Automatic Shoeprint Retrieval Algorithm for Real Crime Scenes

This study is to propose a fully automatic crime scene shoeprint retrieval algorithm that can be used to link scenes of crime or determine the brand of a shoe. A shoeprint contour model is proposed to roughly correct the geometry distortions. To simulate the character of the forensic experts, a region priority match and similarity estimation strategy is also proposed. The shoeprint is divided into two semantic regions, and their confidence values are computed based on the priority in the forensic practice and the quantity of reliable information. Similarities of each region are computed respectively, and the matching score between the reference image and an image in the database is the weighted sum. For regions with higher confidence value, the similarities are computed based on the proposed coarse-to-fine global invariant descriptors, which are based on Wavelet-Fourier transform and are invariant under slight geometry distortions and interference such as breaks and small holes, etc. For regions with lower confidence value, similarities are estimated based on computed similarities of regions with higher confidence value. Parameters of the proposed algorithm have learned from huge quantity of crime scene shoeprints and standard shoeprints which can cover most practical cases, and the algorithm can have better performance with minimum user intervention. The proposed algorithm has been tested on the crime scene shoeprint database composed of 210,000 shoeprints provided by the third party, and the cumulative matching score of the top 2 percent is 90.87.