Automated Firearm Identification

Abstract

The examination of firearm related toolmarks impressed to cartridges and bullets is a well known forensic discipline. The application of three dimensional imaging systems and pattern recognition techniques for automatic comparison and matching of topographic data is a central field of research in the domain of digital crime scene analysis. In this work, we introduce and evaluate a novel Multiple-Slice-Shape (MSS) approach with the objective to closer link the preprocessing and feature extraction stages and improve the automated examinations of firearm toolmark surface data. We employ two existing features which are applied to the topography of firing pin impressions and aim at an automatic matching of the shapes based on multiple line-profile measurement. We suggest several modifications of the original Multiple-Angle-Path (MAP) and Multiple-Circle-Path (MCP) features to achieve an optimal integration into the proposed processing pipeline. Our evaluation approach is three-fold. First, we aim at the determination of an initial parameterization for MSS processing and feature extraction. Second, we evaluate the accuracy of discrimination for two firearms of the same mark and model. Third, we evaluate the accuracy using six different weapons. The test set contains 72 cartridge samples including six guns and three ammunition manufactures. Regarding the first evaluation, the results indicate an improvement of the accuracy for both features. Regarding the second evaluation, the achieved accuracy ranges between 67% and 100% for the MAP feature, and between 92% and 100% for the MCP feature. With respect to the third evaluation, the best result is achieved for MAP 32 with 73% and for MCP 15 with 92% compared to 56% and 82% correct classification rate regarding the original versions. It is supposed that various 3D spatial features can be combined and maybe improved by using the proposed MSS approach. We motivate the evaluation of this question for future work.

Forensic comparison of fired cartridge cases: Feature-extraction methods for feature-based calculation of likelihood ratios

“…Statistical models applied in the published research have primarily been classification models rather then likelihood-ratio models. These classification models have included k nearest neighbors (Fischer & Vielhauer [ 30 ], [ 31 ]; Morris et al [ 32 ]), linear discriminant analysis (Thumwarin et al [ 10 ]; Ghani et al [ 27 ]; Chuan et al [ 28 ]), support vector machines (Zhou et al, [ 22 ]), bagged decision trees (Morris et al [ 32 ]), and neural networks (Li [ 33 ]; Leng & Huang [ 29 ]; Morris et al [ 32 ]; Ghani et al [ 34 ]; Giudice et al [ 35 ]; Razak et al [ 36 ]).…”