Multiview Human Gait Recognition using a Hybrid CNN Approach

Recognizing a person's gait is a challenging task because there are so many factors to consider, such as obstructions due to clothing and bags. As a solution to this problem, a system is proposed for identifying gaits that is based on deep learning and random forests. For feature extraction from video frames, the system employs two popular pretrained models, MobileNetV1 and VGG19. The dimensionality of features is minimized using PCA and mean-based feature fusion is used to combine the reduced features. Six angles were selected from the dataset, and Random Forest was used for classification. The proposed method is put to the test on the CASIA-B dataset, and the results obtained show a mean accuracy of 93.1% for six angles. Experimental findings prove that deep learning and random forests are useful tools for gait recognition.