.png){kind=link}
.png){kind=link}
GitHub repository for the navigation for the Shark Detection code.
See Quantification of shark collective behaviour.docx for further details about the code.
Shark labeling process started by selecting an image to be scrutinized. In order to alleviate labeling all sharks sighted in a single frame, each image was cropped and adjusted to smaller areas of interest containing individuals. Further, an iterative process was initiated for which the cropped area was decomposed into sub-images that were magnified to ensure accurate labeling.
For each sub-image, the observer located the individual shark by clicking on the most distant points of each individual, i.e., the tip of the head (Hx,Hy) and tip of the tail (Tx,Ty). This provided the central position, as the mean of those two points (Cx,Cy), the relative body length (Bl) as the Euclidean distance between head and tail (in pixels) and the swimming orientation (u,v) of each individual. The magnitude of the (u,v) vector was fixed to the labeled shark length, although this parameter could be set as a constant since we were primarily interested on estimating sharks orientation . Once all individuals observed in a sub-image were labeled, the labeling process continued until the whole cropped area was scanned. It is worth noticing that while we divided the labeling process into sub-windows, the final position of all sharks were outputted with respect to the origin of coordinates in the complete image, with (0,0) position corresponding to the upper left corner.
The relative body length of each labeled sharks was calculated as the Euclidean distance between head and tail labeled points. In order to quantify the distance between individuals, the nearest neighbour to each labeled shark was determined using the central position (Cx,Cy) as a reference. For each shark?s central position, our algorithm again calculated the Euclidean distance towards all labeled sharks and only retained index of the individual identity that minimized this distance. Once all nearest neighbours were determined, we calculated the level of alignment between an individual and its closest neighbour using their swimming orientation (u,v) and the magnitude of the vectors for comparison. Based on these values, the swimming alignment was calculated using the Dot Product.
Values of each individual shark (the distance to its nearest companion, alignment level) were compared to the median estimates for the shoal sighted in a sampled image. We assumed that an individual was swimming in a coordinated manner with its nearest neighbour if the relative distance between them was smaller than 2 relative body lengths (in pixels). This criterion (threshold) was based on the median of the set of body lengths, Ls, of all individuals measured on a given sampled image.
Then, the median was calculated as the value of the sorted series, Ls. Distances between two closest individuals smaller than the established threshold are outputted as red colored asterisks, where an asterisk represented the central position (Cx,Cy) of each shark labeled on a sampled frame. Blue asterisks indicated between-individuals distances greater than the threshold. The median angle between all pairs of sharks present in a frame was used as a threshold to evaluate how well sharks were aligned with their closest neighbour.
Shark Detection toolbox was designed and developed by José Carlos Castillo (developer, maintainer).
The license of the packages is custom LASR-UC3M (Licencia Académica Social Robotics Lab - UC3M), an open, non-commercial license which enables you to download, modify and distribute the code as long as you distribute the sources.