Geometric neighborhood model for visual tracking in central catadioptric omnidirectional vision
Yazhe Tang, Youfu Li, Jun Luo · 2014
Central catadioptric omnidirectional vision (CCOV) exhibits serious nonlinear distortion with a quadratic mirror involved. Conventional pinhole model based features perform poorly when directly applied over deformed CCOV. To construct an efficient, distortion involved neighborhood model, a complete catadioptric geometry system which consists of the object and the omnidirectional sensor is analyzed. According to the catadioptric omnidirectional geometry, a neighborhood mapping model that can accurately model the distortion of CCOV is developed. With the analyzed catadioptric geometry, the proposed neighborhood mapping model can efficiently reflect a relationship between the 2D neighborhood of an object and its radial distance on the omnidirectional image. Based on the proposed neighborhood mapping model, a distortion-invariant Haar wavelet transform is proposed for visual tracking in CCOV. Experiments have validated the effectiveness of the proposed neighborhood mapping model.