MD-BiLSTM: An efficient artificial intelligence algorithm for human activity recognition

Currently, with the growing enthusiasm for augmented devices, intelligent exoskeletons are gradually attracting the attention of researchers. Enhancing human activity recognition will undoubtedly propel smart assistive exoskeletons into households. Traditional convolutional layers typically have fixed receptive fields, which limit their ability to capture features at different scales. This paper introduces multi-scale dilated convolution to extract typical features of human activity data across various spatial scales, and combines it with bidirectional long short-term memory networks and self-attention to achieve accurate human activity recognition, known as the MD-BiLSTM model. This design meets the need to capture spatial information at different scales in human movements while leveraging the combination of BiLSTM and Self-attention to build multi-level models of temporal data, significantly improving the accuracy and robustness of complex actions in HAR tasks, which is crucial for detecting complex human activities. The effectiveness and superiority of the MD-BiLSTM model have been verified through the public dataset HuGaDB and our own data. Finally, we further verified the effect of the algorithm on the edge computing platform on our own platform and achieved satisfactory results