Sept |
Helping Conservation of Snow Leopards with Image Processing and Machine Learning
Speaker: Agnieszka Miguel, Associate Professor and Chair, Electrical and Computer Engineering, Seattle University
When: 11:00AM ~ 11:50AM, September 10, 2018
Where: CSB130
Contact: Anthony Maciejewski (Anthony.Maciejewski@ColoState.EDU)
Abstract: Camera traps are one of the primary non-invasive population survey methods for studying snow leopards. Conservation biologists first sort camera trap images into sets with snow leopards and those without. Next, the snow leopard images are used to recognize individual cats based on the characteristics of their spot patterns, such as size, shape, orientation, and coloration. Currently, all of these work is performed manually, which takes up time that could be spent on more advanced data analysis. In our research, we automate the tasks of sorting camera trap images and recognizing individual cats.
To sort camera trap images, we first create motion templates using Robust Principal Component Analysis (RPCA), thresholding, and binary morphological operations. The number of spots found with a Cascade Object Detector that overlap with the motion template and their density are among the features input to the Support Vector Machine classifier. Our results are promising: we achieve an average classification accuracy of 93.74%. Our subsequent work uses Randomized-Subspace RPCA and apply evolutionary multi-objective optimization to find optimal values of several parameters used in the motion computation algorithm to improve execution speed and template accuracy.
To automate the recognition of individual snow leopards, we use an open-source software called HotSpotter, originally developed to identify zebras. The legacy HotSpotter involves many time-consuming manual tasks. We designe autonomous selection of multiple ROIs in motion templates, automate the query process, and introduce a method to build associations between individual ROIs based on clustering of similarity scores using Markov Clustering Algorithm. The proposed technique with its promising results of correctly recognizing individual snow leopards has the potential to save conservation biologists thousands of hours of manual labor.
Bio: Dr. Miguel received her Ph.D. in Electrical Engineering in 2001 from the University of Washington, and MSEE and BSEE from Florida Atlantic University in 1996 and 1994. Dr. Miguel's professional interests involve image processing, machine learning, and engineering education especially diversity and inclusion, retention, recruitment, and active learning. Her teaching interests include circuits, linear systems, MATLAB, digital image processing, and data compression. She is a member of the IEEE, ASEE, SWE, WEPAN, and Tau Beta Pi.
Currently, Dr. Miguel is the Chair of the American Society for Engineering Education (ASEE) Professional Interest Council I, a position that gives her a seat on the ASEE Board of Directors. Dr. Miguel has held several other officer positions across the ASEE including: Vice President of Professional Interest Councils, Division Chair and Program Chair of the ECE and New Engineering Educators Divisions, and ASEE Campus Representative. Dr. Miguel is also a member-at-large of the Electrical and Computer Engineering Department Heads Association (ECEDHA) Board of Directors. She has been a member of the ECEDHA Annual Conference Program Committee since 2012.