A Deep Learning Based Hybrid Approach for Human Physical Activity Recognition in Thermal Imaging

Abstract

One of the tough topics of research is the recognition of human action in the supervision video presently. To classify using traditional algorithms of image processing the human actions comprised of the same patterns sequence that is hard. Video analysis is a significant field of research that implies analytics to the camera. It screens the contents of the video and abstracts intelligent data from it. The majority of the tasks in this field are subjected to constructing the classifying techniques on complex properties that are handcrafted or modelling DL-based CNNs, which work on inputs that are raw and take out important data along with the video directly. In this paper, for the segmentation of human activities in video sequences, k-means clustering is used. To classify and detect various human activities like boxing, carrying, digging, robbing, etc the hybrid combination of ResNet50 and 3D-CNN is utilized. The (Resnet50) Pre-trained technique is utilized as a DL technique in this article. In order to capture the information of motion among the adjoining frames, the 3D-CNN extracts the features in the dimension of temporal together with the dimension of spatial. The performance measures are evaluated for various metrics such as precision, recall, f-score and accuracy.