2023-10-26, 08:10 PM
(2023-10-26, 05:50 PM)Hillzzz Wrote:(2023-10-26, 03:20 PM)barrymac Wrote:(2023-10-26, 12:38 AM)seandepagnier Wrote: Using cameras to see objects that they can alarm humans about is different from using cameras to enhance the autopilot.
Agreed, this thread has veered somewhat from the original intelligent autopilot subject and maybe a new thread regarding object detection is justified.
In my previous post I suggested that you could use object detection to improve navigation. In other words instruct the autopilot to steer the vessel the correct side of a cardinal buoy for example. If however the system also detected a hazard up ahead surely it would be negligent not to raise an alarm?
I don't see that is incompatible with the aims of the original post but suggest its an enhancement however If you prefer I will set up an alternative thread to discuss navigation using object detection.
Well thread management issues aside, I really like the way fancy radars put targets on the chart plotter that look like AIS targets and have the same kind of functionality. It needs high quality gyro input to create a stable overlay on the chart. I think an object detection system could work well like this.
So, with help from GPT-4, the following NMEA 2000 PGNs might be relevant:
"PGN 129808: Radar Data - This PGN is specifically used to communicate radar data. It includes information about radar targets, their position, speed, trajectory, and other related data.
PGN 129039: AIS Class A Position Report - Even though this PGN is typically used for AIS data, a similar format might be used to display radar-detected targets on a chart plotter if they're shown similarly to AIS targets.
PGN 129040: AIS Class B Position Report - Again, this is primarily for AIS, but if the system displays radar targets in a manner similar to AIS Class B vessels, this PGN might come into play.
PGN 129291: Set & Drift, Rapid Update - This could be used for information about the current set and drift, which could affect the trajectory calculations of the radar-detected objects."
Regarding the synthetic training data issue, I believe the practice works something like this.
1. Collect a good set of real data from cameras installed in the field.
2. Manually label the objects using human brains and eyeballs
3. Use that data set as a seed for generating a great many more examples, which of course would be automatically labelled now.
4. Introduce some randomness and visual adversity to make things more challenging for the target model, increase the sea state, add fog etc
I believe there may be services available already that can do this, if not, then surely before long.