Low-rank approximation based abnormal detection in the video sequence

摘要

This article presents a new method for abnormal events detection from video sequences by combing the low-rank approximation and sparse combination learning. Motivated by the structured sparsity of video data, the low-rank approximation is introduced to capture a set of normal dictionaries. With the captured dictionaries, the sparse combination learning is utilized to fit training samples and measure the abnormality of testing samples. Multi-scale 3D gradient features, which encode the spatiotemporal information, are adopted to detect abnormal events. The benefits of the proposed method are three-fold: firstly, the low-rank property is utilized to learn the underlying normal dictionaries, which can represent groups of similar normal features effectively; Secondly, the sparsity based algorithm can adaptively determine the number of dictionary bases, which makes it a preferable choice for interpreting the corresponding dynamic scene semantics; Thirdly, the proposed method is efficient and real time detection can be accomplished. Experimental results on public datasets have shown that the proposed method yields competitive performance comparing with the state-of-the-art methods.