Multi-Fish Tracking of River Herring in ARIS Acoustic Sonar Videos Using Physics-Informed Motion Priors
Notice: Undefined index: name in /tmp/php-events-calendar/views/353ebaae3171acb621b81ad24fa2c8e1 on line 11
Notice: Undefined index: short_desc in /tmp/php-events-calendar/views/353ebaae3171acb621b81ad24fa2c8e1 on line 48
Notice: Undefined index: main_desc in /tmp/php-events-calendar/views/353ebaae3171acb621b81ad24fa2c8e1 on line 50
Dion 303
Dr. Yuchou Change
ychang1@umassd.edu
https://umassd.zoom.us/j/2687322616?pwd=dUlWQldTTG5OaFJFVE5ZVi9zbmxhUT09
Zoom:
https://umassd.zoom.us/j/2687322616?pwd=dUlWQldTTG5OaFJFVE5ZVi9zbmxhUT09
Meeting ID: 268 732 2616
Passcode: Krt0re
Advisor: Dr. Yuchou Chang, Computer & Information Science Department and Data Science Program
Committee Members:
- Dr. Pingguo He, Department of Fisheries Oceanography, University of Massachusetts Dartmouth
- Dr. Ashokkumar Patel, Computer and Information Science Department, University of Massachusetts Dartmouth
Abstract:
River herring including Alewife (Alosa pseudoharengus) and Blueback Herring (Alosa aestivalis) migrate upstream through coastal river systems to reach freshwater spawning habitats, and imaging sonar provides a practical way to monitor these movements under poor visibility conditions. In ARIS acoustic videos, individual fish may change direction, turn sharply, overlap with nearby fish, or temporarily disappear from detection, so it is difficult to maintain a consistent identity and potentially cause the same fish to be counted more than once. DeepSORT as a well-established appearance-based multi-object tracking method was evaluated on a video as a preliminary benchmark. However, its reliance on appearance-based association is not well suited to acoustic fish tracking, since fish targets exhibit weak and inconsistent appearance information compared with conventional optical imagery. In this thesis work, we proposed a multi-fish tracking framework via four physics-informed motion priors. Moving fish candidates are first identified using a particle-based detection method that produces bounding boxes for small moving objects in the sonar images. These detections are then used as the common input to multiple tracking approaches, including centroid-based association and SORT, where Kalman filtering and global assignment are used to maintain fish trajectories across consecutive frames. To improve association during direction changes and other complex movements, we proposed a motion-prior-based tracking method that incorporates four motion cues: direction, velocity, turning behavior, and trajectory smoothness. These priors use recent track history to evaluate whether a new detection is physically consistent with the previous movement of a fish and are incorporated into the track-to-detection association cost before Hungarian matching. The proposed approach is evaluated using manually established fish-event ground truth to determine whether motion-aware association can improve track continuity and reduce duplicate or missed fish counts compared with conventional tracking methods. Experimental results demonstrate the effectiveness of the proposed method relative to a well-established tracking baseline.
All CIS and Data Science Graduate Students are encouraged to attend.
For further questions please contact Dr. Yuchou Chang at ychang1@umassd.edu