HRNetO: Human Action Recognition Using Unified Deep Features Optimization Framework

Abstract

Human action recognition (HAR) attempts to understand a subject’s behavior and assign a label to each action performed. It is more appealing because it has a wide range of applications in computer vision, such as video surveillance and smart cities. Many attempts have been made in the literature to develop an effective and robust framework for HAR. Still, the process remains difficult and may result in reduced accuracy due to several challenges, such as similarity among actions, extraction of essential features, and reduction of irrelevant features. In this work, we proposed an end-to-end framework using deep learning and an improved tree seed optimization algorithm for accurate HAR. The proposed design consists of a few significant steps. In the first step, frame preprocessing is performed. In the second step, two pre-trained deep learning models are fine-tuned and trained through deep transfer learning using preprocessed video frames. In the next step, deep learning features of both fine-tuned models are fused using a new Parallel Standard Deviation Padding Max Value approach. The fused features are further optimized using an improved tree seed algorithm, and select the best features are finally classified by using the machine learning classifiers. The experiment was carried out on five publicly available datasets, including UT-Interaction, Weizmann, KTH, Hollywood, and IXAMS, and achieved higher accuracy than previous techniques.