Moving Objects Segmentation Based on DeepSphere in Video Surveillance

Abstract

Segmentation of moving objects from video sequences plays an important role in many computer vision applications. In this paper, we present a background subtraction approach based on deep neural networks. More specifically, we propose to employ and validate an unsupervised anomaly discovery framework called “DeepSphere” to perform foreground objects detection and segmentation in video sequences. DeepSphere is based on both deep autoencoders and hypersphere learning methods to isolate anomaly pollution and reconstruct normal behaviors in spatial and temporal context. We exploit the power of this framework and adjust it to perform foreground objects segmentation. We evaluate the performance of our proposed method on 9 surveillance videos from the Background Model Challenge (BMC 2012) dataset, and compare that with a standard subspace learning technique, Robust Principle Component Analysis (RPCA) as well as a Deep Probabilistic Background Model (DeepPBM). Experimental results show that our approach achieved successful results than other existing ones.