Human Age Estimation from Multi-angle Gait Silhouettes with Convolutional Neural Networks

Kotcharat Kitchat, Piya Limcharoen, Nirattaya Khamsemanan, Cholwich Nattee

Abstract

Knowing the age of an unknown person is useful information in many fields such as forensicsand commercial fields. Authorities can use a correct age estimation technique to narrow down a searchfor a suspect. Conventionally, facial features are used to estimate the age of a person. However, obtainingthe facial features of a subject needs a high-resolution camera with a close-up image, which is hard todo in a real-world environment. Unlike facial features, gaits, a pattern of locomotion, can be observedfrom a far and unobtrusively. In this work, we propose a new age estimation technique to predict theage of a person from gaits using multi-angle GEIs and a modified Convolution Neural Network (CNN)with Gaussian Mixture (GM). The proposed system consists of two main parts: the moving directionclassification model and the age estimation model. The Gaussian Mixture is used as the loss in our ageregression model to estimate the age. Our age estimation model contains a modified CNN and followswith three sub-networks that calculate three parameters of the Gaussian Mixture. The proposed methodsperform well in multi-viewpoints. The proposed model achieves a Mean Absolute Error (MAE) of 4.08years in a multi-viewpoint dataset (SIIT-CN-B), which outperforms the existing techniques.