Identificação de Trajetórias Espaço-Temporais de Movimentos em Vídeo

Abstract

The goal of this work is to identify flow lines of moving objects in videos. These flows can then be used to represent the underlying movement. Unlike studies using shapebased representations, our work uses the image information inside and outside the object to build denser flows for motion representation. We begin by segmenting each video frame into a large set of non-overlapping uniform-color and texture regions called superpixels. Then, each superpixel is characterized as a model appearance and is represented by its centroid. These superpixels are then tracked over a number of frames to generate flow lines. Tracking is done using a function that calculates the cost of matching corresponding regions in subsequent frames, taking into account both the correspondence between regions and the consistency of their neighborhoods. The ideas reported in this work can be applied to the task of creating motion patterns, which in turn are important for identifying complex movements of objects and predict behavior. Experiments are performed on a set of synthetic and real-world motion sequences.