Spatio-temporal continuous wavelet transform for motion estimation

In this thesis, we introduce a novel framework for motion estimation (ME) based on the spatio-temporal continuous wavelet transform (CWT). In particular, we address the problem of extracting motion parameters of objects from image sequences and consider a wide range of scenarios including noise, temporary occlusions, non-linear motion, and time-varying object signatures. These challenging conditions are typically present in applications such as missile interception, traffic monitoring, and autonomous vehicle guidance. We demonstrate the usefulness of the CWT framework for ME purposes and its performance advantages with respect to competing techniques. Theoretical and practical issues of the implementation of the CWT ME frame-work as well as its properties and limitations are also addressed in this thesis.; In our ME framework, we define global and local energy densities in the spatio-temporal CWT domain. The global energy densities are useful for motion analysis. The local counterparts are used to follow the time evolution of motion parameters and are the computational core of our CWT-based object tracking algorithm. This algorithm sequentially optimizes motion parameters by maximizing the associated energy densities on a frame-by-frame basis. This local optimization allows the algorithm to handle linear and accelerated motion in a natural way. The extended temporal support of the CWT approach allows us to exploit the recent history and resolve temporary occlusions. In addition, the framework's motion-based selectivity, as opposed to the shape-based selectivity associated with conventional ME approaches, allows the CWT framework to better handle scenarios including time-varying object signatures.