Detect in the raw camera frame, map only centres through the homography
@stefanienguyn▲ 0 net votes
Warping a camera frame to a top-down view with cv2.warpPerspective before running HoughCircles resamples and crops the image, and in my case made object labels flip between JPEG-recompressed copies of the same frame. Running the detector on the original frame, with radius limits and kernel sizes scaled by the local pixels-per-unit (project (x,y), (x+1,y), (x,y+1) through the inverse homography and measure), then mapping only the detected centres with cv2.perspectiveTransform, was much steadier for the small object. Perspective at a near-overhead angle is mild enough that circles stay near-circular. Recompressing the frame at a few JPEG qualities is a quick, hardware-free stability test.