
Online State-Time Trajectory Planning
in Highly Dynamic Crowd Environments

Project Description
-
This project aims to address the navigation problem in high-dynamic environments, particularly in dense crowd scenes.
-
This project proposed a gradient-based planner over the state-time space for online trajectory generation.
-
This project porposed timed ESDT arguments that support distance and gradient queries with state-time keys to optimize motion trajectory; defined smooth prior and obstacle likelihood functions compatible with the state-time space.
-
This project valid the method with simulated and benchmark datasets.
-
More information can be found in this paper.
Project Method
Front-End Path Searching
-
State-Time Graph Building:
-
Graph Definition : Instead of searching in pixel level, we define the timed triangle graph, a topological representation of moving obstacles. It is composed of edge and face, which depend on pedestrians' positions and velocities.
-
Dual Graph: We discrete the time dimension into T slices with a resolution dt. We then simplify eachi triangle into a node and build feasible connections for adjacent nodes.
-
State-time Graph Building: We use the optimization method to execute the node-placement procedure when doing the searching.
-
-
State-Time A*:
-
Once the time budget T is used up or the searching process fials, the current optimal path is returned.
-
The neghtor retrival, i.e. SUCC is based on the graph, so the connections across diffrent time slices are built up.
-
The node placement is performed along the progress of the path searching.
-

BACK-END TRAJECTORY OPTIMIZATION
-
Prior Dsitribution of Smooth Trajectories:
-
To parameterize smooth trajectories, a GP prior is defined over the trajectory space, and each trajectory is sampled from this prior distribution.
-
Suppose a near-optimal time assignment t0< t1< t2 < ... < tf, is found by the front end, then the trajectory is discretized into F+1 waypoints accordingly.
-
-
Timed ESDF:
-
ESDF is frequently used cost map, shown in Fig2, which contains distance and gradient information against obstacles.
-
We defined the likelihood function to evaluate the clearance of the trajectory
-
-
MAP Trajectory:
-
Based on GP prior and likelihood function, the back-end trajectory optimization os completed by solving a MAP problem, in equation (17)
-

Experimental Result
Demonstration

Statistical Analyses
