Download or read book Design and Performance of Multi-camera Networks written by Itai Katz and published by . This book was released on 2010 with total page pages. Available in PDF, EPUB and Kindle. Book excerpt: Camera networks have recently been proposed as a sensor modality for 3D localization and tracking tasks. Recent advances in computer vision and decreasing equipment costs have made the use of video cameras increasingly favorable. Their extensibility, unobtrusiveness, and low cost make camera networks an appealing sensor for a broad range of applications. However, due to the complex interaction between system parameters and their impact on performance, designing these systems is currently as much an art as a science. Specifically, the designer must minimize the error (where the error function may be unique to each application) by varying the camera network's configuration, all while obeying constraints imposed by scene geometry, budget, and minimum required work volume. Designers often have no objective sense of how the main parameters drive performance, resulting in a configuration based primarily on intuition. Without an objective process to search through the enormous parameter space, camera networks have enjoyed moderate success as a laboratory tool but have yet to realize their commercial potential. In this thesis we develop a systematic methodology to improve the design of multi-camera networks. First, we explore the impact of varying system parameters on performance motivated by a 3D localization task. The parameters we investigate include those pertaining to the camera (resolution, field of view, etc.), the environment (work volume and degree of occlusion) and noise sources. Ultimately, we seek to provide insights to common questions facing camera network designers: How many cameras are needed? Of what type? How should they be placed? First, to help designers efficiently explore the vast parameter spaces inherent in multi-camera network design, we develop a camera network simulation environment to rapidly evaluate potential configurations. Using this simulation, we propose a new method for camera network configuration based on genetic algorithms. Starting from an initially random population of configurations, we demonstrate how an optimal camera network configuration can be evolved, without a priori knowledge of the interdependencies between parameters. This numerical approach is adaptable to different environments or application requirements and can efficiently accommodate a high-dimensional search space, while producing superior results to hand-designed camera networks. The proposed method is both easier to implement than a hand-designed network and is more accurate, as measured by 3D point reconstruction error. Next, with the fundamentals of multi-camera network design in place, we then demonstrate how the system can be applied to a common computer vision task, namely, 3D localization and tracking. The typical approach to localization and tracking is to apply traditional 2D algorithms (that is, those designed to operate on the image plane) to multiple cameras and fuse the results. We describe a new method which takes the noise sources inherent to camera networks into account. By modeling the velocity of the tracked object in addition to position we can compensate for synchronization errors between cameras in the network, thereby reducing the localization error. Through this experiment we provide evidence that algorithms specific to multi-camera networks perform better than straightforward extensions of their single-camera counterparts. Finally, we verify the efficacy of the camera network configuration and 3D tracking algorithms by demonstrating their use in empirical experiments. The results obtained were similar to the results produced by the simulated environment.