Scope and Objectives:
Human Activity Recognition and Data Generation: Machine learning algorithms learn from past experiences. These past experiences can be in any form of data such as text data, visual data, structured data or sensor data. Choice of data type depends upon the problem and choice of data perception sensor. For example, in surveillance systems we usually deploy cameras to monitor the target area. Cameras perceive the environment in the form of image data of video data.
In case of RGB camera, video is the set of images and image is set of pixels that hold values of Red, Green and Blue colours. A computer algorithm sees image data as a matrix of integer values of between 0-255. Collection of labelled images or videos to perform a particular task is regarded as visual dataset. Such datasets are prepared by larger groups around the world and machine learning algorithms are trained on these datasets.
Human activities are crucial to surveillance of any territory. To develop automated systems that can perform surveillance by their own, human activities need to be recognized correctly. A lot of machine learning algorithms have been proposed to detect human activities (Human Activity Recognition and Data Generation) and many of them has yielded high fidelity. Among these techniques, deep learning is the technique that tries to analogue the human neural system. This technique has got prominence over traditional machine learning techniques due to increased accuracy and performance for various computer vision, speech processing and natural language processing tasks.
Deep learning networks:
Deep learning networks trained using large sets of labelled data. More robust and well prepared dataset, more accurate the performance of deep learning model. Therefore, the need to prepare more challenging dataset with different attributes is always in demand. Research groups prepare the dataset using recording equipment and human resources. This process of generating dataset is highly cumbersome. People are moving towards the automatic generation of datasets with minimal human intervention.
For human activity recognition task through deep learning, dataset collection is kind of monotonous task. Training a deep learning network requires millions of images or videos. Collecting these videos utilizes a lot of time, effort and human involvement. In order to address data generation problem, generative models proposed long time ago. Advent of deep learning, updated the traditional generative models to generative adversarial networks (GAN) . Like other deep learning models, GAN also achieved success over previous generative models. Therefore, generation of activity dataset using generative adversarial networks seems a fast and automated approach. Scope of data generation can retrained to surveillance systems.
Dataset generation encountered by minimal human input, resulting an activity video. Following are the objectives of activity data generation
- Generation of video sequence from single frame
- Generation of video sequence from single frame and label
- Generation of video sequence for two classes “walking” and “running” .
You may also know Feature extraction in Machine learning.