Forensic gender and stature identification from footprint images using machine learning

The analysis of footprints to infer human characteristics and biometric information is a valuable tool in forensic investigation. Traditional methods rely primarily on physical measurements and observational analysis, which requires significant time, effort and specialized expert judgment. This study proposes a novel, automated, end-to-end approach to gender classification and stature estimation from footprints, using image analysis and traditional machine learning methods. Specifically, this we employ a image pre-processing techniques for Region of Interest extraction to segment foot prints in images and identify toe and heel exterior points through convexity and defectiveness points. The study utilized a dataset of 396 footprints from 33 participants (18 males and 15 females, aged 18-48 years, height range 148-182 cm). Hyper-parameter tuning via grid search optimization is employed and traditional Machine Learning (ML) models, including Logistic Regression (LR), Gaussian Naive Bayes (GNB), K-Nearest Neighbor (KNN), Decision Tree Classifier (DTC), and Support Vector Machine (SVM), are benchmarked for the task of both inferring gender, and stature. We specifically focus on traditional ML methods due to their relatively modest training data requirements, with the aim of establishing their feasibility for such forensic analysis. KNN demonstrated better accuracy overall for gender classification achieving 0.91 accuracy, while Extreme Gradient Boosting (XGBoost) outperformed other methods for stature estimation with MAE of 4.10 cm and RMSE of 5.42 cm, however varying strengths and weaknesses of each classifier for gender classification and stature estimation were observed. Our results suggest that the strongest performing traditional ML methods offer a feasible solution for such analysis, however expanding the training dataset to incorporate more footprint examples of more varying quality and depicting a greater diversity of population is likely necessary to fully realise a workable end-to-end solution. Such datasets may also open the door to more advanced deep learning methods.