MobileNet for human activity recognition in smart surveillance using transfer learning

Human activity recognition is a significant and trending research area in computer vision due to its wide range of applications in healthcare, wellness, smart home systems, smart surveillance, and more. Recently, with the rise and successful deployment of deep learning techniques for image classification and object recognition, research has shifted from traditional handcrafted methods to deep learning approaches. Video content has gained considerable popularity on the internet, especially on social media platforms like YouTube, spurring growing interest in video understanding within the research community. This paper analyzes the performance of the MobileNet model for human activity recognition in smart surveillance using transfer learning, a promising approach for leveraging pretrained models. The MobileNet model, pretrained on ImageNet with 101 classes, was modified with additional layers to enhance its performance. The model was further tested using the UCF101 benchmark dataset, which includes RGB images across 101 categories of activities, such as playing guitar, jumping, and typing. The primary goal is to assess the MobileNet model’s performance in improving human action recognition through transfer learning. Key metrics, including accuracy, precision, F1-score, Area under the Curve (AUC), and Root Mean Squared Error, are used to evaluate its effectiveness. The MobileNet model achieved an accuracy of 98.47% on the UCF101 dataset. The experimental results demonstrate that the proposed technique is both robust and efficient on the UCF101 dataset.