Human Actions Recognition System Based on Neural Networks

EasyChair Preprint 11799

Abstract

The recognition of human activities in videos is a relevant area of study due to its real-life applications, such as surveillance, security, healthcare, human-machine interaction, and monitoring. This research compares two recognition approaches, called simple and hybrid, using two specific data sets. The first set includes three classes: yoga, exercise, and dance; the second is a sample from the Kinetics-700 set, with five activities, four of which are violent. Both sets present low variability between classes and high variability within classes. To reduce computational costs, a pre-trained CNN model and simple techniques for reducing computational resources are used. The hybrid approach uses an additional model with three variants: GRU, LSTM, or BiLSTM. Even though all the models presented similar results, the simple approach, using a pre-trained architecture and a reconstructed top-head, proved to be the most effective, reaching an accuracy of 94\%, while the hybrid approach using LSTM layers obtained 90\%. The model demonstrated an adequate classification of violent activities, which could serve as a basis for developing a surveillance and security systems.

Keyphrases: BiLSTM, GRU, LSTM, Pretrained CNN

@booklet{EasyChair:11799,
  author    = {Juan Brito and Rigoberto Salomón},
  title     = {Human Actions Recognition System Based on Neural Networks},
  howpublished = {EasyChair Preprint 11799},
  year      = {EasyChair, 2024}}