| Anadromous Salmonid populations have been declining in North America for several years as the result of habitat degradation, over harvesting, and other factors with several species now listed as threatened or endangered under the United States Endangered Species Act. One major contributing factor in the decline has been the construction of hydroelectric power generating dams on the rivers used for spawning. To help alleviate the negative effects of the dams, various collection and bypass systems have been installed in the hopes of routing fish safely around rather than through the dams. To be effective, the systems must be both attractive to fish and safe for fish passage. Meeting these goals requires an understanding of fish behavior at the entrance to the bypass systems, but there are no currently established effective methods for studying fish behavior at a fine scale in these environments. This thesis explores the use of tracking algorithms processing multibeam sonar data as a method of observing fish behavior. An experiment using a sonar and stereo video camera system to simultaneously record free-swimming fish motion was conducted for this study. The video system provided a high-quality data set used as the “ground truth” to compare the output quality of various tracking algorithms processing the sonar data. Two algorithms are shown to be effective, a constant velocity Kalman smoother when fish motion is approximately a straight line, and the Segmenting Track Identifier (STI) when motion is more complex or in general when the measurement noise is small. The STI algorithm, proposed by Linder [1], is developed into a practical tracking algorithm. This development includes the re-parameterization of the motion model used by the STI to be more general and robust, and development of a method of generating a measurement prediction and associated covariance to allow the STI algorithm to be used in common data association frameworks. An STI based probabilistic data association filter is developed, and it is shown to perform well in simulations of tracking a single fish in clutter and in simultaneously tracking multiple fish from the video data. |
