In this paper pose-based human action recognition is being studied. First, human pose estimation methods, visual structures, principles of operation of convolutional neural networks, components, and the best architectures are examined. Further studies on human pose estimation, their methods and model training operations are presented. The analysis of the studies is followed by an experimental study to compare the performance of the CPM, Hourglass, PoseNet, and OpenPose pre-trained models on mobile devices. The models were compared according to the inference time, the confidence score, and the number of body points detected. The obtained test results showed that PoseNet is the most reliable and fast enough model to be used in human pose based systems for mobile devices. After the experimental study, a prototype of an app for the Android operating system was implemented, which allows recognizing human movements: raising and waving of the left and right hands, and counting the number of pushups performed. During the implementation of the prototype, it was found that the simplest human movements can be recognized by observing the change of the position of key human parts.
Žmogaus judesių atpažinimas realiuoju laiku ; Real-time human motion recognition
2020-06-10
Theses
Electronic Resource
Lithuanian , English
Žmogaus judesių atpažinimas realiuoju laiku ; Real-time human motion recognition
BASE | 2020
|Žmogaus judesių atpažinimas realiuoju laiku ; Real-time human motion recognition
BASE | 2020
|