Table 1.
Behavior condition ethogram of dolphins under professional care.
Fig 1.
Diagram of the experimental setup.
TOP: Illustration of the main habitat, with camera placements (blue enclosures) and fields of view (gray cones). BOTTOM: Top-down individual camera views, with objects in the habitat marked. Yellow { Dolphin bounding boxes, Green—Drains, Red—Gates between regions, Orange—Underwater windows (3 total). Correlated dolphin bounding boxes are indicated by number.
Table 2.
Block time intervals.
Fig 2.
Combined figure demonstrating camera overlap, bounding box meshing, and animal position uncertainty.
TOP: Top-down individual camera views, with dolphin bounding boxes in yellow (correlating boxes are numbered). The habitat-bisecting lines (ls) for each camera frame are indicated in solid red. Distances from Bounding Box 2 (centered on the black and gray crosshair) to the closest frame boundary (db) and the boundary to the bisecting line (dl) are indicated by the white measurement bars. MIDDLE: Meshed camera views including dolphin bounding boxes (yellow), with the location uncertainty distribution (A) overlaid for Box 2. BOTTOM: 2D location uncertainty distribution (A) with major (a-a, black) and minor (b-b, red) axes labeled and separately plotted.
Fig 3.
Illustration of tracklet generation.
TOP: Tracklet segments (red) overlaid on a single video frame, generated by stitching the views from both cameras. Each tracklet in this frame was plotted from its inception to each corresponding dolphin’s current position. While each dolphin can be tracked, the lack of clarity when underwater impedes individual identification. CENTER: x-y view of example tracklets (red and green on gray lines) of two dolphins (highlighted light orange), which are also shown in Fig 1, top. POPOUT-RIGHT: Vector illustrations of the two example tracks. Example notation for tracklet j (red): position (p(j, t′)), velocity (v(j, t′)), yaw (θ(j, t′)), and yaw rate . POPOUT-BOTTOM Illustration of tracklet generation, with detections (stars) and tracklet proximity regions (dashed). Example notation for tracklet j (red): position (p(j, t)), velocity (v(j, t)), Kalman-predicted future position
, true future position (p(j, t+ 1)), and future animal detection (u(j, t+ 1, i′)).
Table 3.
Performance comparison between manual and CNN animal detections for two sessions as part of a separate monitoring exercise, where individual dolphins were tracked as opposed to the entire group.
A1 and A2 refer to specific dolphins, with A1 being tracked over two recordings during Deployment 1, and A1 and A2 being tracked during the same recording during Deployment 2. “Detectability” is defined as the total time each individual dolphin was able to be detected by either the human or CNN trackers over each deployment period.
Fig 4.
Static position distributions for OTS and ITS.
A note on the format of the training sessions: Dolphins spent more time stationed at the main island during public presentations than non-public animal care sessions. During public presentations, ACSs spend a higher portion of the training session on the main island because it is within view of all of the public attending the presentation. Non-public animal care sessions are more fluid in their structure than public sessions. ACSs often use the entire perimeter of the habitat throughout the session.
Fig 5.
Spatial distributions for dynamic OTS, with position distributions along the first column and speed distributions/quiver plots along the second column.
Prior to the first full training session of the day at 09:30, the dolphins were engaged in low intensity (resting) swimming clockwise around the perimeter of the habitat, with the highest average OTS speeds recorded after the 9:30 sessions. From there, speeds trail off for the subsequent two time periods. The 13:30–14:30 time block is characterized by slower swimming in a predominantly counterclockwise pattern. There is an increase in speed and varied heading pattern during the 15:00–16:00 time block.
Fig 6.
Spatial distributions for dynamic ITS, with position distributions along the first column and speed distributions/quiver plots along the second column.
During the animal care sessions (Block 1: 09:30 to 10:00, Block 3: 13:00–13:30), the dolphins engaged in lower intensity swimming throughout the habitat than the presentation sessions (Block 2: 11:30–12:00, Block 4: 14:30–15:00). This difference is qualitatively explained through the discrepancy in ACS requests from the animals: high-intensity behaviors are prompted more often during presentations, while care sessions cover a wider variety of behaviors. Conversely, spatial coverage of the habitat does not have high variance within the ITS blocks, with an expectedly high concentration on the central island where the ACSs are located for all ITS blocks.
Fig 7.
Speed and yaw probability distributions and joint differential entropies, respective to time block.
TOP: Probability density functions of animal speed (m s−1) for OTS (left) and ITS (right). MIDDLE: Probability density functions of yaw (rad) for OTS (left) and ITS (right). BOTTOM: Joint differential entropy of speed and yaw for each block of OTS (left) and ITS (right), with limited-range y-axes to more clearly show value differences.
Table 4.
Speed and yaw joint differential entropy.