Researches On Catadioptric Omnidirectional Non-Uniform Compressive Imaging

Catadioptric omnidirectional imaging systems consisting of conventional cameras and curved mirrors for capturing 360o field of view, owing to the advantage of one-shot seamless panoramic imaging, are widely used in many vision applications, such as robot navigation, surveillance, aerial photographic reconnaissance, medical applications, video conferencing. However, for the further investigation of the research, the problem of low and non-uniform spatial resolution in omnidirectional imaging has become the bottleneck.Recently emerged compressed sensing theory provides a new approach for solving the resolution problem of catadioptric imaging. Based on the sparsity or compressibility of the signal, compressed sensing performs acquisition and compression simultaneously, and caputres the compressive sampling of signal through incoherent measurement. It can reconstruct the original signal from the much less measurements than the traditional method, makes the acquisition of high dimension signal to be possible. This paper investigates the theory of compressed sensing in catadioptric imaging to solve its resolution problem. Our work mainly includes:1. Non-uniform measurement matrix design according to the resolution characteristic of omnidirectional image.Due to the existence of reflective mirror, the resolution of omnidirectional image reduces gradually from the periphery to the center of the image. This makes the rational uniform measurement matrix design unsuitable for processing of omnidirectional image. Therefore, the non-uniform measurement matrix is designed in this paper based on the resolution distribution of the omnidirectional image. We analyze relation of the resolution of catadioptirc sensor and traditional sensor based on solid angle. Then, according to the relation between mean square error and sparsity, measurement number, a non-uniform measurement matrix, which is based on the distribution of the image system’s resolution, is designed in this paper. Based on the designed measurement matrix, we can allocate more sensing resources to the inner pares and fewer to the outer parts of the omnidirectional image. Lastly, experimental results show that the proposed method is feasible and effective for solving the resolution problem of omnidirectional image.2. Omni-gradient based total variation minimization for sparse reconstruction of omnidirectional image.Traditional method for computing gradient is not appropriate for omnidirectional image processing since the distortions by the reflective mirror in catadioptric imaging systems are different. The omni-gradient computing method combined with the characteristics of omnidirectional imaging is proposed in this paper. We analyze the characteristics of gradient in omnidirectional image, and present a novel gradient computing method for omnidirectional imaging based on bidirectional mapping. Then, a new total variation model based on omni-gradient, called omni-total variation, is porposed. In order to reconstruct the image from its compressive samples, the omni-total variation regularization based on omni-gradient is utilized in the image resoration. The experimental results show that the images recovered based on omni-total variation model can provide higher quality both in subjective evaluation and objective evaluation, and the inner part of the image has obviously higher quality.3. Catadioptric omnidirectional compressive imaging hardware platform.From the views of the imaging system design and implementation, we explore to the approach to realize compressed samping through physical devices. Firstly, we describe an idea of compressive coded aperture, which is a parallel compressive imageing. Based on the analysis, we introduce the design of coded mask through random point spread function(PSF) to avoid the storage of high dimensional measurement matrix. Then the implementation of coded aperture is discussed in this paper. We impliment the coded aperture method by using relay lens and the programmable device. At last, we verify the effectiveness of the catadioptric omnidirectional compressive imaging through the physical device experiment.