Gait recognition with hybrid and ensemble deep learning models
Gait recognition is a biometric technology that identifies individuals based on their unique walking patterns. The distinctiveness of a person’s gait arises from various traits such as stride length, arm swing, limb movement, and others, making it challenging to conceal or imitate. Furthermore, gait...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2023
|
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Gait recognition is a biometric technology that identifies individuals based on their unique walking patterns. The distinctiveness of a person’s gait arises from various traits such as stride length, arm swing, limb movement, and others, making it challenging to conceal or imitate. Furthermore, gait recognition does not require any active participation from the subject, rendering it a convenient method of identification. However, gait recognition performance can be influenced by several covariates, including walking speed, viewing angle, clothing, and carrying conditions. These factors can alter the appearance of gait patterns, complicating accurate identification. To address these issues,a gait representation named “windowed GEI” alongside five deep learning-based models are presented: Gait-DenseNet201-MLP, Gait-VGG16-MLP, Gait-ViT-MLP, ensemble CNN-ViT with decision-level fusion, and ensemble CNNViT with feature-level fusion. The windowed GEI involves averaging the gait frames without regard to gait cycle. This technique helps alleviate potential biases associated with gait cycles, leading to improved generalisation of the gait representation. The Gait-DenseNet201-MLP utilises a pre-trained DenseNet201 model combined with a Multilayer Perceptron (MLP). All prior layers are connected to subsequent layers, enhancing the information flow among layers and promoting feature learning. The Gait-VGG16-MLP model employs the VGG16 architecture and reduces the number of parameters involved by using the smallest kernel size, thereby significantly decreasing execution time. The third model, Gait-ViT-MLP, is a Transformer-based approach that leverages the attention mechanism to focus on significant regions of an image, substantially improving performance. As each individual model possesses unique strengths, an ensemble model combining the predictions of these three models is proposed to further enhance performance. It is evident that the ensemble model significantly boosts performance compared to individual models. The final model fuses the representation of the models. The fusion yields a more comprehensive and informative feature representation, which effectively amplifies the performance. The proposed models are evaluated on several gait datasets, including the OU-ISIR treadmill gait dataset D, CASIA dataset B, and OU-ISIR large population dataset. |
|---|