Real-Time Moving Objects Segmentation based on RGB-D camera
PubDate: March 2020
Teams: Beihang University
Writers: Rui Zhu; Yongjia Zhao
Segmentation of moving objects in a scene is difficult for nonstationary cameras, and especially challenging in the presence of fast and unstable environment. To obtain the accurate and real-time segmentation result, we propose an efficient algorithm that combine the ISODATA clustering with scene flow to segment the moving objects using RGB-D cameras. Frist, we apply the ISODATA clustering method to divide the images into geometric clusters. Then, we compute the residuals of the different clusters and labels the clusters to static scene and moving scene. At last, clusters with moving labels are used to compute the scene flow. We segment the moving objects according to the scene flow estimation and residuals. In this way, we make it possible that segment the moving objects from a moving platform in real-time. We test our algorithm on TUM Datasets and Princeton Tracking Benchmark Datasets and result shows that our method can segment the moving objects in a very low runtime without damaging the accuracy at the same time.