Multi-task learning: A solution of small sample size problem in floor-based gait recognition

Abstract

Recognizing users via gait information is a potential approach for user privacy protection in the Internet of Things. Floor-based gait recognition aims to accomplish identification using plantar pressure during walking, and it has advantages of user-friendliness, robustness, and anti-counterfeiting. Convolutional neural network (CNN) is a promising gait recognition method for its advantage of automatic feature learning. However, the application of CNN in floor-based gait recognition is restricted by the small sample size problem, which is due to the difficulty of collecting large-scale plantar pressure data sets. Our aim is to provide an improved CNN model to achieve competitive gait recognition performance with a pretty small group of plantar pressure images. Inspired by multi-task learning, a multi-task CNN that contains gender recognition and BMI level recognition auxiliary tasks is constructed for gait recognition. To verify the proposed method, a dataset that contains 2880 plantar pressure images from 10 subjects is collected through the designed pressure sensing floor. Our method achieves an accuracy of 88.21% in 10-subject recognition and satisfactory results in the recognition of different categories. By comparison, the speed of identifying a sample by our model is about 37.74 ms and the model size is 5.50 MB, which helps with hardware implementation. The proposed method is proved to have advantages in recognition accuracy and efficiency.