Multi-Label Human Activity Recognition on Image Using Deep Learning

Abstract

This paper describes the model of convolutional neural network which is designed for multi-label human activity recognition. The possibilities of using activity recognition systems in the daily life of a person are considered. As part of this work, the study is conducted for the method of recognizing human activity on an image that can be obtained from a surveillance camera. To obtain more accurate recognition results, the network model used technology of transfer learning. Several pre-trained convolutional networks are considered using two types of transfer learning in order to find the best solution. The deep learning networks for solving the problem are implemented in Python using deep learning libraries. Considered models are trained to recognize binary multi-label human activity. Training and testing are performed on images collected by the author. The article also provides the obtained training and testing results of different models of convolutional neural networks. The data obtained are tabulated and also presented in graphical form.

Recognizing Human Daily Activity Using Social Media Sensors and Deep Learning

“…Human physical activity recognition is a promising research field due to its applications in healthcare, athletics, lifestyle monitoring, and computer-human interaction. Usually, it can be implemented from three categories of data, such as video, wearable motion sensor data, and multi-sensors data [11,12,13]. Researchers use traditional classification methods such as SVM (Support Vector Machine) and TF-IDF (Term Frequency–Inverse Document Frequency) models to identify human physical activities and develop a syntactic approach that treats activities as sequential text and uses a grammar syntax such as stochastic context-free grammar to identify physical activity categories [7,8,9,14,15,16].…”

“…Researchers use traditional classification methods such as SVM (Support Vector Machine) and TF-IDF (Term Frequency–Inverse Document Frequency) models to identify human physical activities and develop a syntactic approach that treats activities as sequential text and uses a grammar syntax such as stochastic context-free grammar to identify physical activity categories [7,8,9,14,15,16]. Recently, deep learning has been adopted by combining spatiotemporal dynamic texture descriptors [11].…”