Purchase Predictive Design Using Skeleton Model and Purchase Record

Abstract

The depth camera has enabled the skeleton and joints of the human body to use skeleton data in 3D space. Behavior recognition using skeleton data is mainly based on artificial neural networks such as RNN. This study classifies behaviors observed by the consumer into four categories using skeleton model learning for purchase predictive design. Skeleton model learning collects 25 skeleton joints using several Kinect v2s in unattended stores where four racks of items can be purchased. Torso, left arm, right arm, left leg, and right leg to five body joints are performed by BRNN, and as the layer becomes deeper, each part is then joined to the body. Finally, the 25 joints are grouped together and BRNN-LSTM is performed to solve the vanishing gradient problem (Jun et al. in J Korea Multimedia Soc 21:369–381, 2018, [1]). Supervised learning involves four behaviors used as input and the purchase record status as output. A GRU is employed to reduce computational complexity while maintaining the benefits of LSTM.