Crimes in our society, increasing in volume and sophistication, have determined the need for knowledge and use of scientific methods to their prevention and investigations, deserving significant attention from governments and researchers. The forensic investigation of a crime is a complex process that starts at the scene of the event and continues on the laboratory, providing the court, where the trial is conducted, with information and arguments needed to confirm the occurrence of the offense and to identify suspects. Investigators need technical support and knowledge to achieve efficient and effective results. This thesis presents three Computer Vision methods that can be applied in forensic investigations: one for photogrammetry, which allows measurement of objects’ heights in images; a framework to the recovery of footwear impressions from crime scenes; and a method for image reconstruction from a collection of small unordered fragments. The process of obtaining measurements in images – photogrammetry – is a customary requirement in Forensic Science. In a trusted system, the estimated height of a person can be used to corroborate or refute pieces of evidence. In the second method, multiview stereo techniques are used to obtain a three-dimensional model, reconstructed from photographs taken from footwear impressions at the crime scene. In the third method, the need for reconstructing shredded images is explored. Photographs can be shredded in order to hide information and it is up to the field of Computer Forensics to explore methodologies for the automatic reconstruction of such images, since their manual assembling is very laborious. The contributions of this thesis are: (a) new effective vanishing point detector that achieves better accuracy than the previous methods in the literature and that was designed to facilitate the process of making measurements in single 2D image; (b) a new methodology for the recovery of 3D footwear impression, from images taken from the evidence at different angles, and its validation in comparison to methods currently used in practice (c) new deterministic method, based on quadratic programming, to reconstruct images from a collection of small rectangular fragments, providing more accurate results than other proposed methods in the literature.