Clustering and geometrical features for classi cation of humans in three dimensional data
| dc.contributor.author | Hedin, David | |
| dc.contributor.author | Wendel, Johan | |
| dc.contributor.department | Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper | sv |
| dc.contributor.examiner | Boyraz Baykas, Pinar | |
| dc.contributor.supervisor | Pålsson, Albin | |
| dc.date.accessioned | 2019-10-03T17:23:43Z | |
| dc.date.available | 2019-10-03T17:23:43Z | |
| dc.date.issued | 2019 | sv |
| dc.date.submitted | 2019 | |
| dc.description.abstract | Computer vision is a rich research field that uses images from normal or technically specific cameras to perform tasks ranging from surveillance to autonomous driving. Computer vision with depth images are however relatively new. Depth images add a third dimension to the image by giving every pixel a depth value and can be produced with several different camera types. The implications for depth imagery is that if an object can be classified in a depth image as e.g a trashcan, a cat or a human, the nature of the data immediately also gives us the distance and position of that object. The extra data from these cameras can enable estimating the volumes and sizes of objects in an image with some extra processing. The focus of this master thesis is on the processing and analysis of this depth data to enable object identification and human classification. Much research has been on analysing 3D data from the 2D perspective, in this thesis the captured data is first converted to Cartesian coordinates before attempting classification, yielding further possibilities. The goal of this thesis was to find if there are some anthropomorphic-geometrical features that can describe the human body well enough to accurately classify humans in the Cartesian data. The features are used in two ways, as a Heuristical-geometrical filter and as features for a support vector machine. Furthermore the thesis presents a successful dynamic adaption of the fast-DBSCAN (Density Based Spatial Clustering of Applications with Noise clustering) algorithm for 3D Cartesian data and a slice method for finding local maxima of point cloud objects. The results show that anthropomorphic-geometrical features can to an extent be used to classify Cartesian point cloud data. Low resolution cameras has potential for classification purposes as resolution seem to have little effect on geometrical classification as long as human resolution is no less then 20px vertically. Some further work would be needed to create a anthropomorphic-geometrical for real world application. | sv |
| dc.identifier.coursecode | MMSX30 | sv |
| dc.identifier.uri | https://hdl.handle.net/20.500.12380/300400 | |
| dc.language.iso | eng | sv |
| dc.relation.ispartofseries | 2019:66 | sv |
| dc.setspec.uppsok | Technology | |
| dc.subject | Depth camera | sv |
| dc.subject | moving platform | sv |
| dc.subject | human machine interaction | sv |
| dc.subject | localisation | sv |
| dc.subject | clustering | sv |
| dc.subject | human classification | sv |
| dc.subject | 3D | sv |
| dc.title | Clustering and geometrical features for classi cation of humans in three dimensional data | sv |
| dc.type.degree | Examensarbete för masterexamen | sv |
| dc.type.uppsok | H | |
| local.programme | Biomedical engineering (MPBME), MSc |