This sounds an awful lot like what I saw last year. I've barely looked at the camera code this year, but last year I found that .mY (which Kevin's code was using to determine if the light was in sight) was being returned non-zero even though .confidence was near zero. I modified the code to use confidence rather than mY. Please see
this post and
the thread containing this post.
N.b. Kevin fixed his code, so if you're using Kevin's latest code you have a different problem or your Confidence_Threshold is set too low.