The invention discloses a human body posture recognition method based on a convolutional neural network. The method comprises the steps that original data of a mobile sensor are collected and labeled,data frequency down-sampling and normalization processing are conducted, a training set and a test set are divided, the convolutional neural network is trained, and a model is transplanted to an Android terminal for human body posture recognition. The method is used for human body posture recognition according to a convolutional neural network. According to the method, a Spit-Transform-Merge strategy is introduced into the implementation of the method; a group of Lego convolution kernels with a smaller channel number is provided, the convolution kernels are stacked according to a random mapping and cyclic matrix method so as to realize convolution operation, and finally, generated Lego feature maps are vertically combined and sent to a classifier through a full connection layer for sensordata identification. The method has the advantages of being high in recognition speed, high in recognition accuracy, small in calculation amount, high in generalization capacity and the like, and meanwhile the method plays a very important role in smart home, health detection, motion tracking and the like.