2022
Mike Kusluski
Atlanta, GA
Attendees will learn the strengths and limitations of a machine-learning approach to muzzle-to-target distance estimation, as well as how to optimize reference series when only a limited number of suspect cartridges are available.
Firearms examiners commonly analyze gunshot residues on clothing in order to estimate the muzzle-to-target distance. Most examiners use the Modified Griess Test to visualize GSR patterns on clothing then perform a visual comparison to known-distance reference patterns to derive a distance estimate. Photonic imaging of gunshot residue on fabrics using a high-resolution flatbed scanner modified for blue-green fluorescence and infrared reflectance has been described previously. In this phase of research, the ability to quantify GSR patterns and generate muzzle-to-target distance estimations by machine learning was assessed.
One-hundred-twenty (120) known-distance targets (12 distances at 10 replicates each) were produced at 3-inch (7.6 cm) increments from 3 to 36 inches (7.6 to 91 cm) using a .40 S&W caliber Glock model 22 Gen 3 semiautomatic pistol in a Ransom International® Master Rest and CCI Blazer brand 180 grain FMJ ammunition. An additional 30 unknown-distance targets were prepared within a 3 to 36 inch range, measured to within 0.5 inch (1.3 cm) resolution.
After scanning, target images were enhanced and analyzed using Image Access Forensic Analyzer software. Ten reference series were established using one replicate from each known-distance increment in sequential order. Each of the 30 unknown-distance target patterns was compared against each of the ten reference series independently, for a total of 300 separate analyses.
Success was defined by the following specifications:
1. The acceptance criteria for individual estimates would vary by distance (+/- 3 inches for a shooting distance less than 12 inches; +/- 6 inches for a shooting distance between 12 and 24 inches; and +/- 9 inches for a shooting distance between 24 and 36 inches). This is based on common practice among examiners.
2. The error rate would reflect the percentage of estimates falling outside their distance-based acceptance criteria.
3. A successful combined average of absolute error for all estimates would be below 25%.
Of the 300 analyses, 20 estimates were outside the acceptance criteria for an error rate of 6.7%. It is notable that all 20 of the error estimates were associated with three unknown-distance targets, which were subsequently determined to be anomalous. This indicates that the errors were likely due to shot-to-shot variation in the unknown-distance targets rather than the reference series or analytical method.
Additional analyses were conducted to determine the estimation capability possible had only a limited number of cartridges (3 – 9) been available to produce reference series. This analysis indicated that successful estimates are possible (and in many instances with higher accuracy) with a smaller number of cartridges, if the known-distance increments are properly selected.