SteveMacenski
commented
4 years ago Additionally, cover the general dynamic obstacle case, RADAR support is just an example of one particular source of dynamic obstacle information. Complex YOLO and other machine learning techniques can also provide positions and velocities of dynamic obstacles for use.
This ticket covers how to use dynamic obstacle information in the navigation stack (e.g. how to integrate into costmaps, how to integrate into planners and controllers)
The best solution is to integrate the dynamic detections more formally into the costmap model. Such that when we query a cell (X, Y) we also have some relative timestamp and project the obstacles at time t from the motion model. This way its not dependent on some costmap projection model N seconds into the future but rather projecting the motion model forward at any time queried.
The intermediate solution is to have a costmap layer that takes the tracks and adds them to the costmap and then projects them forward in time N seconds by some covariance matrix which informs the cost value to use (lethal, high, low, etc) into a shape defined by variances. We could make a plugin interface for different motion models as well rather than assuming constant velocity so someone can integrate Ai prediction techniques if they like.
elladuffy17
Radars are getting cheaper and cheaper, they make for excellent sensors for static mounted applications for detecting and tracking moving obstacles. Removes the need for ML for detecting things of interest in the scene and give ranging / velocity profiles of blobs.
Investigate if this works well (and computationally efficiently) on mobile robots as autonomous cars. Integrate into the navigation stack after identifying potential straights and weaknesses. In particular, I think this will be valuable for a radar dynamic obstacle tracking costmap-ish layer, or correlating information from visual or lidar detections, or filter tracking/behavior prediction.