Visual surveillance systems for the identification and characterization of anomalous behavior


The overall objective of the research proposed in this project is the design of vision-based services aimed at facilitating the monitoring of an area with poor visibility. camera is one of the most versatile sensors. With it, you can develop many services to improve visibility in space with obstacles or poor lighting conditions.

We propose a design of a real-time surveillance system working on Graphics Processing Units (GPUs), with ccapabilities to identify and characterize anomalous behaviors in restricted environments by representing and interpreting the movement in support of preventive actions security systems. The high parallel processing power of GPUs allow parallel processing of information from multiple cameras.

The proposed system will also activate mechanisms that facilitate the analysis of situations (eg activate illumination of an area or aiming the camera to a target).

Also will apply techniques of characterization and recognition of movement to facilitate recognition and tracking of people.

The specific objectives to be achieved by the project are:

  • Specification and design of an information processing system of human figures from sequences of images of low quality.
  • Specification and design of an identification system of anomalous situations that allow identify risk behaviours by analyzing a sequence of video images captured.

Development of a prototype. From the design of vision applications and system architecture is proposed to develop a prototype on GPU to validate the system both in a restricted environment such as the lobby of a building or parking of vehicles or public places, and large public settings such as museums and transport systems (railway stations or airports).

The most innovative aspect of the project is the use of active representation models based on self-organizing neural networks for representing objects and their movement. Moreover, these networks can represent actions in the scenes, classify and allow the detection of anomalous behaviors. Given the parallel intrinsic model of neural networks, its implementation on GPUs using CUDA language allow large increases of speed-up in all processes.

In order to comply with legal aspects of the organic law of data protection can provide access to system operators, only the active representation obtained by the neural models to prevent identification and thus preserving the privacy of individuals under observation .