The problem of object tracking is a problem arising in domains such as computer vision (video surveillance for instance) and cinematographic post-production (special effects). There are two major classes of solution to this problem: region of interest tracking, which indicates a coarse tracking, and space-time segmentation, which corresponds to a precise tracking of the region of interest's contour. In both cases, the region of interest must be selected beforehand on the first, and possibly on the last image of the video sequence. In this thesis, we propose two tracking methods (one of each type). We propose also a fast implementation of an existing tracking method on Graphics Processing Unit (GPU). The first method is based on the analysis of temporal trajectories of salient points and provides a region of interest tracking. Salient points (typically of point of strong curvature of the isointensity lines) are detected in all the images of the sequence. The trajectories are built by matching salient points of consecutive images whose neighbourhoods are coherent. Our first contribution consists in the analysis of the trajectories on a group of pictures, which improves the motion estimation quality. Moreover, we use a space-time weighting for each trajectory which makes it possible to add a temporal constraint on the movement while taking into account the local geometrical deformations of the object ignored by a global motion model. The second method performs a space-time segmentation. The object contour motion is estimated using the information contained in an outer-layer centered on the object contour. Our first contribution is the use of this outer-layer which contains information about both the background and the object in a local context. Moreover, the matching using a statistical similarity measure (residual entropy) allows to improve the tracking while facilitating the choice of the optimal size of the crown. Finally, we propose a fast implementation of an existing tracking method of region of interest. This method relies on the use of a statistical similarity measure: the Kullback-Leibler divergence. This divergence can be estimated in a high dimension space using k-th nearest neighbor distance. These calculations being computationally very expensive, we propose a parallel implementation of the exhaustive search of the k-th nearest neighbors using GPU programming (via the programming interface NVIDIA CUDA). We show that this implementation speeds the tracking process up to a factor 15 compared to a classical implementation of this search using data structuring methods.
展开▼
