Daily activity recognition from egocentric images

Abstract

By analysing people way of life we can create methods of prevention and intervention for human behaviour derived diseases. Lifelogging allow us to obtain information, through image capture, of the daily life and the environment in which we move. However, we need to classify those images in order to obtain information, and then to analyse that data to detect behaviour patterns that may be affecting people. But how can we classify thousand of images in a quick way? Automatic classification algorithms, such as convolutional neural networks based techniques and deep learning have shown promising results when classifying images. This work introduces the challenge, first, of realizing a tool for a manual classification, with a website showing images that allow us to easily classify images using batches. Such a tool allows us to create a data set of nearly 20.000 images to, in the second part of the project, realize a fine-tuning over a convolutional neural network trained with ImageNet. After that fine-tuning, the convolutional neural network is combined to obtain the features from the images in order to train a Random Decision Forest classifier. Finally the results are studied. The global accuracy for the CNN system based is that of 58%. A better solution is obtained when combining CNN’s and RDF’s reaching up to 85% of global accuracy. Thus concluding that the classification system based on training a RDF with the data provided by the CNN, image features and probabilities, is the system offering better results.