Search

The challenges of human behavior recognition based on sensor data often require addressing the needs of various new users in real-world situations, which leads to difficulties in personalizing user data and resolving concerns related to resolution. This relates to the fact that persons with different characteristics who engage in the same behavior inevitably provide varying data discrepancies. Models that are faced with unfamiliar users encounter difficulties in accurately recognizing expected behavior, and it is not practicable to gather a large amount of training data every time the model needs to be reconfigured for new users. This study introduces a novel approach known as the FSLBR model, which combines the domains of few-shot learning and behavior recognition algorithms. Initially, a meta-learning technique focused on optimization is employed to categories datasets according to user categories. The paradigm of few-shot learning has shown significant effectiveness by leveraging a limited amount of data for new tasks. The FSLBR method and an attention-centered Memory module are then smoothly integrated into the FSLBR model. In the field of behavior recognition, it is observed that a small selection of data is sufficient for good classification when dealing with new users. The capacity of the model network to extract and extrapolate data features is improved. The empirical experiments conducted on the MEx dataset support the assertion that the proposed FSLBR model demonstrates improved performance under the few-shot learning framework compared to standard deep learning methods. • Applying the ideas and methods of MetaLearning to field of behavior recognition. • MetaLearning method employed to train multiple tasks, and summarizing inductions. • Highly generalized initialization parameters, allowing model update high accuracy. • Building upon MAML-GM framework, enhancing the model's classification capabilities. [ABSTRACT FROM AUTHOR]