Fall intention recognition based on semantic interval localization via hard attention for robot-assisted walking
Junxin Chen, Yaoyao Chen, Pengyao Xu, Chao Chen, Rui Song, Weihua Li, Chong Di · 2024
In the care of the elderly or disabilities, fall prevention through wearable sensors has been a key research direction. However, previous studies mainly focused on the detection of the completed fall activity, and could not effectively recognize the action when it just happened. Therefore, based on the acceleration sensor, this article designs a network for acceleration signals, focusing on the recognition of fall action just occurred, so as to achieve the capture of the action before it is completed, which is called fall intention recognition. Inspired by the field of computer vision— fine-grained classification, the recognition of picture categories only needs to pay attention to the unique feature region of the category. Similarly, the recognition of an action does not need the entire and complete signal sequence, but only needs to pay attention to the semantic interval that is different from other activities. In other words, it is the segment when the fall action starts, because the acceleration data varies with the direction of the fall. At the same time, acceleration data, especially those collected on the body, are prone to external interference, so this article dose not abandon the pre-fall data, also combines with linear acceleration data, which considered all of them are distinguishing features of fall intention. It is verified on the UniMiB SHAR standard dataset, and the experimental results show that the fall intention recognition algorithm in this paper is effective.