Finding the evidential potential of solid target damage from shotguns using red light laser scanning and machine learning.

The use of a shotgun in criminal damage presents unique challenges to the discipline of shooting incident reconstruction because typical evidence associated with a firearm discharge can be missing (such as the casing which provides a lot of information on the calibre and type of ammunition used, which with a shotgun remains with the weapon until manually removed (Haag, 2021)). Investigators typically make use of witness panels and measurement of the radial area to perform distance estimation which provides a range of distances a discharge could have come from. Other information such as muzzle and impact velocity is gathered by using automatic detection methods such as chronography and high-speed photography. The technology of laser scanning has been shown to be applicable for gathering evidence in anthropology, medicine and blood spatter analysis. The use of this technology on a shotgun damage site has not been fully investigated with regards to the use of 3D data as opposed to the traditional 2D data currently used and whether using this data could be of later use by investigators. Furthermore, the use of machine learning to analyse the information recovered has been shown to be of potential within the field of distance estimation. By utilising the laser scanning data recovered, prediction of muzzle-to-target distance, muzzle velocity and impact velocity was investigated to explore if the additional information found could be used in an objective fashion to enhance accuracy and minimalize subjectivity. A method was developed for using a red light laser scanner to capture damage from 3 different common building materials (concrete slab, plywood and sheet steel) and analyse these meshes using Geomagic X metrology software. 12Ga, number 7.5 birdshot ammunition was fired at 45 targets (15 from each material) over 3m, 5m and 7m distances (5 targets from each material at each distance). Velocity data from the muzzle was captured using a ballistic chronograph and impact was calculated from high-speed camera footage. Data was collected, processed, normalised and added to MATLAB regression learner where Principle Component Analysis (PCA) was applied as well as leave one out processing (LOOP). Inputs were tried in differing combinations to find the optimal inputs for prediction of muzzle-to-target distance, muzzle velocity and impact velocity. These predictions were averaged by distance as is done in distance estimation in literature. It was found that scan data was critical to prediction in all outputs and that differing materials need different combinations of input, algorithm and principal component analysis (PCA). The behaviours of the material at impact played a role in affecting the optimal input combination for prediction. The average predicted muzzle velocity differed from the average true muzzle velocity by 3.2m/s (0.75%). Average predicted impact velocity differed from average true impact velocity by 8.3m/s (1.74%) and the average predicted muzzle-to-target distance differed from the average true distance by 0.62m (14.41%). The thesis provides a proof-of-concept study into the use of laser scanning coupled with machine learning for forensic shooting incident reconstruction. The thesis demonstrates that laser scanning is capable of maximising evidence from a discharge damage site whilst the machine learning provides an objective approach to estimation.