Behavior Inference based on Joint Node Motion under the Low Quality and Small-Scale Sample Size
Cheng Ji, Gaohao Zhou, Ge Liu, Qiang Mei · 2021
Behavior inference based on joint node motion is a relatively small field. In this field, the information density is barren and usually accompanied with low quality, because the data acquisition equipment will not use RGB images, but the use of structured light and TOF technology will introduce series problem, like overlap or drift of key nodes. This paper propose a lightweight model combining matrix convolution and continuous frame feature analysis to realize the motion behavior classification of key nodes. The size of our model is controlled within 70MB, and the calculation time on CPU and GPU is satisfactory. At the same time, our model shows linear attenuation in the process of gradually reducing the training sample size, which means that users can predict the model performance according to the sample distribution.