Abstract
dc:descriptionIn this Thesis, a set of 4 methods are presented to deal with the problem of unsupervised human action retrieval of 3D mesh sequences. In the first method, a set of state-of-the-art static shape descriptors are used to represent the human body in each frame of the sequences and each mesh sequence is considered as an N-dimensional curve, where N is the dimension of the corresponding shape descriptor. In this method, the temporal correspondence between the frames of the sequences is not taken into account at the descriptor extraction stage. In the second method, a descriptor which incorporates the temporal evolution in mesh sequences is introduced. This descriptor consists of two sub-descriptors: the first is based on a set of differential kinematic features of the trajectories of the centroid extracted in each frame of the sequences, while the second sub-descriptor is based on the first method. An important contribution of this method is an accompanying novel artificial dataset, which was made publicly available for research purposes. The actions contained in this dataset include common human actions and their variations, making the corresponding retrieval problem more challenging. Given the small number of publicly available 4D human action datasets of meshes, this dataset is a very practical contribution of the present Thesis. The third proposed method is based on the observation that the human visual system is able to discriminate human actions without observing the whole body of a moving human but only its joints. A new algorithm for extracting the trajectories of the protrusions of human body is thus introduced. The trajectory-based descriptor proposed in the second method is used to extract a descriptor for each of six salient points of the human body (i.e. the top of the head, the ends of upper and lower limbs and the centroid). The selection of the specific points originates from the fact that the human body can be segmented in six basic parts (i.e. the head, the upper limbs, the lower limbs and the core), so each of the extracted points (as well as the centroid) is a representative point of these segments. The last method is an extension of the third one. Specifically, the Wavelet transformation in various scales is applied on the features extracted by the third method. The final sequence descriptor is composed of a set of statistics, forming a descriptor with constant length, independently from the initial length of the sequences. The experimentation of the two first methods was based on all related publiclyavailable datasets, both real and artificial, including the dataset introduced with the second method. On the other hand, the salient points-based methods can be applied only on clean mesh sequences. In all cases, the retrieval performance is evaluated using standard scalar (Nearest Neighbor, First Tier, Second Tier, Discounted Cumulative Gain) and precision-recall diagrams. The corresponding retrieval performance is at or beyond the state of the art and in some cases it is ideal for the experimentation datasets.
Degree
thesis:*- Grantor dc:publisher
- Democritus University of Thrace (DUTH)
- Year dc:date
- 2019
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Veinidis, Christos
- Βεϊνίδης, Χρήστος
Subjects
dc:subject × 10Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/45970