c++ - Find point of interest on image -

i'm trying track little white dots on edges table. in of case works. i'm using cornerharris function it's used in tutorial : http://docs.opencv.org/doc/tutorials/features2d/trackingmotion/harris_detector/harris_detector.html . sometimes, have got problem : reflection of light on edges creates point of interest have not consider. example : i'm searching 2 nearest points of top corners, can see on right edges, have find dots(red , green dots) , on left edges, light noise problem (cyan , blue dots).

does knows method keep dots white on picture ? thankyou , sorry english

on purely image processing part, recommend using kind of shape feature analysis(like comparing histogram in 8x8 around examined point of interest precomputed ones of features want .

this mean first points harris corners, compare features dismiss unwanted ones ( euclidian distance in 8x8 = 64d ?). of course assumes existence of strong feature (read "taking time find one") assumes know feature points beforehand.

alternative more on computer vision side : use geometry of corner points repartition advantage : want distorted rectangle, make sure find 1 ! surely can compute function gives validity of last feature point assuming 3 others ? (distance of intersection of 2 lines generated other 3 points ...)

the typical , coolest approach apply ransac : try random (but not !) combinations of points , check 1 fits best using function, , consider good.

if intend on tracking on time or on several images, have tune bit, ransac can fail (statistics of random combinations ...), , use points successful run guestimate position.

last idea moment : use color-aware derivation technique : compute harris corner of rgb image or of flattened version gray ? gradients use color information tip discern edges, , i'm not sure corners you're finding use of those. again might mean reimplementing harris corners algorithm (try it, it's fun, , not hard if have algebra library heavy work)

i recommend geometric test of fitting uses wisely model-info of system rather assumptions on how reflections like.

really funny introduction ransac : danielwedge.com/ransac/

edit : trusty photoshop knows mean : highlighted invalid shapes

valid grid, photoshop says so

invalid grid, logical, right ?