Skeleton-based human activity recognition using ConvLSTM and guided feature learning

• DOI: 10.1007/s00500-021-06238-7
• Publication Date: 2021-10-01T10:17:17Z
• Authors: Santosh Kumar Yadav; Kamlesh Tiwari; Hari Mohan Pandey; Shaik Ali Akbar
• Keywords: Human skeleton; Position (finance); Skeleton (computer programming); Feature (linguistics)

Abstract Human activity recognition aims to determine actions performed by a human in an image or video. Examples of include standing, running, sitting, sleeping, etc . These activities may involve intricate motion patterns and undesired events such as falling. This paper proposes novel deep convolutional long short-term memory (ConvLSTM) network for skeletal-based fall detection. The proposed ConvLSTM is sequential fusion neural networks (CNNs), (LSTM) networks, fully connected layers. acquisition system applies detection pose estimation pre-calculate skeleton coordinates from the image/video sequence. model uses raw along with their characteristic geometrical kinematic features construct guided features. are built upon using relative joint position values, differences between joints, spherical angles selected angular velocities. spatiotemporal-guided obtained trained multi-player CNN-LSTM combination. Classification head including layers subsequently applied. has been evaluated on KinectHAR dataset having 130,000 samples 81 attribute collected help Kinect (v2) sensor. Experimental results compared against performance isolated CNNs LSTM networks. Proposed have achieved accuracy 98.89% that better than LSTMs 93.89 92.75%, respectively. tested realtime found be independent pose, facing camera, individuals, clothing, code will made publicly available.