Affiliation:
1. Haldia Institute of Technology, India
Abstract
Human activity recognition (HAR) is a study area concerned with the voluntary detection of routine human activity using sensor-based time-series data. However, sensor-based HAR systems have higher misclassification rates for complex actions like running, jumping, wrestling, and swinging because of sensor reading errors. The HAR system's overall performance is decreased by these sensor errors, which produce the worst classification outcomes imaginable. For complex tasks, better accuracy can be attained using vision-based HAR systems. This technique builds a deep convolutional neural network, and then uses it to extract features from the input sequence in order to gather data. Then, the temporal connections between the images will be ascertained using LSTM. This model's accuracy was suggestively higher than that of other cutting-edge deep neural network models after it was effectively validated on the UCF50 dataset. The implementation of the models to maximize their efficacy has been covered in this chapter.