Camera position offset

I’m using an external RTK antenna with centimetric accuracy. When using in car, the antenna is on the roof while smartphone is capturing pictures form inside. Typical offset along the track is about 50 cm while vertical one is around 35 cm. How to take this into account?


Fantastic! The accuracy you are getting is already way better than the recorded coordinates and timestamp in the photos. I tried to add corrections from SBAS. But the timestamps vary +/- several seconds and they are rounded off to the nearest second in spite of the filename having millisecond precision. The lat/long is only to the nearest second too which is several metres. I was assured that it does not matter because some adjustment is done by the structure in motion processing. At least my tracks disnt wander over rooftops.

Yes, I’m using RTK since Mar 2021. Usually, your track is taped to the road as you can see:

Moreover, on the above sample, traffic signs (speed limit) were new, so they were only seen on my pictures. Positions reported in JOSM are also very good (<< 1 m), in opposite to usual positioning where panels may appear very far from real position.

For the capture, I normally use Mapillary app under Android, although I can’t guarantee that position along the track is very accurate. Last summer, because Mapillary was not able to store data on the SD card, I used OpenCamera in timelapse option. Positions were rounded to the arc-second so not good (~30 m). I did post-processing with JOSM, allowing me to add a camera position offset. The drawback is an important time offset of several seconds, not stable. So, I add to fine tune this offset, using road curves and other marks like line painting. Optimistically, I would say I may reached a 0,2 s accuracy (so driving at 108 km/h → 30 m/s → 6 m along the track). I’m still looking for a method to accurately positioning my pictures.

Red cross is the estimated position looking at the picture and landscape (bridge, painted lines…). Position estimated by Mapillary is 23 m away so nearly 0,75 s bias. So, my wet-finger estimate is much better (although much more time consuming for me).