| Abstract [eng] |
In this Master thesis, the identification and evaluation of Tinetti transition exercises using human pose based deep neural networks is being analyzed. The identified existing datasets are not suitable for correctness estimation, therefore a new Tinetti dataset is created containing examples of performing correct and incorrect exercise versions. The collected dataset is augmented with data from several existing datasets. Human pose keypoints are extracted using MediaPipe BlazePose. Identified and proposed the examples of multiobjective CNN neural network architectures capable of determining the Tinetti exercise and its correctness. The best suggested model achieved 77.65% exercise classification accuracy and 65.7% exercise correctness accuracy. |
