Abstract: In order to obtain an accurate description of human action in videos,a novel human action recognition algorithm using spatio-temporal co-occurrence features and improved VLAD is proposed.In the feature extraction step,the traditional dense interest points in the videos are refined by the expanded human rectangular box.Dense trajectories are obtained by tracking the refined points using optical flow field.And spatio-temporal co-occurrence features are extracted in the spatio-temporal volume along the trajectory.In the feature encoding step,each feature vector is assigned to multiple visual words from the codebook,which brings the complementary shape information within the encoding process.Then a linear combination of these basis to reconstruct each descriptor is learned under the least square error criterion.The linear combination coefficients are taken as the weights of VLAD to accumulate the residuals.Experimental results show that the proposed algorithm achieves the state-of-the-art result on KTH,YouTube and HMDB51 datasets.