Deep learning based action recognition

Abstract

In recent times, surveillance is becoming more and more prevalent, with it being used to aid law enforcement, ensure home safety and allow caregivers to track the movement of the elderlies at home. The detection and recognition of human actions, through surveillance and recording devices, is informative — it allows the user to monitor the actions of the target subject at any point of time. One of the more successful ways of recognizing human action is through the use of skeleton data. This method provides an excellent representation for describing human activities. Recently, the Graph Convolutional Network (GCN) has become a popular topic, and graph-based skeleton action recognition significantly improved in performance, because GCN can learn the neighbourhood information and body joints’ interaction in a directed way. Thus, GCN has exceptional potential in skeleton action recognition. This report consists of three sections: Firstly, several ablation studies are conducted on the baseline model – Two-Stream Adaptive Graph Convolutional Network (2s-AGCN) to better understand the performance of the baseline model, Secondly, several experiments are performed where we explore adding some modification to the baseline model to improve on its performance when benchmarking the model against datasets that contain skeleton data such as – NTU RGB+D 60 dataset, Lastly, experiments on several models with new ideas incorporated are conducted to test their performance against the baseline model.