Video event classification and anomaly identification using spectral clustering

Abstract

This paper proposes a spectral clustering based methodology to classify video events and to detect anomalies. Feature trajectories from objects in a video are modelled, compared and clustered in order to classify the detected object events. Principles of normalized spectral clustering are used with modifications to affinity structure. A novel method for determining spectral clustering parameters based on Eigen structure of the affinity matrix is introduced. Employment of unsupervised learning for event classification is made possible by the proposed successive cluster identity labelling algorithm. A mechanism to identify abnormal events under the context is also introduced. The effectiveness and the robustness of the proposed methodology are demonstrated through experiments conducted on video streams focusing on human motion patterns.

Visual event classification with human like perception

“…Hence, any automaton on this task should incorporate this adjustability of its opinion as for user's intension. Many work in literature [9,10] have introduced a tunable threshold in capturing this sense, where any event exceeding the disparity threshold from existing classes would be marked as an anomaly. However, this proposition tends to violate our intension of holding event connectivity.…”

“…2: Defined W to be diagonal matrix whose (i,i) element is the sum of A 's ith row, and compute the normalized Laplacian L, number of cases belonging to K classes, the algorithm uses a disparity matrix, D of size nxn with the initial parameters K and (J.The parameters D , K and (J are calculated as for [10].…”

“…Existing approaches for marking an anomalous detection is highly depended upon a crisp threshold distance from a representative event from identified existing classes [9,10,11]. However, such firm consideration of a representative case on defining a divergent case would tend to violate our own original intention of clustering to infer events on their connectivity.…”

“…Anomaly detection has been a major concern in many event classification schemes [3,9,10]. Existing approaches for marking an anomalous detection is highly depended upon a crisp threshold distance from a representative event from identified existing classes [9,10,11].…”

Comparison of Two Algorithms for Land Cover Mapping Based on Hyperspectral Imagery

“…No further dimensionality reduction was used in this stage. Instead, to capture the slight variations of the spectral characteristics, spectral clustering was used [19]- [21]. In this, the points in the 69 dimensional space, representing each pixel of the image (except the pixels containing water) were clustered into two groups, based on their similarity in Hyperspectral Image characteristics.…”