Automatic detection of fish and tracking of movement for ecology

Abstract

1. Animal movement studies are conducted to monitor ecosystem health, understand ecological dynamics and address management and conservation questions. In marine environments, traditional sampling and monitoring methods to measure animal movement are invasive, labour intensive, costly, and measuring movement of many individuals is challenging. Automated detection and tracking of small-scale movements of many animals through cameras are possible. However, automated techniques are largely untested in field conditions, and this is hampering applications to ecological questions. 2. Here, we aimed to test the ability of computer vision algorithms to track small-scale movement of many individuals in videos. We apply the method to track fish movement in the field and characterize movement behaviour. First, we automated the detection of a common fisheries species (yellowfin bream, Acanthopagrus australis) from underwater videos of individuals swimming along a known movement corridor. We then tracked fish movement with three types of tracking algorithms (MOSSE, Seq-NMS and SiamMask), and evaluated their accuracy at characterizing movement. 3. We successfully detected yellowfin bream in a multi-species assemblage (F1 score = 91%). At least 120 of the 169 individual bream present in videos were correctly identified and tracked. The accuracies among the three tracking architectures varied, with MOSSE and SiamMask achieving an accuracy of 78%, and Seq-NMS 84%. 4. By employing these emerging computer vision technologies, we demonstrated a non-invasive and reliable approach to studying fish behaviour by tracking their movement under field conditions. These cost-effective technologies potentially will allow future studies to scale-up analysis of movement across many underwater visual monitoring systems.

Nonintrusive and automatic quantitative analysis methods for fish behaviour in aquaculture

“…According to various requirements for the analysis of specific fish behaviours, different network structures have been designed, but they are all based on the original neural network (Lopez-Marcano et al, 2021). CNN architectures are popularly applied to analyse fish behaviours (Han et al, 2020) and can be used to track the behaviour of fish (Li et al, 2019;Wang et al, 2017).…”

Optimising sampling of fish assemblages on intertidal reefs using remote underwater video

“…In addition, extracting data from the full videos can develop a ‘bottleneck’ in the analysis workflow due to observer fatigue, with 1 h of video taking up to 13 h to process ( Cappo, Harvey & Shortis, 2006 ; Campbell et al, 2015 ). Recently, researchers have been recognising the value of time-saving applications of deep learning to automatically count and identify fish in videos ( Christin, Hervet & Lecomte, 2019 ; Ditria et al, 2021 ; Lopez-Marcano et al, 2021 ). However, models do not perform as well in videos with complex backgrounds, poor visibility, differing light conditions, or cryptic camouflaged fish ( Salman et al, 2019 ).…”