Gait Classification with Transfer Learning (ImageNet)

Walking gait is defined as the cyclical pattern in walking. Gait can be used as a basis for biometric classification.

The following supervised classifier is a DenseNet classifier which is trained on multiple still frames of labeled walking subjects. Transfer learning demonstrates a very high accuracy ~95% with just 2 epochs for large number of classes (10 in this case).

The dataset is obtained from CASIA (dataset C) which contains multiple backgrounds. The classifier performs better irrespective of the background, thanks to the transfer learning neural network which was already trained with large dataset of images.

You can check out my GitHub code here.