Algorithms for trajectory integration in multiple views

Abstract

This thesis addresses the problem of deriving a coherent and accurate localization of moving objects from partial visual information when data are generated by cameras placed in di erent view angles with respect to the scene. The framework is built around applications of scene monitoring with multiple cameras. Firstly, we demonstrate how a geometric-based solution exploits the relationships between corresponding feature points across views and improves accuracy in object location. Then, we improve the estimation of objects location with geometric transformations that account for lens distortions. Additionally, we study the integration of the partial visual information generated by each individual sensor and their combination into one single frame of observation that considers object association and data fusion. Our approach is fully image-based, only relies on 2D constructs and does not require any complex computation in 3D space. We exploit the continuity and coherence in objects' motion when crossing cameras' elds of view. Additionally, we work under the assumption of planar ground plane and wide baseline (i.e. cameras' viewpoints are far apart). The main contributions are: i) the development of a framework for distributed visual sensing that accounts for inaccuracies in the geometry of multiple views; ii) the reduction of trajectory mapping errors using a statistical-based homography estimation; iii) the integration of a polynomial method for correcting inaccuracies caused by the cameras' lens distortion; iv) a global trajectory reconstruction algorithm that associates and integrates fragments of trajectories generated by each camera.