Human Activity Recognition with Noise-Injected Time-Distributed AlexNet.

This study investigates the integration of biologically inspired noise injection with a time-distributed adaptation of the AlexNet architecture to enhance the performance and robustness of human activity recognition (HAR) systems. It is a critical field in computer vision which involves identifying and interpreting human actions from video sequences and has applications in healthcare, security and smart environments. The proposed model is based on an adaptation of AlexNet, originally developed for static image classification and not inherently suited for modelling temporal sequences for video action classification tasks. While our time-distributed AlexNet efficiently captures spatial and temporal features and suitable for video classification. However, its performance can be limited by overfitting and poor generalisation to unseen scenarios, to address these challenges, Gaussian noise was introduced at the input level during training, inspired by neural mechanisms observed in biological sensory processing to handle variability and uncertainty. Experiments were conducted on the EduNet, UCF50 and UCF101 datasets. The EduNet dataset was specifically designed for educational environments and we evaluate the impact of noise injection on model accuracy, stability and overall performance. The proposed bio-inspired noise-injected time-distributed AlexNet achieved an overall accuracy of 91.40% and an F1 score of 92.77%, outperforming other state-of-the-art models. Hyperparameter tuning, particularly optimising the learning rate, further enhanced model stability, reflected in lower standard deviation values across multiple experimental runs. These findings demonstrate that the strategic combination of noise injection with time-distributed architectures improves generalisation and robustness in HAR, paving the way for resource-efficient and real-world-deployable deep learning systems.