Learning and recognizing human actions from video via poisson equations

A novel algorithm of learning and recognizing human activities from visual data based on Poisson image and nearest neighbor classifier is presented in this note, where the Poisson image is defined by solving the Poisson equations to re-interpret the proposed spatial motion accumulative image and temporal motion accumulative image. In more detail, firstly, the spatial motion accumulative image and temporal motion accumulative image, which are robust to uncertain duration of action, are abstracted from the binary video with moving human blobs therein; then, the Poisson image is acquired by solving the nine-point difference discretization Poisson equations, which improves the convergence speed greatly compared with the five-point difference discretization, defined on the spatial motion accumulative image and temporal motion accumulative image; thirdly, the Poisson image is partitioned into cells and the distribution histograms of each cell are calculated, lifting which leads to the histogram feature vector describing the action segment; and finally, the nearest neighbor classifier is employed to recognize the human actions through the leave-one-person-out cross-validation benchmark. Experimental results on the public Weizmann activity database confirm the recognition performance of the proposed algorithm.