Research On View-invariant Action Recognition Based On Transfer Learning

View-invariant action recognition is a hotspot and one of the most difficult research areas in computer vision. The recognition rate of existing algorithms vary greatly with the variant viewpoint, especially the recognition effect of top view is not ideal. Knowledge transfer has proven to be an effective solution for bridging different fields, based on which two view invariant algorithms are proposed in this paper.The first one is view-invariant action recognition via bilingual bag of dynamical systems. To begin with, spatio-temporal patches are extracted as low-level feature with the combination of interest point detection and dense sampling algorithm, and each patch is represented as a linear dynamical system(LDS); Then apply nonlinear dimensionality reduction and clustering algorithm to build codebook, consequently, bag of dynamical systems(BoDS) combined with soft-weighting is employed for middle-level representation; Finally, By employing K-singular value decomposition(K-SVD) algorithm, a Bo DS pair corresponding to two view is transformed into a transferable dictionary pair. After that, an orthogonal matching pursuit(OMP) algorithm is applied to the dictionary pair to generate the sparse representation of the action, thus ensure that the same action from different views has the same high-level representation.The second one is a topic-based knowledge transfer algorithm for view-invariant action recognition. Firstly, Spatio-temporal descriptors are extracted from the action videos and each video is modeled by bag of visual words(BOVW). Secondly, Latent Dirichlet Allocation(LDA) is employed to assign topics for the BoVW representation. The topic distribution of visual words(ToVW) is normalized and taken to be the feature vector. Thirdly, knowledge transfer algorithm is adopted to find transferred dictionary pairs, which helps to gain sparse representation of actions under two view pairs as high-level feature for recognition.Multi-class SVM is employed for training and recognition in the above-mentioned two methods. Experimental results on the IXMAS multi-view data set and MuHAVi-MAS data set show the stability and effectiveness of the proposed algorithm.