect changes that would otherwise be below the threshold. 
Specific locations in the colour space of the images are used to 
identify the objects of interest. An operator will select these 
depending on the feature or species to be detected. The 
detected candidate regions are discriminated from noise using a 
region size range specified by the operator. 
Region growing is subsequently used to either complete the 
outlines of candidate features detected with motion analysis, or 
can be used to grow the outline of a feature manually selected 
by an operator (Adams and Bischof, 1994). The region 
growing algorithm can be configured to use colour, colour 
statistics and texture, which are the most readily identified 
visible signatures of benthic communities and sessile organisms. 
Figure 6. Example of a 3D measurement of surface area using a 
triangulation mesh. 
It is also possible to use stereo-image matching to determine 
volumes and surface areas of complex structures such as 
animals or physical features. This process is semi-automatic 
with the region of interest in one of the images defined initially 
by motion analysis processing. Operator selection of key points 
followed by epipolar searching and image matching (Gruen and 
Baltsavias, 1988) is then used to provide additional 3D 
locations within the boundary on the left and right images. The 
3D data points are used to define the surface based on a 
Delaunay triangulation, from which surface area and volume 
can be derived (see figure 6). An accumulation of such 
measurements can be used as an estimator of biomass of a 
particular features or species of interest within a transect. A 
critical factor in the effectiveness and robustness of the 
algorithms will be the improvement of image quality and 
resolution to be provided by the digital progressive scan 
cameras and direct-to-disk system. As can be seen from figures 
7 and 8, the image quality from the standard video system and 
the general reduction in image contrast caused by attenuation 
through the multiple refractive interfaces and water medium is a 
limiting factor. 
5. APPLICATIONS 
The vast majority of deep seabed is not mapped in detail, 
although acoustic multi-beam technology and photographic 
methods are increasingly providing data for key areas (Kloser et 
al. 2007). A primary contribution of video data to multi-scale 
surveys of the seabed is the definition of habitat at fine scales. 
Video transects can be used to target contrasts in acoustic maps 
to validate changes between habitats (see figure 7). Information 
on the biological associations with physical components of 
habitats enable mapped acoustic data, which has large coverage 
and is relatively inexpensive to collect, to be used as a proxy for 
the distribution of biodiversity (Kloser et al., 2007). Based on 
analysis of the video sequences, abundance measures such as 
density or cover can be made at a variety of scales of biological 
resolution, and can be related to habitat types at a variety of 
spatial scales (Williams et al., 2007). A key step in the use of 
image data in deep water habitat mapping is the move from 
qualitative to quantitative applications. 
Figure 7. Fine scale habitat identification by video within 
terrains defined by multi-beam acoustics. 
The non-extractive nature of video sampling gives it a 
significant advantage over conventional physical sampling with 
an epibenthic sled or trawl, particularly for monitoring. While 
biodiversity mapping relies on initial physical collection to 
provide an inventory of fauna, sensitive environments such as 
seamount coral communities (figure 8) benefit greatly from 
subsequent monitoring that is non-extractive, especially in 
conservation areas. Video surveys will never replicate the 
species-level resolution possible from collections of benthic 
fauna, but it is often possible to capture data for distinctive 
species. 
Where species have strong habitat associations and habitats 
have high spatial heterogeneity at scales of tens to hundreds of 
metres, video sampling will also provide more robust measures 
of abundance because the data are continuous and do not 
integrate across habitats. In contrast, samples from mobile 
collecting devices such as sleds or trawls do integrate across 
habitats, mixing the fauna and adding considerable uncertainty