real world anomaly detection in surveillance videos

Our project lies within the domain of computer vision and deep learning, and its interactive applications have enabled the world to progress scientifically with the aim of bringing innovation to life. Anomaly detection is one of the application in this

Project Title

real world anomaly detection in surveillance videos

Project Area of Specialization

Artificial Intelligence

Project Summary

Our project lies within the domain of computer vision and deep learning, and its interactive applications have enabled the world to progress scientifically with the aim of bringing innovation to life. Anomaly detection is one of the application in this interdisciplinary scientific field  that would observe the high level interpretation from the digital videos. These are used to detect the anomalous behavior to automate the tasks that human can do. Our goal is to extract the relevant information from the real-world videos in order to produce numerical or symbolic form to classify them in the respective output. In this project, we have 13 different classification labels based on which the model will predict the presence or absence of anomaly.

Project Objectives

The goal of real-world anomaly detection is to identify anomalies by exploiting both normal and anomalous videos. We have used with large-scale, first-of-its-kind dataset of 128 hours of videos. It consists of 1900 long and untrimmed real-world surveillance videos, with 13 realistic anomalies such as fighting, road accident, burglary, robbery, etc. as well as normal activities. The goal of a practical anomaly detection system is to timely signal an activity that deviates normal patterns and identify the time window of the occurring anomaly. As some of the work has already been done, our core purpose of this project is to bring these computer vision models in mobile version using different techniques and framework including mobilenet, opencv, tensorflow and keras. We aim to build an android mobile app to directly upload the real time Surveillance video and detect its anomaly rate with respect to the 13 classification as trained in the deep learning phase using conventional neural networks. As our top priority is to bring this innovative application to mobile, our main concerns are in terms of memory consumed and hence accuracy of class prediction can be a tradeoff here.

Project Implementation Method

Some previous work has been done on this project and we aim to expand it even further with some added functionalities.

It is a project which utilizes convolutional neural networks as its main framework. Anomaly is basically referred as a deviation from normal behavior. Previous work includes detection of anomaly based on weakly labeled training data. Violence and aggression in videos was detected by exploiting motions and limb orientation of people. It is a web based application which does not support mobile environment due to its excessive memory consumption and heavy weighed models.

The machine learning model used for feature extraction is MobileNet V2. It is a pre-trained model on the ImageNet dataset [1]. For Feature extraction, we used the common technique of using the last layer i.e. ‘the bottleneck layer’. All the classification layers are frozen from the top and this becomes the ideal condition for feature extraction. Unfortunately, the features extracted from this layer did not prove to be very beneficial for us. In this project, we have chosen two major tools for development: visual studio code and anaconda. However, we have been shifting among different development tools to check the compatibility with our goal. For major backend work, we have opted for an anaconda navigator to use the Jupyter notebook environment.
Through the mobile app, a video can be uploaded which is then processed to extract the features in single dimension, those features are then transferred to the server to predict the anomaly rate and the results are then showed on the screen to the user.
 

Benefits of the Project

The users of this particular system could have been security authorities only but our system is designed for all types of people. They can upload anomalous or non-anomalous videos and detect the presence or absence of anomalies within seconds.

Technical Details of Final Deliverable

The most unique feature of our project that makes it distinctive is the mobile friendly feature. We would be deploying a mobile app to classify the videos using light weighted models.

The second unique feature is localizing anomaly by temporal means.

In this project, we would be using 13 different classification labels to detect the presence of anomaly (binary classification) in the input video by segmenting it into frames.

Final Deliverable of the Project

Software System

Core Industry

IT

Other Industries

Core Technology

Artificial Intelligence(AI)

Other Technologies

Sustainable Development Goals

Industry, Innovation and Infrastructure

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com