Description
- Stereo Motion (7 pts): In stereo imaging we can compute a point’s world coordinates from left and right images as
𝐵𝐵⃗2
𝑋𝑋⃗𝑊𝑊 = 𝑋𝑋⃗𝐴𝐴𝐴𝐴𝐴𝐴
𝐵𝐵⃗ ∙ ∆⃗
- Show that if the world point is in motion, we can compute its velocity as
𝐵𝐵⃗2 𝐵𝐵⃗
𝑉𝑉⃗𝑊𝑊 = 𝑉𝑉⃗𝐴𝐴𝐴𝐴𝐴𝐴 − 𝑋𝑋⃗𝑊𝑊
𝐵𝐵⃗ ∙ ∆⃗ 𝐵𝐵⃗ ∙ ∆⃗
- Suppose that a moving world point is imaged as
𝑐𝑐 −𝑐𝑐
𝑋𝑋⃗𝐿𝐿𝐼𝐼 = 20𝑑𝑑 , 𝑋𝑋⃗𝑅𝑅𝐼𝐼 = 20𝑑𝑑
𝑓𝑓 𝑓𝑓
With the usual imaging geometry of 𝐵𝐵⃗ = [𝑏𝑏 0 0]𝑇𝑇, 𝐹𝐹⃗ = [0 0 𝑓𝑓]𝑇𝑇, what is 𝑉𝑉⃗𝑊𝑊?
Express 𝑉𝑉⃗𝑊𝑊 in the simplest terms.
- Shading (6 pts): Consider the two surfaces
𝑧𝑧1 = (𝑥𝑥2 + 𝑦𝑦2) and 𝑧𝑧2 = 2𝑥𝑥𝑦𝑦
- Find 𝑝𝑝(𝑥𝑥, 𝑦𝑦) and 𝑞𝑞(𝑥𝑥, 𝑦𝑦) for both surfaces.
- Show that 𝑧𝑧1 and 𝑧𝑧2 give rise to the same shading when a rotationally symmetric reflectance map applies, that is, when 𝑅𝑅(𝑝𝑝, 𝑞𝑞) = 𝑅𝑅(𝑝𝑝2 + 𝑞𝑞2).
- SLAM: Following Bailey & Durrant-Whyte Part II, at time k the vehicle pose 𝐱𝐱𝑣𝑣𝑘𝑘and landmark locations m can be combined into a single state 𝐱𝐱𝑘𝑘.
𝐱𝐱𝑣𝑣𝑘𝑘
𝐱𝐱𝑘𝑘 = 𝐦𝐦
The covariance matrix of 𝐱𝐱𝑘𝑘, denoted by 𝐏𝐏𝑘𝑘|𝑘𝑘, is partitioned into
𝐏𝐏𝑘𝑘|𝑘𝑘 = 𝐏𝐏𝐏𝐏𝑣𝑣𝑣𝑣𝑇𝑇𝑣𝑣𝑣𝑣 𝐏𝐏𝐏𝐏𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣 𝑘𝑘|𝑘𝑘
Control input 𝐮𝐮𝑘𝑘 only affects the vehicle pose, not the landmark locations, via 𝐱𝐱𝑣𝑣𝑘𝑘 =
𝐟𝐟𝒗𝒗(𝐱𝐱𝑣𝑣𝑘𝑘−1, 𝐮𝐮𝑘𝑘). Thus, the update equation for state 𝐱𝐱𝑘𝑘 is
𝐱𝐱𝑘𝑘 = 𝐟𝐟(𝐱𝐱𝑘𝑘−1, 𝐮𝐮𝑘𝑘) = 𝐟𝐟𝒗𝒗(𝐱𝐱𝑣𝑣𝑘𝑘−1, 𝐮𝐮𝑘𝑘)
𝐦𝐦
When control 𝐮𝐮𝑘𝑘 is applied, but before any landmark measurement updates are made, the covariance prediction 𝐏𝐏𝑘𝑘|𝑘𝑘−1 is given by
𝐏𝐏𝑘𝑘|𝑘𝑘−1 = ∇𝐟𝐟𝐱𝐱𝐏𝐏𝑘𝑘−1|𝑘𝑘−1∇𝐟𝐟𝐱𝐱𝑇𝑇 + ∇𝐟𝐟𝐮𝐮𝐔𝐔𝑘𝑘∇𝐟𝐟𝐮𝐮𝑇𝑇
𝝏𝝏𝐟𝐟 𝝏𝝏𝐟𝐟 where ∇𝐟𝐟𝐱𝐱 = , ∇𝐟𝐟𝐮𝐮 = , and 𝐔𝐔𝑘𝑘 is the control convariance. Show that the covariance
𝝏𝝏𝐱𝐱𝑘𝑘−1 𝝏𝝏𝐮𝐮𝑘𝑘
∇𝐟𝐟𝒗𝒗𝐱𝐱𝐏𝐏𝑣𝑣𝑣𝑣,𝑘𝑘−1|𝑘𝑘−1∇𝐟𝐟𝒗𝒗𝑇𝑇𝐱𝐱 + ∇𝐟𝐟𝒗𝒗𝐮𝐮𝐔𝐔𝑘𝑘∇𝐟𝐟𝒗𝒗𝑇𝑇𝐮𝐮
𝐏𝐏 𝐏𝐏𝑘𝑘|𝑘𝑘−1 = 𝑣𝑣𝑣𝑣𝑇𝑇 ,𝑘𝑘−1|𝑘𝑘−1∇𝐟𝐟𝒗𝒗𝑇𝑇𝐱𝐱 𝝏𝝏𝐟𝐟 where ∇𝐟𝐟𝒗𝒗𝐱𝐱 = 𝝏𝝏𝐱𝐱𝑣𝑣𝒗𝒗 and ∇𝐟𝐟𝒗𝒗𝐮𝐮 = 𝝏𝝏𝐮𝐮𝝏𝝏𝐟𝐟𝒗𝒗𝑘𝑘. |
∇𝐟𝐟𝒗𝒗𝐱𝐱𝐏𝐏𝑣𝑣𝑣𝑣,𝑘𝑘−1|𝑘𝑘−1
𝐏𝐏𝑣𝑣𝑣𝑣,𝑘𝑘−1|𝑘𝑘−1 |
prediction simplifies (!) to
𝑘𝑘−1
Hint: Expand ∇𝐟𝐟𝐱𝐱 and ∇𝐟𝐟𝐮𝐮. This problem is not as hard as it looks once you understand what is being asked.