Kalman Filtering Motion Prediction Forrecursive Spatio - Temporal Segmentation

In the framework of computer vision, the spatio-temporal se-gmentation procedure plays a central role. It aims at identifying in the input image, semantically meaningful features that are relevant for the problem at hand. In this paper, these features are selected to be the objects forming the scene. The objects are deened by their properties of temporal and spatial coherence through the video sequence. They provide a complete partition of the scene into its constituent components. Furthermore, the characteristics of the objects permit to track them through time. In this paper, a technique based on a discrete Kalman lter algorithm is proposed to follow the trajectory of the objects. The aim is to obtain a precise prediction of their position and motion. The accurate prediction improves both the recursive spatio-temporal segmentation and object tracking performances, enabling a high level understanding of the scene dynamics. The derived scene representation obtained nds applications in various domains. For instance, it is very well suited for dynamic scene analysis where a deep scene understanding is required. It is also very appealing in the context of second generation video coding. In this paper, objects are seen as entities which are both temporally and spatially coherent over multiple frames throughout the sequence 1, 2, 3]. In order to nd these objects, we use the algorithm proposed by Moscheni 4, 5]. This is an unsupervised technique, that robustly segments a video sequence in terms of multiple moving objects and tracks them through time. The algorithm is composed of two parts, namely, a recursive spatio-temporal segmentation method and an object tracking method. These two parts interact and mutually innuence each other. Starting from the segmentation of the previous frame, the recursive spatio-temporal segmentation aims at robustly partitioning the current frame. This is achieved in three successive steps: partition projection, region validation and constrained spatio-temporal segmentation. Through these three successive steps,