Visual Odometry through Appearance- and Feature-Based Method with Omnidirectional Images

Figure 4

Given a detected point ^→𝑝₁ in the first image reference system, a point distribution is generated to obtain a set of multi-scale points 𝜆_𝑖→𝑝₁. By using the predicted motion transformation 𝑀_𝑗, they can be transformed to 𝑞^→𝑖 in the second image reference system through 𝑀_𝑗∼N(𝜃̂,𝜙,𝜎_𝜃,𝜎_𝜃). Finally 𝑞^→𝑖 are projected into the image plane to determine a restricted area where correspondences have to be found.