DEVELOPMENT OF A FOOT PRINT DETECTION SYSTEM USING IMAGE PROCESSING AND MACHINE LEARNING TECHNIQUES

Abstract

Identifying people especially criminals or crime suspects with footprints is still in the emerging state especially in third world or developing countries like Nigeria. Forensic experts still place much attention on other features of the body that can be secured at crime scene other than footprint which is mostly inevitable. The unfortunate thing is they are getting smarter by protecting this features from being left behind as evidence against them.. Few had been done on this subject, like comparing the insole of footwear with footprint. This study looks at comparing the outsole print of footwear with the barefoot print of persons to establish the relationship that exist if any between their patterns, for the purpose of person identification. Also barefoot morphology for person identification or authentication purpose was considered. Machine learning techniques were adopted for this task. Over 200 dataset were gathered of both footprints and footwear prints from volunteers, these were scanned into computer systems as images for processing, they were de-noised using Gaussian linear filter, segmented into clusters using k-means clustering techniques. Features were extracted using Principal Component Analysis PCA and Gabor filter, the feature vectors were compared using pattern machine techniques. The research employed Euclidean Distance, Artificial Neural Network ANN and Support Vector Machine SVM. The features considered in the research was the intensity of the print as that depicts the weight bearing areas of either the footwear print or the barefoot print. The extracted features were stored in a database. The barefoot prints were tested against footwear prints in the database and vice versa for any identification. The performance metrics employed to evaluate the efficacy of the system are the receiver operating characteristics ROC, true positive rate, false positive rate and accuracy. The result showed that the features extracted with PCA gave a better identification of footwear from barefoot prints and vice versa compared to those features extracted using Gabor filter. The performance metrics reported Euclidean distance and SVM as having 100% sensitivity and specificity on features that were extracted from PCA; this is called perfect classification, while ANN was noted to have random performance. This was basically due to size of data used. Meanwhile, on Gabor feature extract perfect classification was not recorded by any of the techniques. The time spent on the identification process was equally shorter for those samples whose features were extracted with PCA. Therefore it is concluded that extractions from PCA gave better feature vectors than those from Gabor filter. Considering the above, developing nations, or organisations with low income, can afford to set up a forensic system based on foot print recognition, for crime detection.