Human detection and tracking in central catadioptric omnidirectional camera system

Abstract

Catadioptric omnidirectional vision (COV) as a novel imaging system has drawn significant interest in recent decades. Compared with the conventional imaging system, a catadioptric omnidirectional sensor can achieve a global field of view from a single image with compact configuration. Therefore, this sensor can be used in various applications such as public surveillance, intelligent vehicles, and robotics, among others. However, the conventional pinhole model has difficulties in describing the imaging geometry of a catadioptric vision system because of the inherent distortion introduced by the catadioptric quadratic mirror. Consequently, most conventional methods cannot be easily implemented in COV. To solve this problem, this thesis presents solutions from the non-parameterized to parameterized model to realize human detection and tracking in COV. First, this thesis addresses the problem by using a non-parameterized solution. Since this method does not require the system model, it is suitable for the systems that are difficult to be calibrated, such as thermal omnidirectional camera. In the past few years, several widely used approaches have been developed based on the non-parameterized unwrapped panoramic image. Unwarpping based methods are simple to implement, but it can introduce noise during unwrapping despite interpolation and risk splitting the targets located on the boundary of the panoramic image. To maximally maintain information integrity, a contour coding-based rotating adaptive model (RAM) is proposed, which can directly extract the contour information from original catadioptric image. RAM can use the relative angle based on the characteristics of COV to automatically change the sampling order. Without interpolation, the RAM can improve the efficiency of algorithms and avoid introducing noise. Most state-of-the-art contour coding-based features can be integrated with the proposed non-parameterized RAM for application in COV. COV exhibits inherent nonlinear deformation because of the quadratic mirror involved. Therefore, accurate modeling of uneven deformed catadioptric vision by non-parameterized approaches may be difficult to implement. To analyze the neighborhood of an object on the sensor plane, a completely catadioptric geometry system including the object and the omnidirectional sensor is built. Distortion involving a neighborhood mapping model is developed based on the catadioptric geometry system. This model reflects a useful and important mapping relationship between the distorted neighborhood of an object and its radial distance in the image plane. The proposed neighborhood mapping model can accurately model the distortion of COV because the system parameters are comprehensively considered. Some state-of-the-art features can be integrated with the proposed neighborhood mapping model to achieve distortion invariance in catadioptric vision. Moreover, the efficiency of the algorithms combined with the proposed neighborhood mapping model can be improved significantly. Experiments verify that the proposed neighborhood mapping model provides a satisfactory performance compared with the state-of-the-art approaches. A spatial-contour Gaussian mixture model (GMM) based on the proposed neighborhood mapping model is presented to robustly handle partial occlusion in distorted omnidirectional vision. Normalized chamfer matching is employed to measure the contour fragment information of the target. The spatial-contour joint probability distribution is modeled using a mixture of Gaussians. The contour fragments can be organically unified into an integrated probabilistic model by shape clustering. A weight contribution mechanism is proposed to adaptively weigh the spatial-contour fragments based on the responses of the joint features. A more robust and stable performance can be observed using flexible weighted fragment models than the global coding model.

Date of Award	2 Oct 2013
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	You Fu LI (Supervisor)