JUCS - Journal of Universal Computer Science 31(5): 494-518, doi: 10.3897/jucs.131543

A Novel GA-based Approach to Automatically Generate ConvLSTM Architectures for Human Activity Recognition

Sarah Khater^‡, Magda B. Fayek^‡, Mayada Hadhoud^§‡

‡ Cairo University, Cairo, Egypt§ School of Computational Sciences and Artificial Intelligence (CSAI), Cairo, Egypt

Corresponding author: Sarah Khater ( srashad@cu.edu.eg )

This is an open access article distributed under the terms of the Creative Commons Attribution License (CC BY-ND 4.0). This license allows reusers to copy and distribute the material in any medium or format in unadapted form only, and only so long as attribution is given to the creator. The license allows for commercial use.

Citation: Khater S, Fayek MB, Hadhoud M (2025) A Novel GA-based Approach to Automatically Generate ConvLSTM Architectures for Human Activity Recognition. JUCS - Journal of Universal Computer Science 31(5): 494-518. https://doi.org/10.3897/jucs.131543

Abstract

Human activity recognition (HAR) is a challenging computer vision problem that requires recognizing and categorizing human actions using spatiotemporal data. In recent years, ConvLSTM has shown distinctive advances in manipulating spatiotemporal data. ConvLSTM-based architectures, as any deep learning architecture, require deciding on many hyperparameters apart from trainable weights. State-of-the-art designs for general purpose datasets already exist, but specific purpose applications require architecture designs that perform well on application-dependent datasets. The design of such architectures requires either many trials and errors, which consume time and resources, or an experienced architect. Neural architecture search (NAS) meth-ods have been introduced to automate the design process and address the challenge of relying on expert knowledge when creating neural architectures. NAS enables rapid prototyping and experimentation, reducing the time spent on trial and error in manual design. One of the leading approaches in NAS is Genetic Algorithm (GA), which plays a significant role in optimizing neu-ral architectures. In this paper, a novel GA-based approach is proposed to automatically design ConvLSTM-based architectures from scratch for HAR applications. Our approach is based on multi-objective GA that maximizes recognition accuracy and minimizes the number of trainable parameters and overfitting measure. The experiments are held on KTH, Weizmann, and UCF Sports datasets. The best classification accuracies from the generated models are 97.92%, 96.77%, and 94.87% for KTH, Weizmann, and UCF Sports datasets, respectively. The experimental results show that the automatically generated models with the proposed approach outperform some of the state-of-the-art manually designed ConvLSTM-based architectures with percentages up to 9.92%, 5.77% and 23.64% for KTH, Weizmann, and UCF Sports, respectively. We also compared our approach with other NAS approaches. Our approach is found to outperform some of the introduced approaches with percentages approximately 2%, 11%, and 4% for KTH, Weizmann, and UCF Sports, respectively.

Keywords

ConvLSTM, HAR, KTH, Multi-objective fitness, NAS