Minimal-latency human action recognition using reliable-inference

Abstract

We present a probabilistic reliable-inference framework to address the issue of rapid detection of human actions with low error rates. The approach determines the shortest video exposures needed for low-latency recognition by sequentially evaluating a series of posterior ratios for different action classes. If a subsequence is deemed unreliable or confusing, additional video frames are incorporated until a reliable classification to a particular action can be made. Results are presented for multiple action classes and subsequence durations, and are compared to alternative probabilistic approaches. The framework provides a means to accurately classify human actions using the least amount of temporal information. q

Descriptive temporal template features for visual motion recognition

“…For human activity or behaviour recognition, many efforts have been concentrated on using state-space methods (Farmer et al (1991)) to understand human motion sequences (Bobick and Davis (2001); Campbell and Bobick (1995); Gao et al (2004); Nascimento et al (2005); Davis and Tyagi (2006)). However, these methods usually need intrinsic nonlinear models and do not have a closed-form solution.…”

Graphical modeling and decoding of human actions

“…The proposed modeling system is substantially differentiated from and possesses advantages over the previously proposed methods based on postures (or key-frames) [1], [2], [3], [4] and Hidden Markov Model (HMM) [5], [6], [7]. Firstly, our model shares postures among the actions and, hence, enables efficient learning from a small number of samples rather than modeling each action with individual HMM which often requires large number of samples to train.…”

“…In an implicit model, action descriptors are extracted from the action sequences of silhouettes such that the action recognition is turned from a temporal classification problem to a static classification one. Proposed action descriptors include moments of Motion Energy Images (MEI) and Motion-History Images (MHI) calculated from silhouettes [6], [13], GMMs to capture the distribution of the moments of the silhouettes [7] or the five extremities [3] corresponding to the arms, legs and head over the period of an action, ignoring the temporal order of the silhouettes in the action sequence, an ensemble of GMMs of cetegory fearure vectors (CFV) [14], the differential geometric surface properties [15] of the spatiotemporal volume formed by the sequence of silhouettes, and the space-time features [16] by utilizing the properties of the solution to the Poisson equation.…”

“…Divis and Tyagi [7] used moments to describe shapes of a silhouette and continuous HMM to model the dynamics. In [4], Kellokumpu et al chose Fourier shape descriptors and classified the postures into a finite number of clusters.…”

Multiview fusion for canonical view generation based on homography constraints

“…[3,4,5,6,10]), with an exception to some efforts in view-independent action recognition ( [21,18]). All view based algorithms require the object of interest to be imaged from a particular vantage point which is not always possible in an unconstrained environment.…”