How do self-driving bots usually deal with transient objects, e.g., parked cars on the side of roads when they can come and go? These aren't moving objects at the time of capture, but they do move at a later time. Is it common to run a CNN to detect static/dynamic objects prior to feature extraction? It's certainly much more computationally expensive, but otherwise how are dynamic objects dealt with or filtered out?
$\begingroup$
$\endgroup$
1
-
$\begingroup$ Also see OPs related question: robotics.stackexchange.com/q/18426/16622 $\endgroup$– DavidJAug 9, 2019 at 20:15
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
How do self-driving bots usually deal with transient objects, e.g., parked cars on the side of roads when they can come and go?
No. In most of the open-source slams, dynamic objects are ignored which means they are just mapped as a stationary object. But there are few papers that deal with this in the way you think.
These aren't moving objects at the time of capture, but they do move at a later time. Is it common to run a CNN to detect static/dynamic objects prior to feature extraction?
It is not common. There are some papers that introduced object-level features but simply it is not very useful.
It's certainly much more computationally expensive, but otherwise how are dynamic objects dealt with or filtered out?
Any features from dynamic objects should be filtered out by either deep learning approach or conventional soft outlier rejection scheme. You will find the most of optimization-based slams are using M-estimator whereas some of the recent papers introduced deep learning approach.
This kind of topic is refered as semantic SLAM. You can find some related papers in ICRA, IROS 17,18,19.
$\begingroup$
$\endgroup$
The answer of @C.O Park is pretty good. I will just add that when you SLAM, either you are doing it to localize the vehicle and thus the map is a consequence, in this case using transient objects which are stationary at the moment is good because they will help with the scan matching.
If you are SLAMing to localize the vehicle currently and then use the map later on, the transient objects are useful for the current run, however they may not be there for the following runs. In this case they might affect the results of the scan matching for subsequent runs. In this case it might be interesting to filter such objects to improve the quality of the generated map as we suppose that the map generated contains static objects useful for localization.
