An investigation into incorporation of temporal awareness through convolutional recurrency
Examiner: Irene Gu, Dept of Electrical Engineering
Finding the exact pose of a human is the focus of, or requirement for many different fields of study. This is traditionally done with markings attached to the body of the subject(s). Recently there has been an emergence of markerless solutions using machine learning. The pose is then often found by locating certain joints of the subject(s) in 2D image space.
In this thesis, we have focused our study on how the temporal data of a video can be utilised to increase the performance in such markerless human pose estimation systems. We have used deep neural networks by employing convolutional layers as well as two different convolutional-recurrent memory cells. Several different network architectures of the systems were tested. Our experiment results have shown that the usage of temporal data does reduce the estimation error as compared to the baseline. There seemed a positive effect on the precision of the estimation over time if several recurrent layers are stacked, however, at the cost of increased computations. More details and results will be shown in the presentation.
Student project presentation
Landahlsrummet (room 7430), Hörsalsvägen 9, 7th floor
28 January, 2020, 10:30
28 January, 2020, 11:30