TOP  >  Thesis/Dissertation  >  {sotsurontitle}

Simulation for Designing a Person Retrieval System in Non-overlapping Camera Network

Recently, CCTV camera networks are increasingly popular. They are useful for criminal investigation, marketing analysis and so on. Usually application using camera networks requires to search for a person of interest among a database of recorded images. Such system is called retrieval system. In this thesis, we look at the problem of applying retrieval system to large scale real scenarios.

The retrieval process is made of two stages. In the observation stage, the person appearing in the camerafs videos are detected and their appearancefs features are registered into a database. In the search stage, the user inputs the appearance features of a target person to the system. After comparison with a database of registered personfs features, the system returns a set of likely matching candidates.

There are many factors affecting the performance of such system. For example, the retrieval is likely to get more complex when the number of person walking in the observed area increases. However it is unclear to which extend which factors affect the system performances. Our research intends to provide guidelines and insight in the design of retrieval system in real environment.

In order to assess the influence of various parameters independently, we designed an advanced simulator. Our simulator allows us to modify parameters such as, the positions and number of cameras, the topography of the area, the number of pedestrian etc... The changes can be made very easily at no financial cost. This is some of the main reasons why we choose simulation over tedious real world experiments.

We first show the validity of our simulator by comparing results from our simulations with a real world experimentfs results. We then increase the scale of our simulated environment to analyze the impact of various factors on large scale retrieval system.