MULFE: A Sensor-Based Multilevel Feature Enhancement Method for Group Activity Recognition

The recognition of group activities from sensor data faces challenges in effective feature extrac-tion, especially the difficulty of expressing the dynamic changes in member actions, the positional relationships and collaborative relation-ships among members. To address this, this paper pro-poses a sensor-based multi-level feature enhancement method (MULFE) for group activity recognition. MULFE utilizes an individual action feature extraction network to extract individual action features, and constructs a group location-level feature enhancement module to capture the group location interaction features among individuals. By combining group location interaction features with individual action features using attention-weighted fusion, location-level enhanced group activity features are achieved, with which the representation of multiple individual features within the group and the complex relational features of individual spatial locations are enhanced. Furthermore, a group spatiotemporal-level feature enhancement module based on CAMLP-Mixer network is designed, using multi-layer perceptron (MLP) to achieve feature interaction and integration to further obtain group spatiotemporal features. The group spatiotemporal features are combined with the location-level enhanced group activity features to generate the multi-level enhanced group activity features, making the model more suitable for understanding complex group activities. Experiments are conducted on two self-built datasets, UT-Data-gar and Garsensors, to validate and analyze the performance of MULFE. The experimental results demonstrate that MULFE can effectively recognize group activities, particularly maintaining high accuracy and strong robustness in situations with random changes in group size.