Human Action Recognition using Skeleton features

Many research works proposed in the field of action recognition based on the pose estimated in each frame. But most of these research works demanded a large amount of computing to store and process the collection of video frames. So we observed that Action recognition is computationally expensive. Our main objective was to find suitable features and classify human behavior based on action recognition. So to address these problems of storing and processing video input RGB frames, we came up with a new approach that is computationally not expensive. We convert the videos into skeleton-based Spatial-temporal Graphical (ST_GCN) frames that can detect human action more accurately than the direct RGB input. To derive solid development designs from these diagrams, a broad range and multi-scale setting total and perceptual demonstrating region unit significant parts of a solid feature extractor is required. We initially used tensor flow pose estimation CMU, Mobile-Net Pre Trained-Models by Open Pose and Tensor flow are used to train on the COCO dataset to extract key feature points. Then, NTU RGB D 60 3D skeleton data from Rapid-Rich Object Research Lab and MS_G3D model, which has more substantial generalization capability, have opted, and performance measured based on the model's accuracy on test data. Moreover, our models could profoundly classify a human's abnormal and normal behavioral patterns of a human based on their actions.