Front and Back Views Gait Recognitions Using EfficientNets and EfficientNetV2 Models Based on Gait Energy Image

• Main Authors: Tengku Mohd Afendi, Zulcaffle, Fatih, Kurugollu, Kuryati, Kipli, Annie, Joseph, David Bong, Boon Liang
• Format: Article
• Language: English
• Published: University of Bahrain 2023
• Subjects: TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://ir.unimas.my/id/eprint/44802/1/Front%20and%20Back%20Views%20Gait%20-%20Copy.pdf; http://ir.unimas.my/id/eprint/44802/; https://journal.uob.edu.bh/handle/123456789/5178; http://dx.doi.org/10.12785/ijcds/140157

Front and back views gait recognitions are important, especially for narrow corridor applications. Hence, it is important to experiment with new algorithms on the front and back views gait recognitions. In this paper, we present the experiments on gait recognition using the pretrained EfficientNets and EfficientNetV2 models and Gait Energy Image. These models are chosen because they are among the best deep learning models in computer vision. The pretrained models were used in this experiment because it can produce faster and better accuracies compared to training the models from scratch. In addition to the pretrained models, we also propose ensemble models so that they can produce better accuracies. The result shows that the EfficientNetB7-Augm+ EfficientNetB6-Augm is the best overall accuracy (79.59%). However, combining the models slow down the inference speed. So, for recognition speed, EfficientNetB6 and EfficientNetB6-Augm are the best with 87.01ms speed per input image. The results produced are very good considering no cross-view algorithms applied to the Gait Energy Image. Future works will include the cross-view algorithms to further improve the accuracies of the proposed method.