Preliminary results with BERT for human activity recognition using skeleton sequences features

The use of artificial intelligence techniques for the recognition of human activity has been an important area of research. Several approaches have been proposed and a large part of them address this problem through vision with conventional RGB cameras. Some of the most significant problems in human activity recognition systems are privacy, the limitations of the operating devices and the comparison of machine learning and deep learning techniques for the prediction of said activity. The present document presents an activity recognition system by means of a recurrent neural network BERT, based on camera vision, using image sequences containing the information of the detection of the pose of the human skeleton for the extraction of characteristics. The proposed method is evaluated with the UP-Fall data-set, outperforming the results of other activity recognition systems that use deep learning techniques and the same data-set. This method achieves an accuracy of 99.14% and a F1-Score of 80.95%.