A unified model for activity recognition from video sequences

Abstract

We propose an activity recognition algorithm that utilizes a uni ed spatial-frequency model of motion to recognize large-scale differences in action using global statistics, and subsequently distinguishes between motions with similar global statistics by spatially localizing the moving objects. We model the Fourier transforms of translating rigid objects in a video, since the Fourier domain inherently groups regions of the video with similar motion in high energy concentrations within its domain to make global motion detectable.Frequency-domain statistics can be used to isolate the frames that both adhere to our model and contain similar global motion, thus we can separate activities into broader classes based on their global motion. A leastsquares solution is then solved to isolate the spatially discriminative object con gurations that produce similar global motion statistics. This model provides a unied framework to form concise globally-optimal spatial and motion descriptors necessary for discriminating activities. Experimental results are demonstrated on a human activity dataset.

Learning Self-Similarities for Action Recognition Using Conditional Random Fields

“…These accuracy results are very encouraging, specially since we are using very sparse descriptors for the pose (just two per frame). Although higher accuracy results have been reported Ikizler & Duygulu (2007), accuracy of the proposed method is comparable to the well known work of Niebles et al (2006) and also with the recently reported results in Resendiz & Ahuja (2008) posted for the same dataset.…”

“…These surfaces are computed on an extracted foreground of a person performing an action. In contrast to ; Niebles et al (2006); Resendiz & Ahuja (2008), we compute only a few of these surfaces per frame, in fact just two features per frame. Experimental validation on Weizman datasets confirms the stability and utility of our approach.…”

Self-similarity based action recognition using Conditional Random Fields

“…These surfaces are computed on an extracted foreground of a person performing an action. In contrast to [12,10,11], we compute only a few of these surfaces per frame, in fact just two features per frame. Experimental validation on Weizman datasets confirms the stability and utility of our approach.…”

“…5. Cross-validation results for action recognition of the Weizman dataset when the foreground patch is divided into an upper and a lower part for computing the self-similarity surface.reported results in [11] posted for the same dataset.…”