With the rapid development of the Internet of Things, smart devices carry sensors to collect user data and extract relevant information from the data to track, predict and analyze user activities. However, these methods are usually based on shallow features, which leads to suboptimal performance due to the complexity and the similarity of human activity data. To solve this problem, we propose a new method based on convolutional neural networks and attention mechanisms. This approach integrates multi-head convolution with the attention mechanism, applying attention to multi-head convolutional neural network for improved feature extraction and selection, thereby enhancing the precision of activity recognition. We used the publicly available data set of the WISDM laboratory for validation experiments. By analyzing the experimental results, the proposed method is more accurate than the existing methods, achieving 96.6% of the recognition accuracy.