Joint detection and tracking of multi-objects from image observation under unknown background dynamics

© 2013 Dr. Haseeb MalikThis dissertation applies two independent information fusion frameworks to jointly detect and track multi-objects in an image sequence using random sets (RFS). The first framework fuses information by treating underlying observation and associated models as unambiguous processes. The second framework allows integration of ambiguous information with conventional unambiguous information using a novel approach based on the parametric statistics. A major contribution lies in treating the camera dynamics as completely unknown throughout the dissertation. Furthermore as an approximate solution to tre Bayesian filtering equations, the extended target track-before detect PHD filter as well as multi-sensor CPHD filter are derived in a mathematically rigorous manner. Finally both frameworks are successfully applied to multi-object tracking in image sequence. Other contributions of this dissertations include but not limited to: • A Novel relative velocity based approach, termed sum of relative velocity magnitude (SRVM) to detect moving objects. • A novel 2D to 3D and vice versa velocity transformation for state estimation within image processing phase of tracking. • A SRVM intensity image based visualization and detection approach for multiple moving objects. The final two algorithms include 1. Fusion of unambiguous information using random sets and associated track before detect PHD filter for extended multi-targets. 2. Fusion of ambiguous information using random sets and associated multi-sensor CPHD filter with assumed known measurement to object association (MOA) (see chapter 1 for details on MOA)