The invention discloses human body behavior recognition based on a logarithmic Euclidean space BOW (bag of words) model, and belongs to the technical field of digital image processing. The recognition comprises the steps: firstly enabling an input video to be divided into video segments which have a fixed length and are overlapped; secondly cutting each video segment into space-time cubic blocks which have the fixed size and are partly overlapped; thirdly extracting a gradient and a light stream feature covariance or a shape feature covariance of each space-time cubic block, and carrying out the dimension reduction of a covariance matrix through employing a symmetric positive definite matrix dimension reduction method. The method carries out the logarithmic change of the covariance matrix, extracts the triangular features of a logarithmic covariance matrix, and converts the triangular features into a logarithmic Euclidean space vector. The method carries out the behavior modeling for the logarithmic Euclidean space through employing the BOW model, carries out the clustering of behavior characteristics through employing spectrum clustering to generate a codebook, and codes the behavior characteristics through employing the LLC (Locality-constrained Linear Coding) technology. A nonlinear support vector machine is used for the training, recognition and classification of the behavior characteristics. The method is used for the recognition of human body behaviors, and is great in robustness.