-
Notifications
You must be signed in to change notification settings - Fork 8
/
dih original abstract
11 lines (8 loc) · 3.11 KB
/
dih original abstract
1
2
3
4
5
6
7
8
9
10
11
Human Activity Recognition using Deep Learning based approaches
Abstract
Automatic recognition and classification of human activity has been an active research field in the last years due to its potential applications in variety of domains. This task of activity recognition entails the automatic identification of human behaviour from images or video sequences. We propose a deep learning based approach to automatically recognize human activities. The objective is to learn the various spatiotemporal features which are to be extracted from a video stream similar to that from a CCTV camera.
We propose a two-step model for feature extraction and human action classification respectively. In the first part, we utilize Convolutional 3D (Conv3D) Networks to automatically learn visual features from sequential image data that is the video feed. Through previous research in this field, it has been found that Conv3D Networks outperform Conv2D Networks for video analysis due to their ability to model temporal features better.
In the second step we plan to use a Long Short-Term Memory (LSTM) network based model to use these learned features to classify the human activities. LSTM Networks are more suitable for sequential data as they are capable of learning long-term dependencies among the features by remembering information for long periods of time. By the combined use of Conv3D and LSTM, it would be possible for our model to retain both spatial and temporal features from the video stream input.
In this field, currently many labelled datasets are available having a variety of classes of routine human activities like walking, running, jumping, etc. We plan to select a dataset from the available ones on the basis of features which are more suitable for sequential analysis. We are going to use a subset of the classes available for training our pilot model. Afterwards, we will apply the similar algorithm to the complete dataset.
Deliverables:
There are great societal benefits of Human Activity Recognition System, especially in real-life human centric applications such as eldercare and healthcare. There is also wide applicability of Human Activity Recognition in the task of automated video surveillance. Using CCTV cameras for monitoring and surveillance is common for security purposes. However, most of the present-day video surveillance systems have a common flaw: they need human operator assistance to constantly monitor the large volume of information generated by the cameras. Moreover, number of surveillance cameras and the region of activity are both limited by the availability of security personnel. In today’s AI driven world, there is still consistent lack of an automated robust video surveillance system. The method we are proposing can be utilised to classify the activities of persons in the video frame. This will enhance the performance of the surveillance system by real time activity recognition. It is both diligent and cost cutting method. This method may be further developed in future to create a description of the events happening within their area and generate warnings if they detect a suspicious person or unusual activity.