Predictive Analytics for Recognizing Human Activities Using Residual Network and Fine-Tuning

Abstract

Human Action Recognition (HAR) is a rapidly growing study area in computer vision due to its wide applicability. Because of their varied appearance and the broad range of stances that they can assume, detecting individuals in images is a difficult undertaking. Due to its superior performance over existing machine learning methods and high universality over raw inputs, deep learning is now widely used in a range of study fields. For many visual recognition tasks, the depth of representations is critical. For better model robustness and performance, more complex features can represent using deep neural networks but the training of these model are hard due to vanishing gradients problem. The use of skip connections in residual networks (ResNet) helps to address this problem and easy to learn identity function by residual block. So, ResNet overcomes the performance degradation issue with deep networks. This paper proposes an intelligent human action recognition system using residual learning-based framework “ResNet-50” with transfer learning which can automatically recognize daily human activities. The proposed work presents extensive empirical evidence demonstrating that residual networks are simpler to optimize and can gain accuracy from significantly higher depth. The experiments are performed using the UTKinect Action-3D public dataset of human daily activities. According to the experimental results, the proposed system outperforms other state-of-the-art methods and recorded high recognition accuracy of 98.25% with a 0.11 loss score in 200 epochs.