| Abstract [eng] |
The dissertation presents three algorithms that solve the problems of the dissertation. The first algorithm, Agrast-6 neural network, automatically segments depth images and finds the human body in them. Agrast-6 is based on the ideas of the SegNet neural network but uses a lot fewer parameters. The proposed neural network can be applied in larger systems where one of the data processing steps is extracting human silhouettes from depth images. The second algorithm also segments the human body in depth images but also uses user input, therefore it is semi-automatic. It is based on the ideas of Euclidean clustering. Three improvements are proposed – skipping segmented nodes, skipping fully-segmented branches, and using an auto-expanding bounding box instead of a set of spheres. These improvements greatly reduce the processing time. This algorithm was successfully applied to prepare a dataset of 220k images in a relatively short time. Both algorithms greatly speed up data processing at a cost of segmentation accuracy. The third algorithm fuses skeletons from different Kinect sensors into a single, more accurate super-skeleton. This algorithm increases the accuracy of Kinect’s skeletal data compared to the use of a single sensor. The skeleton fusion algorithm was applied in human skeleton load analysis during physical activities. |
