Abstract:

Falls have become the primary cause of disability due to injury for the elderly. Timely and accurate warning of fall events is an important link to rescue work. In order to improve the accuracy of fall detection, a fall detection method based on a fusion strategy is proposed, which considers both the integrity of high-dimensional digital sequences and the specificity of different dimensions. The input digital sequences obtained from the wrist portable sensor are processed by window segmentation according to the saliency of resultant acceleration, so as to ensure the timing of the data and improve the identifiability of the fall information. The features of air pressure difference and body temperature are introduced to formulate a normalized digital sequence with nine-axis characteristics and explore more information related to fall detection. According to the regression classification results of gradient boosting decision tree (GBDT) and random forest (RF) in ensemble learning, the fusion strategy is considered to obtain the classification identification of whether fall events have happened. Experimental results illustrate that the proposed method achieves higher accuracy of fall detection than the GBDT model and the RF model on self-testing data. Moreover, the proposed fusion strategy also achieves an excellent accuracy of fall detection in the UR Fall and UMA Fall public datasets, validating the effectiveness and generalization of the proposed method.

Key words: fall detection, ensemble learning, fusion strategy, gradient boosting decision tree, random forest