Digital crime scene analysis: automatic matching of firing pin impressions on cartridge bottoms using 2d and 3d spatial features

The examination of forensic toolmarks impressed on shot cartridges and bullets is a well known and broadly accepted forensic discipline. The underlying concept is based on two main hypotheses: every firearm owns unique toolmark characteristics which lead to consistent and reproducible impressions on cartridges and bullets. Furthermore, it is possible to differentiate between markings of two different firearms. The application of optical 2D and 3D sensing technologies for acquisition, as well as pattern recognition techniques for automated toolmark examination are currently emerging fields of research in the domain of digital crime scene analysis. In this paper we propose and evaluate a pattern recognition approach for automated firearm identification based on central-fire firing pin impressions. The entire pattern recognition chain is addressed, starting with a confocal microscope for optical data acquisition. The preprocessing covers image enhancement, as well as necessary registration and segmentation tasks for cartridge bottoms. Feature extraction involves 18 firing pin related features from 2D and 3D spatial domain. The classification accuracy is evaluated by using 10-fold stratified cross-validation. Our evaluation approach is two-fold, during the first part we examine how well it is possible to differentiate between two firearms of the same mark and model. During the second part the evaluation is extended to analyze the accuracy of discrimination using six different weapons, whereby each two guns are of the same model. The test set contains 72 cartridge samples including three different ammunition manufactures and six individual 9mm guns. Every possible combination of weapon model, instance and ammunition type is represented by four samples within the test set. Regarding the first evaluation goal a classification accuracy between 87.5% and 100% is achieved. For the second evaluation goal the achieved classification accuracy equates to 86.11%.