Download presentation
Presentation is loading. Please wait.
1
Visibility-based Motion Planning Lecture slides for COMP 790-058 presented by Georgi Tsankov
2
Topics Art gallery problem Sensor placement Map generation Finding a target (pursuit-evasion) Following a target
3
Topics Art gallery problem Sensor placement Map generation Finding a target (pursuit-evasion) Following a target
4
Art gallery problem Question: how many cameras do we need to guard a polygonal art-gallery ? NP-hard Bounds: O(n/3), examples http://www.cs.wustl.edu/~pless/506/l6.html
5
Topics Art gallery problem Sensor placement Map generation Finding a target (pursuit-evasion) Following a target
6
Sensor Placement Definition Art gallery + more visibility constraints [1]: Free path Range constraint (min, max) Incidence constraint
7
Sensor Placement Algorithm Also NP-hard, so a good and fast approximation is needed. Random Sampling: Sample M random points and decompose the boundary in R 1,… R N – intervals visible by different sets of points. Optimal set cover problem – also NP-hard, but efficient approximations exist.
8
Sensor Placement Set cover problem
9
Sensor Placement Inspection Routes Extension: generating inspection routes. [7] Cost of camera vs. cost of movement Definition: Compute the shortest path for a single camera to inspect the whole boundary. Optimal solution for simple polygons: O(N 4 ) NP-hard when there are holes!
10
Sensor Placement Inspection Routes Approximate algorithm: Repeat until the boundary is covered Choose an unguarded point P i on the boundary and build its visibility region V(P i ) Sample K points inside V(P i ) and pick the one with the best gain. Connect the points: TSP with triangle inequality - a 2-approximate algorithm. Extension to 3D Very hard to compute visibility polyhedron! Simplifications
11
Topics Art gallery problem Sensor placement Map generation Finding a target (pursuit-evasion) Following a target
12
Map Generation Definitions Goal: Build a map of unknown environment. [9] On-line version of the Sensor Placement problem Partial map, disconnected. Problem: Given a partial map, where should we go next (NBV) ?
13
Map Generation Next Best View How do we compute the next position? Set of candidates on the free boundary value = -motion cost + visibility gain
14
Map Generation Example
15
Topics Art gallery problem Sensor placement Map generation Finding a target (pursuit-evasion) Following a target
16
Finding a Target (pursuit-evasion) Definitions Problem: Given: [3] A known 2D polygonal environment with obstacles A target robot with unrestricted speed and unknown initial position plan the movement of a pursuer, so that it eventually sees the target. Applications: Surveillance Search and rescue operation Search for other autonomous robots
17
Finding a Target (pursuit-evasion) Definitions Contaminated/Clear regions Information state: (q, S), the goal is (q, {})
18
Finding a Target (pursuit-evasion) Definitions Visibility region: V(q), gap edges, solid edges. B(q) = (0,1,0,…) - binary vector describing the information state.
19
Finding a Target (pursuit-evasion) Definitions Discretize the space Conservative regions All points inside see the same edges. (the information state stays the same) How do we build them?
20
Finding a Target (pursuit-evasion) Conservative regions Constructing the conservative regions Put lines whenever a new obstacle edge enters/leaves the visibility region. Extend all edges Extend pairs of vertices
21
Finding a Target (pursuit-evasion) Information graph Once we have the regions, build a graph G, and information graph G I :
22
Finding a Target (pursuit-evasion) Information graph How do we compute the edges ? A gap edge disappears – ok A gap edge appears – make it clear (0) q1->q2 – a gap edge disappears q2->q1 – a gap edge appears
23
Finding a Target (pursuit-evasion) Information graph Computing the gap transitions (cont’d): Gap edges merging - OR them One gap edge splits – assign the same value q3->q4 – two gap edges merge q4->q3 – a gap edge splits
24
Finding a Target (pursuit-evasion) Traversing the graph Just find a path to some goal node (00..0) This is a complete algorithm for one pursuer! Assumes unrestricted visibility.
25
Finding a Target (pursuit-evasion) Multiple pursuers What if one pursuer can’t make the job? How many pursuer do we need: H(F) ? Depends on the geometry of F. Upper bound O(sqrt(H) + log(N)) H(F) can not be computed exactly.
