Privacy Preserving in Video Surveillance Systems Using Regression Residual Convolutional Neural Network in Private and Public Places

1. Introduction And Problematic

The video surveillance allows a fast localization of threats, thus protecting people and enterprise. This technology ensures the surveillance of perimeter and the outer shell against vandalism, burglary and espionage. It is constantly evolving, a wide product range that is now available to all. We can adopt cameras that allow just to deter, others that can film at night using infrared mode, it is also possible to buy a kit with a number of cameras to equip your home. In general, the use of video surveillance systems allowed a reduction in crime either in public and private places. Everyone was winner of using this type of device because the drop-in crime is positive for everyone.

Merely, In the context of video surveillance in public places, this practice is a violation of privacy. According to statistics in England every day a person walking in the city center is filmed by 300 cameras. Unfortunately, there are people who use these systems to film reality TV which shows us the daily life of people or to blackmail for their personal benefit. It is for this reason, the surveillance of the public place by the police must be explicitly indicated by signs such as the example in the following figure. SVS is indeed an intrusive aspect, and it is on this point that the authorities have set limits as required by global law No. 95-73 of January 21, 1995 of guidance and programming relating to security. In order to protect the privacy of citizens, we need to find a technique to satisfy the interests of all parties, including the person filming and the one being filmed. The installation of a video surveillance device must be done within a legal framework and with respect for the privacy of individuals.

The main idea of our work is to construct a new security policy in order to have private video surveillance systems by:

• •

  Using regression recurrent neural network to detect person in dangerous scene;

• •

  Hide people present in normal scenes (hide the private information of these people);

• •

  Unmask people present in dangerous scenes;

• •

  Allows special authorities to recover original videos in the event of an investigation;

• •

  Normal agents can only see people in scenes detecting as dangerous;

• •

  Compare your proposal with other deep learning models such as AlexNet, LetNet5 et VGG16.

The manuscript is structured in five sections: 2) review of literature, 3) proposed system, 4) results, discussions and comparisons, 5) general conclusion.

2. Review Of Literature

Recently The technology of video surveillance has received great attention as a very active search area, this section contains a detailed description of several techniques that exist in the literature. Various methods have been proposed in this domain.

In (huang, 2010) a method of classifying scenes based on an improved standard model function. The results obtained are more robust, more selective and less complex (huang, 2010). An image scene sequence is taken in order to detect the objects present in these images with the extraction of spatio-temporal characteristics without preserving the privacy of the people filmed. In (diehl, 2002) The authors have developed a low-cost real-time system. The proposed system consists in interpreting the activity in the environment according to the context of origin and thanks to an incremental automatic learning method. A new solution has been proposed to classify objects based on the detection of recurring movements for each object (Zhang, 2007).