HUMAN ACTIVITY RECOGNITION - A COMPARATIVE STUDY USING TRADITIONAL AND NEXTGEN TRANSFER LEARNING PRE-TRAINED MODELS

Abstract

Human Activity Recognition (HAR) is very vital in appreciating human de meanor and finds applications in healthcare, sports analytics, and surveillance systems. Increasingly the data driven insights are being utilized HAR plays a key role in identifying patterns, trends and anomalies associated with human ac tivities. The use of machine learning and deep learning techniques has helped to improve significantly HAR methodologies leading to higher accuracy and ef ficiency. This study provides an extensive insight into traditional and advanced transfer learning pre-trained models for exploring intricacies of HAR. Each of the different model architectures in the research was assessed in depth by this evaluation, with its own strengths and capabilities. VGG16 or VGG19 and EfficientNetV2S, Xception are examples of old pre-trained models which would be compared with ConvNeXt frameworks such as ConvNeXtSmall, ConvNeXtBase, ConvNeXtLarge, and ConvNeXtXLarge. The main objective of this study is to comparatively analyze the efficiency of these models using human activity recognition. The benchmarking used a well-curated dataset that involved 12000 images annotated and classified into fifteen activities The Kaggle dataset sourced is useful for evaluating any performance changes made to different pretrained models. In order to avoid partiality and control external factors; like biases each model had exactly the same number of layers as others. The experiments are carried out in the Google Colab environment, which is cloud-based and therefore allows for extensive experimentation and analysis.