A Prototype Learning Framework Using EMD: Application to Complex Scenes Analysis

Abstract

Abstract-In the last decades, many efforts have been devoted to develop methods for automatic scene understanding in the context of video surveillance applications. This paper presents a novel non-object centric approach for complex scene analysis. Similarly to previous methods, we use low-level cues to individuate atomic activities and create clip histograms. Differently from recent works, the task of discovering high-level activity patterns is formulated as a convex prototype learning problem. This problem results into a simple linear program that can be solved efficiently with standard solvers. The main advantage of our approach is that, using as objective function the Earth Mover's Distance (EMD), the similarity among elementary activities is taken into account in the learning phase. To improve scalability we also consider some variants of EMD adopting L 1 as ground distance for one and two dimensional, linear and circular histograms. In these cases only the similarity between neighboring atomic activities, corresponding to adjacent histogram bins, is taken into account. Therefore we also propose an automatic strategy for sorting atomic activities. Experimental results on publicly available datasets show that our method compares favorably with state-of-the-art approaches, often outperforming them.

Exploiting Sparse Representations for Robust Analysis of Noisy Complex Video Scenes

“…The proposed method has some similarity with our previous works on EMD clustering [5,6], as nonnegative matrix factorization is practically a clustering algorithm. However, with respect to [5,6], this method is more scalable and produces more interpretable results due to the sparse constraints.…”

“…This paper follows recent works on non-object centric analysis of complex scenes [1][2][3][4][5][6]. However it departs from many previous works [1][2][3][4] as we do not rely on Probabilistic Topic Models for inferring high-level activities.…”

“…Many of them explicitly focused on analyzing complex and crowded scenes depicting public spaces in video surveillance applications [1][2][3][4][5][6]. Visual analysis of such complex scenes faces two main problems.…”

“…These features are inevitably uncertain and noisy. Recent works [1][2][3][4][5][6] have shown how powerful statistical machine learning approaches can be used to implicitly handle noisy, uncertain visual information, leading to excellent results in terms of salient behaviors and anomalous activities identification. However none of the previous works have tackled explicitly this issue, i.e.…”

“…However in this paper, to calculate motion features, we do not rely on noisy optical flow vectors but adopt a representation based on short trajectory snippets. Differently from previous works [1][2][3][4][5][6], we model the task of extracting salient activities as a matrix factorization problem and we consider as objective function the Earth Mover's Distance (EMD) [7], which is well-known to be a robust metric in case of noisy histogram comparison. To further reduce the influence of noisy data we also constrain the computed vector basis to be sparse.…”