Human Action Recognition Based on Self- Organizing Map in Surveillance Cameras

In recent years, Deep Learning, a Machine Learning algorithm, has been successfully used in many fields, and there have been reports of such a development in the field of behavior recognition. However, in general, high performance hardware is required to execute Deep Learning models in real-time. In this paper, we propose a surveillance camera system for behavior recognition using a self-organizing map (SOM) based on a time-series inference skeleton extracted from image data acquired by a monocular camera using a Machine Learning algorithm on an edge AI processor. The proposed surveillance camera system can be considered as a hybrid twin system. We performed skeleton extraction on an edge AI processor and evaluated experimentally in terms of processing performance and power consumption, and experimented with and evaluated an action recognition method using time-series inference skeleton information. An experimental evaluation of skeleton extraction on an edge AI processor shows the usefulness of the edge AI processor for use in a surveillance camera system. Experimental evaluation of action recognition by SOM showed that it is possible to recognize basic actions. We also found that each region on the U-matrix needs to be clarified in order to improve the percentage of correct answers.