Real-Time Age and Gender Estimation Combining Tracking System on Embedded Device

• Main Authors: Chao, Chih-Fan, 趙之璠
• Other Authors: Lee, Chen-Yi
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/vu782y

碩士 === 國立交通大學 === 國際半導體產業學院 === 107 === Age and gender estimation is a really popular study in recent years. Many applications can combine with it, especailly in the computer vision field. For example, we can apply it on survilliance system, human computer interaction or business advertisement. However, the most important thing among these application is that whether we can get the estimation result in time, which makes the issue become a challenging one. Nowadays, there are several research focusing on age and gender estimation. Most of them are through the deep learning model. After training, most of them can reach a good gender estimation result, however; the age estimation still needs to be strengthened. Though some of the state-of-the-art can reach a lower mean absolute error on age estimation, nevertheless; they use a too deep architecture so that their model is complicated and large. This makes it harder to apply on the real-time case. The thesis consider the age and gender as a regression problem. We reference the DEX[3] and SSR[11] to classify the age and gender into many groups, after multiplying the probability and index, we can get the final expected value for age and gender. Here, we modify the algorithm to make the model can learn dynamically in fewer parameters, and also can learn the low-level features to high-level features. The total model size is not larger than 1MB, which is highly possible to apply on embedded system. Furthermore, the thesis add the face tracking system. It raises the stability of age and gender estimation. This is the first one to add tracking system in this issue which makes the system will not vary the age and gender all the time. Also, we focus more on the real-time videos. Unlike the past research, which often focus on 2D images, we have experimented on the 3D real-time case. After experiment, the final results for real-time video is that we can reach about 2.0 MAE for age estimation and almost 100% accuracy on gender estimation. Finally, we apply our model on the TX1 and can reach 24 fps.