Research Group on Visual Computation

Description

Smart phones with integrated cameras have became very popular in the last couple of years. Introduced as a new development platform, yielded many new possible applications. The main advantages of smart phones are the availability of a powerful embedded processor, networking capabilities and multiple type of sensors (e.g. GPS, accelerometer, gyroscope). Using a set of mobile phones, a distributed, highly scalable intelligent camera network can be created, which is desirable in many tasks of vision based applications, like localization, reconstruction, object recognition and tracking, HDR imaging of dynamic events. Many of these applications requires a calibrated camera system.

Since mobile phones typically have built-in fixed focus cameras, their intrinsic parameters can be easily calibrated using various algorithms and simple calibration patterns. Therefore we focus on the estimation of extrinsic camera parameters (i.e. pose estimation) of such ad-hoc mobile camera networks.

The main challenges in calibrating an ad-hoc camera network are the following

1. Every computation task is done on the smart phones, without any central server.
2. We have to handle events like camera appearing or disappearing.

The main steps of our algorithm

1. Calculate the relative pose within the network. For this step we are using extracted SIFT/SURF features and the relative pose is factorized from the essential matrix.
2. Relate an arbitrary camera of the network to a planar surface. In our solution we are assuming that, the planar surface is containing a low-rank texture (Z. Zhang, A. Ganesh, X. Liang, Y. Ma, TILT: Transform Invariant Low-Rank Textures, IJCV, 2012).
3. Calibrating the whole network w.r.t. the 3D world. In this phase, the proposed approach tries to find the extracted and reconstructed planar patch using normalized mutual information in every camera of the system.

Results