Due to the serious problem of the world’s aging population and the harm caused by unintentional falls to elderly individuals, the question of how to precisely identify falls has steadily attracted the public’s interest. In this article, a novel deep learning (DL) model, a combination model of a temporal convolutional network and gated recurrent unit (TCN-GRU) architecture, is proposed to obtain high-level features for classification. We evaluate its relative performance against two widely used machine learning (ML) based classifiers and six DL architectures using two popular open-source datasets collected using inertial sensors. Our algorithm results show that the proposed method outperformed other algorithms in nearly all four performance metrics we examined, for the datasets we tested. For the MobiAct dataset and Mosi-F dataset (which is a mixture of the MobiAct dataset with the Sisfall dataset), the prediction accuracy reached 99.5% and 97.6%, and the F1_Score reached 98.9% and 97.6%, respectively, demonstrating satisfactory performance. Moreover, the proposed algorithm had higher detection accuracy despite a small data volume, and it correctly detected all types of fall events from ten primary daily activity groups. •A novel model is proposed to capture the different elderly activities accurately.•A new enriching fall-detection classifier with less motion information is proposed.•The method is better with superior classification accuracy and better convergence.