Hybrid Deep Learning Models for Gait Recognition: A Comparative Analysis of CNN, CNN-LSTM, and HOA Techniques

Abstract

Gait recognition is a critical biometric technique with applications in surveillance, healthcare, and security. This study proposes a hybrid deep learning framework combining Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and the Hippopotamus Optimization Algorithm (HOA) for robust gait recognition. By leveraging spatial feature extraction, temporal dynamics, and metaheuristic hyperparameter optimization, the proposed HOA-CNN-LSTM model achieves superior performance. Experimental results on the TUM-GAID dataset show that the hybrid model outperforms standalone CNN and CNN-LSTM approaches in accuracy, processing time, and error rates. The findings suggest that HOA-optimized architectures provide scalable and efficient solutions for gait recognition tasks in real-world settings.

Introduction

Gait recognition, the analysis of individuals' walking patterns, is gaining prominence as a non-intrusive biometric technique suitable for applications like surveillance, healthcare, and human-computer interaction. Unlike traditional biometrics (e.g., fingerprint, iris), gait can be captured remotely and passively, making it ideal for use in uncontrolled or non-cooperative environments.

Challenges and Deep Learning Approaches
Recognition accuracy is hindered by variations in clothing, footwear, speed, and viewing angles. Initially, gait recognition relied on handcrafted features, but these were inconsistent. The adoption of deep learning, particularly Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) like LSTM, improved performance by learning spatial and temporal gait patterns. Hybrid CNN-LSTM models have shown superior results but are sensitive to hyperparameter settings.

Metaheuristic Optimization with HOA
Traditional optimizers like SGD and Adam face issues such as slow convergence. Metaheuristic algorithms—like Genetic Algorithms, PSO, and the newer Hippopotamus Optimization Algorithm (HOA)—offer better exploration of the solution space. HOA, inspired by hippopotamus foraging behavior, helps fine-tune complex model parameters.

Proposed Methodology
The study proposes a hybrid CNN-LSTM model optimized using HOA. The process includes:

• Data Preprocessing: Using the TUM-GAID dataset, gait silhouettes are extracted and enhanced via techniques like augmentation and Dynamic Time Warping (DTW).

• Feature Extraction: A ResNet-50 CNN captures spatial gait features.

• Temporal Modeling: A Bidirectional LSTM captures sequence dynamics.

• HOA Optimization: HOA fine-tunes key hyperparameters (learning rate, batch size, dropout, etc.) using a fitness function based on classification performance.

Experimental Results
Evaluated on TUM-GAID with 10-fold cross-validation, the HOA-optimized CNN-LSTM model outperformed baseline models (CNN-only and CNN-LSTM) in terms of:

• Accuracy

• Genuine Acceptance Rate (GAR)

• False Acceptance Rate (FAR)

• False Rejection Rate (FRR)

• Processing Time

The proposed system is robust, generalizable, and computationally efficient, making it well-suited for real-world biometric authentication applications.

Conclusion

This study introduced an advanced hybrid gait recognition framework that combines ResNet-based CNN for spatial feature extraction, Bi-LSTM for capturing temporal dynamics, and the Hippopotamus Optimization Algorithm for hyperparameter tuning. The HOA-CNN-LSTM model outperformed standalone CNN and CNN-LSTM models in accuracy, efficiency, and error rates across diverse gait scenarios. The results validate the effectiveness of HOA in enhancing deep learning models\' robustness and scalability. Future work will explore integrating attention mechanisms, deploying models on edge devices, and extending to multimodal biometric systems for enhanced accuracy and security.

