Age estimation from gait data using machine learning

Research in the field of gait biometrics has gained momentum in recent years due to the demand for automatic visual surveillance systems. In the past, human age estimation were based on face recognition, but the system performance could be low due to the low resolution and lack of sufficient detail of face images/videos collected by surveillance systems. Human age estimation from gait data has systematic advantages. In this study, local binary pattern (LBP), transient binary pattern (TBP) and deep transfer learning methods were implemented to estimate human age from gait data. The reason for using the LBP method is that this feature extraction method was successful in the face recognition problem. LBP was used to estimate age from gait energy images by implementing different diameters and number of points. Extracted features are separated by support vector machine, k nearest neighbor and random forest classifiers and validation data performance is obtained. TBP histograms are grouped into regional blocks by image tiling and represented as a histogram for each square block. Feature extraction was performed with two- and three-bit depth histogram graphics and 6×6, 8×8 and 10×10 square block separation methods, and performance measurement was made on the validation data with the same classifiers. Finally, age classification performance was obtained with the Resnet50 deep transfer learning method. The best-performing cases of the three methods are explained. The results of LBP, TBP and deep transfer learning were compared. The classification accuracy obtained with the TBP is higher than that of LBP. The classification accuracy obtained by the deep learning method is higher than the LBP and TBP methods. As a result, LBP and TBP are advantageous in terms of the computational cost and the deep learning transfer is advantageous in terms of accuracy performance.