Bottom-Up Saliency Detection in Crowds Shane Grant Kevin A. Heins Department of Computer Science Department of Mathematics University of California San Diego University of California San Diego La Jolla, CA 92093 La Jolla, CA 92093 [email protected] [email protected] Abstract This project aims to explore the topic of saliency in computer vision; specifically to create a bottom-up saliency based detector for events of interest in a crowded environment. 1 Statement of Qualifications Shane Grant is a third year undergraduate student studying computer engineering at the University of California, San Diego. He has taken introductory courses to artificial intelligence (CSE 151) and has worked on object detection for the UCSD Unmanned Aerial Systems team for two years, utilizing the Viola-Jones algorithm for robust real-time object detection.1 Kevin Heins is a third year undergraduate student studying probability and statistics at the University of California, San Diego. He has taken both undergraduate and graduate level courses in statistical and probabilistic learning (CSE 151, 250A). Kevin is also a member of the UCSD UAS team working on object detection. 2 Project Outline The goal of our project is to create a bottom-up saliency detector and apply it specifically to crowds, though the detector should be general enough for use in other applications. We intend to draw on research done by Timor Kadir2, Dashan Gao3, Nuno Vasconcelos4, and Christopher Kanan5. 2.1 Learn Relevant Algorithms and Software Most essential to the project is developing an understanding of the concept of discriminant saliency. Understand the creation of saliency maps for each discriminant used and the creation of a "super-map" composed of the aforementioned saliency maps. Deadline: Week 2 2.2 Assemble a Data Set Training and Testing Since this project will focus on saliency detection and its applicability to crowds, data will be drawn from databases of crowd in both static image and video format. The pedestrian crowds database from UCSD SVCL6 and data utilized in unsupervised detection of motion in crowds7 will be combined with video and still image data obtained at the UCSD campus. Partition the data set into distinct test and training sets.Deadline: Week 4 2.3 Train the Algorithm Utilizing existing algorithms for bottom-up based saliency detection, train based upon the data collected previously. Deadline: Week 6 2.4 Test and Modify Run algorithm on test data set and observe the results. Modify the algorithm as necessary to improve detection or better fit expectations. Analyze results and compare against more traditional methods of detection. Deadline: Week 10 3 Division of Labor Since the topic of visual saliency is new to both Shane and Kevin, collaboration will be common for the majority of tasks involved with the project. Since Shane has more experience programming and Kevin more in mathematics, topics that dip more heavily into one of these areas may be appropriately divided. 4 Experimental Questions Does bottom-up saliency improve upon search methods of detection (accuracy, computation speed, etc)? Does bottom-up saliency effectively find objects of interest (especially in crowded environments)? Can this be applied to areas other than crowds? Is this a biologically viable method of detection? References [1] Viola, Paul & Jones, Michael (2001) Robust Real-Time Object Detection, Second International Workshop on Statistical and Computational Theories of Vision - Modeling, Learning, Computing, and Sampling http://research.microsoft.com/~viola/Pubs/Detect/violaJones_IJCV.pdf [2] Kadir, Timor & Zisserman, Andrew & Brady, Michael (2004) An affine invariant salient region detector, European Conference on Computer Vision 2004, Pp 228-241 http://www.robots.ox.ac.uk/~timork/Saliency/eccv_asalscale_final.pdf [3] Gao, Dashan & Mahadevan, V & Vasconcelos, Nunan, (2007) The discriminant center-surround hypothesis for bottom-up saliency, Neural Information Processing Systems (NIPS), Vancouver, Canada http://www.svcl.ucsd.edu/publications/conference/2007/nips2007/nips2007_budiscsal.pdf [4] Gao, Dashan & Vasconcelos, Nunan, (2007) Bottom-up saliency is a discriminant process, Proceedings of IEEE International Conference on Computer Vision (ICCV), Rio de Janeiro, Brazil http://www.svcl.ucsd.edu/publications/conference/2007/iccv07/iccv07_gao_budiscsal_psycho.pdf [5] Kanan, Christopher, (2007) NIMBLER: A Model of Visual Attention and Object Recognition With a Biologically Plausible Retina, CSE 252C Fall 2007 http://www-cse.ucsd.edu/classes/fa07/cse252c/projects/ckanan.pdf [6] Chan, Antoni & Vasconcelos, Nuno Understanding Video of Crowded Environments, Statistical Visual Computing Lab at UCSD http://www.svcl.ucsd.edu/projects/crowds/ [7] Brosow, Gabriel J. & Cipolla, Roberto Unsupervised Bayesian Detection of Independent Motion in Crowds, IEEE CVPR, June 2006, New York, NY
View Full Document