**4.1. Computer vision and Forensic analysis**

used in forensic science are ranked according to their precision and the size of the studied object and/or scene. A more detailed description of the individual techniques and their variants

The portability is a common relevant factor, due to the need of moving sensors around the object or scene to carry out a proper data acquisition. Although in some cases, the evidence could be moved to the forensic lab, geomatic sensors should provide a medium degree of

Although some geomatic techniques have showed to complement a full spectrum of possibil‐

In order to surpass the limitations of precision and resolution of TLS, focused on the recording of medium or large scenes, sometimes more precision and resolution are required for the 3D modelling of small objects or evidences. In these cases, the coordinate measuring machine (CMM) or handheld scanner could offer an alternative. These systems allow tactile and discrete point measurement and massive point acquisition using a linear scanner. However, the main

CMM and handheld techniques are highly useful in soft and nonparametric surfaces where any contact will disturbance the previous measurements, or worse, affect the preservation of the evidence. The unique geometric counterpart is the limitation in depth acquisition, being

On the other hand, sometimes the precision is not always a critical factor, being crucial the ability to discover hidden information in the scene. At this point, the ground penetrating radar (GPR) has been proved as useful technique for detection of underground structures, mainly clandestine graves [13] but not limited, being other alternatives the search for buried weapons,

The main advantages are the non‐invasiveness property and the capability to cover wide areas (outdoors and/or indoors). However, its applicability is limited to the specialized training in data acquisition, processing and data interpretation [14], being also critical the data acquisition planning, since the bigger the suspicious area the lower efficiency in its detection. Moreover, this sensor should be combined with an alternative system to georeference it according to an external reference frame. Due to its simplicity, portability, and appropriate precision, the GNSS

For a more extensive review of geomatic/geoscience methods applied to forensic search, please

This section will be focus in three new approaches into geomatic applications in forensic analysis. Firstly, recent advances in computer vision and photogrammetry are described; secondly, highly portable active sensors, known as gaming sensor, as introduced as an low‐

ities, some of them could provide added value in forensic analysis.

problem of CMM is its portability and its lack of radiometry.

unable to acquire data in holes of small diameter or deep holes.

systems are chosen to connect the GPR data with a global frame.

can be found in [12].

16 Forensic Analysis - From Death to Justice

mobility and portability.

drugs, hazardous waste, etc.

**4. Progress and challenges**

refer to [15].

The recent advancements on flexibility, automation and quality due to the algorithmic evolution in photogrammetric and computer vision techniques will be the topics of this subsection.

Although for post‐mortem examination, two‐dimensional photography has been established as 'gold standard' [16], the possibility of getting 3D information overpasses the 2D photogra‐ phy. The main reason is that there is not any data projection, so it is possible to measure different invariant (e.g. distances, surfaces) without any deformation. Some authors have worked with the orthophoto as a metric support of the forensic analysis, but the lack of depth or Z coordinate limits its exploitation in forensic analysis. Other authors have tried to provide 3D metric capabilities directly to the image, generating the solid image, which encloses the RGB values together with *XYZ* coordinates. However, this approach requires other sensors such as laser scanners.

For these reason, the complete 3D documentation and modelling based on images are being employed in several forensic analysis of evidences, such as pattern injuries against injury‐ inflicting instrument in weapon analysis [17], bite mark identification [18], wound documen‐ tation and analysis [19], or forensic pathology [20].

Although image‐based modelling methods have required a long learning curve, the recent advances in photogrammetry and computer vision algorithms and software tools have allowed the automation of workflows opening the use of these tools to non‐experts users [21]. In [21], it is shown how the image approach for 3D reconstruction and dimensional analysis of crime scenes fulfil the forensic requirements in terms of automation, flexibility and quality (**Figure 9**).

The hybridization of both disciplines has made it possible advances on three milestone issues: (1) automation of features extraction, matching entities and image orientation under unfav‐

**Figure 9.** Example of use of the photogrammetry in a crime scene.

ourable conditions; (2) the guarantying of quality in results by means of robust procedures providing greater accuracy and reliability; (3) flexibility, by making it possible to work with any type of image (visible, thermal, infrared, etc.) and any type of camera (calibrated and non‐ calibrated). However, the main milestones regarding image‐based modelling currently are twofold: (i) cope with texture‐less objects (ii) obtaining CAD models from point clouds, which are more useful in forensic analysis.

Since texture‐less objects could appear in any kind of forensic scenario, in the first stage of 3D reconstruction (interest point detection), the matching results will be poor. That is because the local image features (through their descriptors) are unable to provide a robust correspondence among them. For this reason, they present a significant challenge in computer vision and photogrammetry, especially for the image detectors and descriptors. If high amounts of texture‐less object are present in the scene, the number of wrong correspondences in the image orientation process (see Section 3.1) could surpass the efficiency of robust methods, and as a result, the camera orientation would be incorrectly determined.

The most novel automatic approaches to afford nonparametric forms are based on 2D and 3D triangulation strategies, which generate a surface model. Nonetheless, this falls far from what is expected of CAD models in forensic analysis, which must be shaped as solid models with topological relations and properties. Nowadays, there are only semi‐automatic approaches to generate solid models based on three methodologies: (1) fitting of basic primitives for those simple objects which are represented by means of a parametric shape; (2) performing cross‐ sections that, accompanied by shape extrusion operations, enable the generation of the corresponding solid model; (3) fitting of more complex functions of B‐spline type (NURBS‐ non‐uniform rational B‐spline) which, through cross‐sections and sweeping operations all along a path make it possible to generate the solid model. However, all of these methods need manual interaction at the moment of point clouds segmentation and results refinement.