26
Finding a Target (pursuit-evasion) Bounded target velocity Evader with bounded velocity [2] The contaminated regions are more restricted
27
Finding a Target (pursuit-evasion) Summary For 2D: Complete algorithm for 1 pursuer (with ideal visibility). Bounds for number of pursuers, strategy for multiple observers. For 3D: How to compute visibility polyhedron ? Are gap transitions the same as in 2D ? Demos - http://robotics.stanford.edu/groups/mobots/pe.html http://robotics.stanford.edu/groups/mobots/pe.html
28
Topics Art gallery problem Sensor placement Map generation Finding a target (pursuit-evasion) Following a target
29
Following a Target Applications Applications: Camera movement in VE Surveillance Monitoring (automated) processes Medicine Military needs Relation to Game theory (the target actively tries to escape visibility)
30
Following a Target Definitions Visibility constraints: free space, range, incidence Visibility sweeping line – the line passing through the target and a reflex vertex. [1]
31
Following a Target Types of problems Critical vs. Average tracking Critical: Must never lose the target (can’t find it later). The goal is to maximize the escape time. Average: OK to lose it (probability to re-acquire it at some time). The goal is to maximize the average visibility in some time interval.
32
Following a Target Types of problems Predictable vs. Unpredictable target motion. [4]
33
Following a Target Known target motion Algorithm for predictable motion: For each time step t, determine V(t) – set of points seeing the target. The pursuer should be in V(t) for all t. Restrict V(t) to points, reachable from V(t-1) in one time step Do this for all t. Complete algorithm, but exponential on dimensions. The space must be discretized.
34
Following a Target Known target motion
35
Following a Target Unpredictable target Unpredictable target motion No complete algorithm, need approximation The time horizon is reduced. (one time step). Average vs. Critical tracking
36
Following a Target Unpredictable target – average tracking Maximizing the probability of visibility in the future (one) time steps. Random motions for the target are considered (heading, velocity) for the time step Disc with uniform density Random next positions for the observer are sampled and intersected with the disc. Choose the best one – maximizes visibility in the next time step.
37
Following a Target Unpredictable target – critical tracking Maximizing the escape time Shortest distance to escape (SDE) Select next position, which maximizes SDE
38
Following a Target Other problems Other kinds of problems: Unknown environment [5] Stealth tracking [8] Multiple observers / targets [6]
39
Following a Target Unknown environment Unknown environment – have only local map Build Escape Path Tree (EPT)
40
Following a Target Unknown environment Assign “Escape Risk” to each free edge. Local coordinate system: n: towards the occluding vertex t: orthogonal, increases SDE Compute the gradient which minimizes the risk
41
Following a Target Unknown environment The direction for the observer is the average gradient over all escape paths Benefits of the tree (EPT)
42
Following a Target Stealth tracking Stealth tracking The observer tries to stay outside of the target’s visibility region. Lookout region. Escape risk.
43
Following a Target Multiple observers / targets Multiple observers / targets Most of the algorithms can be reused. In [6] the algorithm is exponential on the number of observers. Choosing a metric: worst vs. average risk.
44
Following a Target Summary What we have: Complete algorithm for predictable target. Good approximations for unpredictable target. Algorithm for multiple observers/targets is not efficient enough, and is centralized. Future work: 3D environments Improved algorithm for multiple observers.
45
Topics covered Art gallery problem Sensor placement Map generation Finding a target (pursuit-evasion) Following a target
46
Acknowledgments Pictures were taken from: www.cs.cmu.edu/~motionplanning/lecture/Chap6- CellDecomp_howie.pdf www.cs.cmu.edu/~motionplanning/lecture/Chap6- CellDecomp_howie.pdf Lectures from Latombe’s course: http://robotics.stanford.edu/~latombe/cs326/2004/schedule. htm http://robotics.stanford.edu/~latombe/cs326/2004/schedule. htm The papers on the next page.
47
References 1. Motion Planning: Recent Developments. - Gonzalez Banos, D. Hsu, Latombe (pp. 19-32). 2. Visibility-Based Pursuit-Evasion with Bounded Speed. - B. Tovar, S. LaValle. 3. Finding an Unpredictable Target in a Workspace with Obstacles. - S. LaValle, D. Lin, L. Guibas, Latombe, R. Motwani. 4. Motion Strategies for Maintaining Visibility of a Moving Target. - S. LaValle, H. Gonzalez-Banos, C. Becker, Latombe. 5. Real-time Combinatorial Tracking of a Target Moving Unpredictably Among Obstacles. - H. Gonzalez-Banos, D. Hsu, Latombe. 6. A Sampling-Based Motion Planning Approach to Maintain Visibility of Unpredictable Targets. - R. Murietta-Cid, B. Tovar, S. Hutchinson. 7. Randomized Planning for Short Inspection Paths. - T. Danner, L. E. Kavraki. 8. Stealth Tracking of an Unpredictable Target among Obstacles. - T. Bandyopadhyay, Y. Li, M. H. Ang, D. Hsu. 9. Navigation Strategies for Exploring Indoor Environments. – H. Gonzalez- Banos, Latombe 10. Visibility-Based Pursuit-Evasion in Three-Dimensional Environments. – S. Lazebnik
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.