Online motion segmentation using spatially-constrained J-linkage in dynamic scene

Abstract

We present a method for online motion segmentation in dynamic scenes. Here the dynamic scene is a scene without restrictions on the motion of objects and the motion of a camera. Such a scene causes complex interaction of objects and a camera, leading to the generation of trajectories corrupted by noise and outliers. Moreover, no prior knowledge of the number of objects is given, and the number of objects can vary as the motion of objects. In this paper, we deal with clustering of trajectories in dynamic scenes, while estimating the number of objects at the same time. The basic idea is to find motion models that support the motion of trajectories, and cluster the trajectories according to the support motion models, instead of handling trajectories directly for clustering. To do so, we adopted online J-linkage algorithm proposed in [7], an online multiple model estimation method. Based on the observation that points on one object are located nearby, we applied spatially-constrained agglomerative clustering to the J-linkage algorithm. This spatial constraint can drastically reduce searching time in clustering. The proposed method operates in an online and incremental fashion, so that it is applicable to real-time applications. We tested our method on the Hopkins datasets, demonstrating the effectiveness of the method in dynamic scenes.