Type of Document	Master's Thesis
Author	Khatod, Hemant Narendra
Author's Email Address	hkhato1@paws.lsu.edu, hemantkhatod@hotmail.com
URN	etd-11102004-121616
Title	Towards Automation of Forensic Facial Reconstruction
Degree	Master of Science in Mechanical Engineering (M.S.M.E.)
Department	Mechanical Engineering
Advisory Committee	Advisor Name Title Warren N. Waggenspack Jr. Committee Chair Dimitris E. Nikitopoulos Committee Member Michael C. Murphy Committee Member
Keywords	local fitting automatic registration NURBS
Date of Defense	2004-11-05
Availability	unrestricted
Abstract Forensic facial reconstruction is a blend of art and science thus computerizing the process leads to numerous solutions. However, complete automation remains a challenge. This research concentrates on automating the first phase of forensic facial reconstruction which is automatic landmark detection by model fitting and extraction of feature points. Detection of landmarks is a challenging task since the skull orientation in a 3D scanned data cloud is generally arbitrary and unknown. To address the issue, well defined skull and mandible models with known geometric structure, features and orientation are (1) aligned and (2) fit to the scanned data. After model fitting is complete, landmarks can be extracted, within reasonable tolerance, from the dataset. Several methods exist for automatic registration (alignment); however, most suffer ambiguity or require interaction to manage symmetric 3D objects. A new alternative 3D model to data registration technique is introduced which works successfully for both symmetric and non-symmetric objects. It takes advantage of the fact that the model and data have similar shape and known geometric features. Therefore, a similar canonical frame of reference can be developed for both model and data. Once the canonical frame of reference is defined, the model can be easily aligned to data by a euclidian transformation of its coordinate system. Once aligned, the model is scaled and deformed globally to accommodate the overall size the object and bring the model in closer proximity to the data. Lastly, the model is deformed locally to better fit the scanned data. With fitting completed, landmark locations on the model can be utilized to isolate and select corresponding landmarks in the dataset. The registration, fitting and landmark detection techniques were applied to a set of six mandible and three skull body 3D scanned datasets. Results indicate the canonical axes formulation is a good candidate for automatic registration of complex 3D objects. The alternate approach posed for deformation and surface fitting of datasets also shows promise for landmark detection when using well constructed NURBS models. Recommendations are provided for addressing the algorithms limitations and to improve its overall performance.
Files	Filename       Size       Approximate Download Time (Hours:Minutes:Seconds)     28.8 Modem   56K Modem   ISDN (64 Kb)   ISDN (128 Kb)   Higher-speed Access    Khatod_thesis.pdf 10.79 Mb 00:49:57 00:25:41 00:22:28 00:11:14 00:00:57