References

[1] K. Jain, A. Ross, and S. Prabhakar, “An introduction to biometric recognition,” IEEE Trans. Circuits Syst. Video Technol., vol. 14, no. 1, pp. 4–20, Jan. 2004. [2] W. Kusakunniran, “Gait recognition under various viewing angles based on correlated motion regression,” IEEE Trans. Circuits Syst. Video Technol., vol. 22, no. 6, pp. 966–980, Jun. 2012. [3] J. Han and B. Bhanu, “Individual recognition using gait energy image,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 28, no. 2, pp. 316–322, Feb. 2006. [4] D. Tao, Y. Wang, X. Li, and X. Wu, “General tensor discriminant analysis and Gabor features for gait recognition,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 29, no. 10, pp. 1700–1715, Oct. 2007. [5] Y. Li, W. Zhang, and X. Wang, “DeepGait: A Learning-Based Gait Recognition Framework Using CNN,” in Proc. Int. Conf. Image Process. (ICIP), 2018, pp. 1398–1402. [6] Y. Zhang, L. Wang, and Z. Wang, “Combining CNN and RNN for Gait Recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. Workshops, 2019, pp. 70–77. [7] C. Wu, A. Wang, and F. Gao, “Hybrid CNN-LSTM Model for Gait Recognition,” in Proc. Int. Joint Conf. Neural Netw. (IJCNN), 2020, pp. 1–8. [8] M. Alzubaidi et al., “Review of deep learning: Concepts, CNN architectures, challenges, applications, future directions,” J. Big Data, vol. 8, no. 1, pp. 1–74, 2021. [9] D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2015. [10] X. S. Yang, “Nature-Inspired Optimization Algorithms,” Elsevier, 2014. [11] M. A. El-Dosuky, “Hippopotamus optimization algorithm,” Int. J. Intell. Comput. Inf. Sci., vol. 18, no. 2, pp. 1–10, 2018. [12] S. Bansal, R. S. Anand, and S. Tiwari, “Metaheuristic optimization in deep learning: A survey,” Artif. Intell. Rev., vol. 55, pp. 2181–2225, 2022. [13] A. H. Hussain et al., “Optimization of CNN using HOA for medical image analysis,” Appl. Soft Comput., vol. 105, p. 107258, 2021. [14] M. Hofmann, S. Bachmann, B. Pflugfelder, and G. Rigoll, “TUM GAID: A gait dataset for cross-view gait recognition,” in Proc. Int. Conf. Comput. Vis. Theory Appl., 2014, pp. 995–1001. [15] S. Kumar and P. Ramesh, “Lightweight CNN for Real-Time Gait Recognition on Edge Devices,” IEEE Trans. Biom. Behav. Identity Sci., vol. 4, no. 1, pp. 30–41, Jan. 2022. [16] J. Li, T. Sun, and Z. Yu, “Multi-Scale CNN for Gait Recognition Under Clothing Variations,” Neural Comput. Appl., vol. 35, no. 7, pp. 12489–12499, Apr. 2023. [17] W. Wang, K. Jin, and X. Ren, “Residual CNN for Hierarchical Gait Feature Learning,” Comput. Vis. Image Underst., vol. 234, p. 103489, May 2023. [18] X. Zhang, L. Chen, and Y. He, “Temporal Gait Analysis Using LSTM Networks,” Pattern Recognit. Lett., vol. 165, pp. 85–92, 2022. [19] S. Lee and H. Choi, “Gait Cycle Segmentation Using LSTM Temporal Encoder,” Sensors, vol. 22, no. 14, pp. 4890–4905, Jul. 2022. [20] M. Amin, H. Alwan, and L. Naji, “Bidirectional LSTM for Robust Gait Recognition in Smart Environments,” J. Ambient Intell. Humaniz. Comput., vol. 14, pp. 227–241, 2023. [21] X. Zhang, Y. He, and L. Cheng, “Hybrid CNN-LSTM Model for Cross-View Gait Recognition,” IEEE Access, vol. 10, pp. 74012–74023, 2022. [22] F. Liu, M. Chen, and Y. Wang, “Attention-Augmented CNN-LSTM for Gait Recognition,” IEEE Access, vol. 11, pp. 45678–45691, 2023. [23] E. Okonkwo, F. Balogun, and A. Yusuf, “Residual Dropout-Based Hybrid CNN-LSTM for Gait Analysis,” Appl. Intell., vol. 54, pp. 9982–9995, 2024. [24] Y. Chen, H. Zhao, and T. Ma, “GA-Optimized CNN for Biometric Recognition,” Expert Syst. Appl., vol. 208, p. 118215, Jan. 2023. [25] L. Das and K. Roy, “Gait Recognition Using PSO-Tuned Deep Networks,” Appl. Soft Comput., vol. 132, p. 109889, Oct. 2023. [26] M. Rahman, S. Amin, and H. Chowdhury, “Hippopotamus Optimization Algorithm for CNN Parameter Tuning in ECG Classification,” IEEE Trans. Instrum. Meas., vol. 74, pp. 1–10, 2023. [27] A. Akintola, S. Yusuf, and T. Idris, “HOA-Tuned Attention Networks for Long-Sequence Gait Recognition,” IEEE Trans. Neural Netw. Learn. Syst., early access, 2025. [28] R. Silva, L. Gomez, and C. Martins, “Benchmarking Deep Gait Models: Cross-View Analysis on TUM-GAID,” IEEE Trans. Image Process., vol. 33, pp. 115–127, Feb. 2024. [29] J. Huang et al., \"Time warping and data augmentation for gait cycle alignment,\" IEEE Trans. Cybern., vol. 54, no. 2, pp. 890–902, 2024. [30] H. He and J. Li, \"ResNet variants for gait analysis,\" Neurocomputing, vol. 520, pp. 80–89, 2022. [31] K. Nakamura et al., \"Temporal context modeling with Bi-LSTM for action and gait analysis,\" Sensors, vol. 23, no. 5, p. 2203, 2023. [32] F. Alazab and M. Kaur, \"Cross-entropy loss optimization in deep gait recognition systems,\" Applied Sciences, vol. 13, no. 6, p. 3451, 2023. [33] E. Guillen Pinto, L. Ramirez Lopez, and C. Ramos Linares, \"GaHu-Video: Parametrization system for human gait recognition,\" Mendeley Data, 30-May-2020. [Online]. Available: https://data.mendeley.com/datasets/gprg4s73v4/1. [Accessed: 14-Nov-2024]. DOI: 10.17632/gprg4s73v4.1.

Copyright

Copyright © 2025 Atansuyi N., Ajaegbu C., Adekola O., Akande O.. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.