Stereo Object Detection and Tracking Using Clustering and Bayesian Filtering Texas Tech University 2011 NSF Research Experiences for Undergraduates Site Project James Smith Faculty Advisor: Dr. Mohan Sridharan Abstract Robots equipped with sensors are being increasingly deployed in real-world scenarios Vision is a rich source of information for a mobile robot compared to other sensors Algorithms to process visual inputs computationally expensive Primarily focused on implementing image clustering to detect objects Secondary research into applying Bayesian filtering to object tracking Introduction Methods - Clustering The process of clustering, or grouping, has long been used in image analysis. The process allows simple object grouping, usually based on various similarities between pixels. We applied a generic clustering algorithm to add disparity as a third dimension. Search radius around each point to determine similar points Group similar points as potential objects Similar process to K-mean clustering Provides rough estimate of objects in 3-dimensional space *This research is supported by NSF Grant No. CNS Opinions, findings, conclusions, or recommendations expressed in this paper are those of the author(s) and do not necessarily reflect the views of NSF. Stereo Imaging and Clustering Left Stereo Image Right Stereo Image Disparity ImageClustered Image Erratic Robot Wheeled On-board Computer Battery Operated Stereo Cameras Back-Facing Camera Laser Range-Finder Clustering disparity images allows quick and accurate object detection. Research into Bayesian filtering shows promising outcomes in object tracking Conclusion Future Work Combine clustering with other techniques to improve the accuracy of object detection. Implementation of a Bayesian filtering system to track objects through time. Eventual integration with other sensor systems to produce more intelligent robots Bayesian Filtering Bayesian filtering works on the principal of creating a probabilistic prediction of future values of data, and correcting those predictions based on how closely the prediction matches reality. Estimates state through time Takes various sources of error into consideration Easily modifiable to trade off speed and accuracy Sources Greg Welch and Gary Bishop An Introduction to the Kalman Filter University of North Carolina, Sebastian Thrun, Wolfram Burgard, and Dieter Fox Probabilistic Robotics Cambridge, MA: MIT, 2005 Nikos Vlassis, Aristidis Likas, and Jakob Verbeek The Global K-Means Clustering Algorithm Pattern Recognition: Vol. 36 Issue 2, 2003 Stereo imaging has been used recently as an effective method of providing distance information in robotic applications. Previously with single image technology, many techniques were created to find and track objects. Our research consisted of applying these techniques to a stereo-vision system. Object Distance Right Camera Base bel(x t ) = ∫ P(x t | u t, x t-1 ) bel(x t-1 ) dx t-1 bel(x t ) = ŋ P(z t | x t ) bel(x t ) Left Camera x = State u = Control / Motion z = Observation Disparity-Distance Relation Disparity values are created on requested by the stereo- on-chip camera. An equation relating disparity to physical distance was determined experimentally. Real Distance = (591) +.02 Disparity Real Distance = m(Theoretical Distance) + B *m, B = Constants