The invention discloses a human body action recognition method of a neural network based on an attention mechanism, and provides an end-to-end trainable network which comprises a deep convolution sub-network and an attention sub-network and is used for recognizing human actions from skeleton data. The human body action recognition method comprises the following steps: firstly, encoding a skeletonsequence into a color space-time diagram, feeding the color space-time diagram into a deep convolutional sub-network to extract deep features, and mapping the deep features into a label space by usinga full connection layer; in the attention sub-network, extracting manually-made features representing the importance degree of joint movement, learning attention weights through simple and effectivelinear mapping, and mapping the result into a label space through a full connection layer; and multiplying and fusing the results of the two to obtain the final recognition accuracy. According to thehuman body action recognition method, effective deep features can be automatically extracted from the data to the maximum extent. The network structure disclosed by the invention comprises two sub-networks, and the two networks are simultaneously subjected to joint training in an end-to-end manner without post-processing.