Human Activity Recognition with Noise-Injected Time-Distributed AlexNet

Area of Science:

• Computer Vision
• Deep Learning
• Biologically Inspired Computing

Background:

• Human Activity Recognition (HAR) is crucial for applications in healthcare, security, and smart environments.
• Traditional AlexNet architectures are designed for static images and require adaptation for temporal video data.
• Overfitting and poor generalization limit the performance of time-distributed AlexNet models in HAR.

Purpose of the Study:

• To enhance the performance and robustness of HAR systems.
• To integrate biologically inspired noise injection with a time-distributed AlexNet architecture.
• To evaluate the impact of noise injection on model accuracy, stability, and overall performance.

Main Methods:

• Adapted AlexNet architecture using a time-distributed approach for video classification.
• Incorporated Gaussian noise injection at the input level during training, inspired by biological sensory processing.
• Conducted experiments on EduNet, UCF50, and UCF101 datasets to assess model performance.

Main Results:

• The bio-inspired noise-injected time-distributed AlexNet achieved 91.40% accuracy and 92.77% F1 score.
• The proposed model outperformed existing state-of-the-art models on the tested datasets.
• Hyperparameter tuning, especially learning rate optimization, improved model stability and reduced result variance.

Conclusions:

• The strategic combination of noise injection and time-distributed architectures significantly improves HAR generalisation and robustness.
• This approach offers a pathway towards resource-efficient and deployable deep learning systems for real-world HAR.
• Biologically inspired noise injection is an effective technique for enhancing deep learning model performance in dynamic environments.