Traditional security systems work to avoid crimes as much as possible. Real-time Surveillance gives an opportunity to prevent crimes before they can happen. Implementing security measures are also very time consuming and usually requires human interference. An autonomous security system will make security economically viable and it works quickly. Using facial, object and behavior recognition on the video feed provided by CCTV cameras, various criminal activities can be detected, and authorities will be assisted to take action. Covering a large number of CCTV’s distributed over wide space can generate lots of data and requires tremendous processing power to process this data. Hence, we will use Hadoop’s image processing interface to distribute the processing task over the cloud network, so communication between authorities of various areas is enhanced.
In the current time, at almost all locations, the security systems work in a rather passive way. CCTV cameras installed in these system record videos and feed them to a human supervisor. Such a security system is prone to human errors. Quick actions are not possible which are necessary for many conditions to prevent adversary. The entire security works locally and provides with limited cloud capabilities. Such a static system is outdated and itself is under security threat of being misused and hacked. Hence we propose a modern, dynamic system with capabilities to work in the cloud with powerful real-time surveillance and arguably cheaper than the existing system. Footages from multiple CCTV cameras will reach to a local station. These video feed will be provided to a preliminary object recognition algorithms and will undergo the process of culling in the local station.
After the initial process of object recognition, the video feed will be divided into a small unit, which comprises multiple images. This images will be mapped to the respective nodes for processing and their results will be reduced to get the final output.
The Authors in  proposed a scalable video processing system over the Hadoop network. The system uses FFmpeg for video coding and OpenCV for Image processing. They also demonstrate a face tracking system, which grouped multiple images of the same people together. Video feed captured is stored in the Hadoop Distributed file system. The system does not state proper security mechanisms and storing such huge amount of data in the HDFS will not be cost-efficient,
The system in  used Nvidia CUDA enabled Hadoop clusters to improve server performance by using the parallel processing capability of CUDA cores present in Nvidia GPU’s. They demonstrated an AdaBoost-based face detection algorithm in the Hadoop Network. Although equipping the clusters with Nvidia GPU’s might increase the cost of clusters, CUDA cores potentially provide massive improvements in Image processing jobs. Although we aim to implement the system into existing hardware to minimize the cost.
The Authors in  used the Hadoop Framework to process astronomical Images. They implemented a scalable image-processing pipeline over Hadoop, which provided for cloud computing of Astronomical Images. They used an existing C++ Library and JNI to use that library in Hadoop for Image processing. Although they achieve success, many optimizations were not made and Hadoop was not Integrated properly with the C++ Library.
A survey in  describes various security services provided in the Hadoop Framework. Security services, which are necessary for the framework such as Authentication, Access Control, and Integrity, are discussed including what Hadoop provides and what it does not. Hadoop has multiple security flaws which can be exploited to initialize a replay attack or view the files stored in the HDFS node. Hence as per the scholarly, a good Integrity check method and Authorization control method are necessary.
The object recognition stated in  provides an efficient way of recognizing a 3-Dimensional Object from a 2-Dimensional Image. In his stated methodology, certain features of the object remain constant regardless of the viewing angle. Extracting these features specifically will save a tremendous amount of resources as compared to the older object recognition systems that recreate the entire 3-D objects using Depth Analysis.
As depicted in , the original eigenfaces fail to accurately classify faces when the data is coming from different angles and light sources like in our problem. Hence, we use the concept of TensorFace. A vector space of different Images trained at multiple angles is applied to N-mode SVD to Multilinear Analysis to recognize faces.
Behaviour Recognition can be carried out as stated in . The features will be extracted from the video feed and applied to feature descriptors, model events, and Event/behaviour, models. The output will be mapped from feature space to behavior label space where a classifier will map it as normal or abnormal.
The system proposed in  states an economic, reliable, efficient and scalable surveillance system where data is stored using P2P concept. It avoids load on a single Data Centre and divides the load into multiple Peer Nodes. It also provides authentication as a module between the Peer Nodes and the directory nodes. The system doesn’t present any method to implement computer vision and integrity check.
Proposes an open source Hadoop Video processing Interface integrate C/C++ applications in the Hadoop Framework. It provides R/W interface for developers to store, retrieve and analyze video data from the HDFS. Using the available security in the Hadoop framework for video data can give poor performance and security was not mentioned in the HVPI.
TensorFlow, a machine Learning System, stated in , provides multiple tools to implement multiple training algorithms and optimizations for multiple devices on a large scale. It uses data flow graphs for computation states and operations that change those states. TensorFlow can work very well with Hadoop Framework to distribute the processing in the existing hardware.
To provide real-time recognition various pre-processing is done to improve Hadoop and neural network performance. The entire process can be divided into the following phases:-
Video Collection: – The video feed coming from the video capture device like CCTV will be converted to the Hip Image Bundle (HIB) object using various tools like Hib Import, info. After that, HIB will undergo preprocessing using a video coder like Culler class and FFmpeg. In this stage, various user-defined conditions like spatial resolution or the criteria for Image metadata can be applied. Filters like a greyscale filter provide improvements for various face detection algorithms. The images surviving the culling phase will undergo the preliminary object detection phase using object detection algorithms like tensor flow or provided by a library like OpenCV. Weapons, Cars, and Humans will be detected in this phase.
The collected Image will be mapped to MapReduce programming model using the HibInputFormat class. The individual Images are presented to Mapper as objects derived from the HipiImage abstract class associated with HipiImageHeader. The header will determine the what data to map to the respective data node in the network.
Mapping Phase: – Images, which are flagged as humans, will be mapped to the facial recognition and behavior recognition algorithms in the respective data nodes. Images recognized as cars will be mapped to object detection. Various algorithms for recognition in the mapping phase can be derived from OpenCV, which also inherently used Nvidia CUDA and OpenCL for increased performance in the recognition. OpenCV provides Java interface and can be directly used with Hadoop. Although a self-developed can be used and if required, will be written in C++ and JNI ( Java Native Interface) can use to integrate with Hadoop.
Reduce Phase: – Criminal faces will be detected during facial detection since the node with the highest confidence value will be declared as the winner. Stolen cars will also be detected in the similar fashion. Human behavior will classify and detect specific suspicious behavior.
Although the above paper only discusses specific applications, the entire architecture is scalable to be implemented in specific environments. The system can find applications in various companies offices, police department and various high-security facilities for real-time computer vision assistance. The system can also be implemented over the existing hardware either as a complement to the existing system or as a substitute to the existing system. Once enough test samples are collected, various optimizations can be used like different neural networks, more suited to the specific applications. Optimizations can also be made to the Java Native Interface (JNI) to improve further performance. Various pre-processing techniques in the video coder can be applied to improve the neural network performance.