Abstract:
Human Activity Recognition (HAR) is a vital area of Computer Vision. HAR focuses on various activities carried out by humans. Information relative to the human activities is collected by smart sensors and wearable devices. HAR is classified into two categories, e.g. (a) Vision-based, i.e. human activities are captured in form of image and video and (b) Sensor-based, i.e. human activity input can be taken from wearable devices and object tagging techniques. Human activity recognition is an extensive thrust area for Content-based video analysis, Human-machine interaction, animation, healthcare fields. The paper presents a comprehensive analysis of various deep learning-based approaches adopted to implement human activity recognition based on accuracy. It is observed that for the vision-based category the performance of the Depth Camera-based Recurrent Neural Network model is 99.55% accuracy with 12 activities for MSRC-12 datasets and for the sensor-based category, the performance of HAR by Wearable sensors using Deep Neural Network model is 99.93% accuracy with 03 activities for SHO datasets. It is also observed that for Opportunity dataset, InnoHAR: A DNN for complex HAR model gives good performance with 94.6% accuracy along with 18 activities, for PAMAP2 dataset, Multi-input CNN-GRU model gives good performance with 95.27% accuracy along with 12 activities, for WISDM dataset, ConvAE-LSTM model gives good performance with 98.67% accuracy along with 6 activities, and for UCI-HAR dataset, ConvAE-LSTM model gives good performance with 98.14% accuracy along with 6 activities.
