Utilising AI predictions supplemented with Monte Carlo Markov Chain simulations, we assess the properties of the popular metric MaxN against the rarely used MeanCount to analyse underwater monitoring video data. We find that estimates of MaxN are biased towards the movement and schooling behaviour of a species, as well as the length of a video, whilst MeanCount is unaffected by these factors and has greater power to detect effects of interest. However, both metrics are biased by imperfect detection (e.g. low water visibility), and although correlated with abundance, they do not estimate absolute abundance. By extracting distance information from stereo RUVs, we apply distance sampling techniques to develop methods for estimating fish density that account for imperfect detection. Distance sampling, widely used in terrestrial wildlife studies, records distance from the observer for each detection, and hence estimates how the probability of detection varies with distance from the observer. This allows us to estimate absolute abundance, correcting for imperfect detection. Despite stereo RUVs containing distance information and being widely used by marine ecologists to obtain fish length and biomass, they have never been analysed in this fashion before. Our simulations show that abundance estimates derived from distance sampling are found to be unbiased under imperfect detection, thus it is possible to obtain accurate estimates of fish density in a broad range of practical settings, using equipment and software that marine ecologists are already frequently implementing, but in new ways.