An Optimized Deep Transfer Learning Framework for Robust Human Action Recognition in Challenging Environments

Abstract

The most important research area in computer vision is action recognition (AR) based on AI reasoning. During the past few years, big breakthroughs have been witnessed in this field, also the interest of research in this field is evolving like understanding of actions and scenes, studying the human joints, and human posture recognition. The precision of AR has been raised due to the growing of learning-based AI. Even though various innovations are being witnessed in AI technology, but still there exist quite a few challenges also in this field. This field requires large datasets and corresponding labels for their learning-based algorithms. Human action recognition (HAR) is rapidly gaining popularity in the field of pattern identification due to its unobtrusive features acquired from a distance and the fact that, unlike some other biometric applications like iris and fingerprint detection, it does not require the participation of individuals. Robotics, location estimation, sports analysis, pedestrian detection, human-computer interaction, video games, and video surveillance, are some examples of real-world applications where HAR is applied. Although different research approaches for action recognition have been presented, these methods still face obstacles like different viewpoints, environment conditions, and temporal variations, etc. The main purpose of this study is to deal with these addressed challenges. The suggested technique consists of a few core stages i.e. frames initial preprocessing, train using transfer learning concept, features extraction, fusing extracted features, features optimization, and classification. The first step is image preprocessing, in which the image dataset will be normalized by using color transformation and resizing. Pre-trained deep networks namely InceptionV3 and Resnet50 are used to extract deep features and then fused these feature vectors. Optimization based on IWOA is performed on fused vector and finally, classifiers are used for classification. Four publicly accessible datasets such as UCFT youtube, wizman, Ixmas, and UCF Sports are used to validate the HAR method and attained an accuracy of 100%, 99.9%, 99.1%, and 100% respectively.