Currently, research on data acquisition hardware systems for human activity recognition is limited, and there is a problem of a lack of diverse and generalized reference datasets. In this paper, we design a low-power data acquisition system that supports real-time data transmission and propose a data acquisition method with randomness and crossover. Firstly, a low-power acquisition platform is built for data acquisition, wireless transmitting and receiving, and pre-processing; secondly, a comprehensive and accurate data acquisition scheme is developed to improve the generalization of the new dataset; and finally, a 2D-CNN neural network is used to train a model for the acquired dataset in different modes. The experimental results demonstrate that the designed data acquisition system has a reasonable structure and low power consumption, ensuring real-time data transmission. The application of this system greatly improves the quality of the dataset. The obtained dataset achieves an accuracy of 92.54% on deep learning models. Compared to traditional datasets, the new dataset shows significantly better performance in human activity recognition tasks. The development of this data acquisition system and dataset provides convenience for neural network applications.