Median Trajectories
Authors
- First Online:
- Received:
- Accepted:
DOI: 10.1007/s00453-012-9654-2
- Cite this article as:
- Buchin, K., Buchin, M., van Kreveld, M. et al. Algorithmica (2013) 66: 595. doi:10.1007/s00453-012-9654-2
Abstract
We investigate the concept of a median among a set of trajectories. We establish criteria that a “median trajectory” should meet, and present two different methods to construct a median for a set of input trajectories. The first method is very simple, while the second method is more complicated and uses homotopy with respect to sufficiently large faces in the arrangement formed by the trajectories. We give algorithms for both methods, analyze the worst-case running time, and show that under certain assumptions both methods can be implemented efficiently. We empirically compare the output of both methods on randomly generated trajectories, and evaluate whether the two methods yield medians that are according to our intuition. Our results suggest that the second method, using homotopy, performs considerably better.
Keywords
Trajectories Geometric algorithms Homotopy1 Introduction
A relatively new type of geometric data that is being collected and analyzed more and more often is the trajectory: a path through space and time that a certain object traverses. This is due to technological advances like GPS, RFID tags, and mobile phones, and has caused an increase in demand for analysis possibilities. New analysis methods for trajectory data have been developed in the last few years, but a number of basic concepts are still lacking a satisfactory study. One of these concepts is the median trajectory for a given collection of trajectories. Intuitively, a median trajectory is a trajectory that uses pieces of the trajectories of the collection and is somehow in the middle. However, it is not clear how this concept should be defined. In this paper we establish criteria that we believe a median trajectory should meet, and we develop two median definitions that meet these criteria. Furthermore, we give algorithms to compute the median trajectory according to these definitions, and analyze experimentally whether our definitions give useful output.
1.1 Trajectories
Trajectories are a type of geographic data that have a temporal and a spatial component. Trajectories describe the locations over time of an entity that can move. The entity can be a person, animal, vehicle, hurricane (eye of), shopping basket (with an RFID tag), or any other moving object. We assume that the movement is continuous, but is measured at a discrete set of times.
Formally, a trajectory is the time-stamped path taken by a moving object, and is typically represented by a sequence of n+1 tuples of points and time stamps, (p _{0},t _{0}),…,(p _{ n },t _{ n }), which are points in space-time, where space is two- or three-dimensional. In this paper space is always two-dimensional. A collection of m trajectories τ _{1},…,τ _{ m } therefore gives rise to an input size of Θ(nm). In some applications, the time stamps of the m trajectories are exactly the same, while in other applications they are different. In general, trajectories can be collected with different or irregular sampling rates, at different times, and data can be missing. In between time stamps, we have no knowledge of the movement of the entity. The standard assumption is that the moving object moves with constant velocity from a time-stamped point to the next time-stamped point. Therefore, the path of a trajectory is a polygonal curve with n edges that can self-intersect, and can have repeated vertices if the entity does not move. Often, the number of edges of a single trajectory is much larger than the number of trajectories in a set, that is, n≫m.
1.2 Trajectory Analysis
Analysis methods for trajectories have been developed in Geographic Information Science and in Data Mining. Sets of trajectories can be analyzed in a variety of ways. They can be clustered into a collection of subsets that have a high within-subset similarity and a low across-subset similarity (e.g. [19, 28] and many more). They can be classified if a clustering is given [29]. Also, movement patterns on them can be computed [8, 20, 27]. Movement patterns that have been defined and for which algorithms have been suggested are flocking, convoys, herds, leadership, commuting, encounter, and various others. These intuitively represent similar movement in a group (same location), similar movement over a time span (same heading), or movement to the same position (same destination). An overview and classification of movement patterns relevant to trajectory data can be found in [13].
Several analysis tasks require a definition of (and algorithms for) similarity of trajectories. For instance, a simple similarity measure for trajectories is the average distance at corresponding times. With a similarity measure, or its inverse, a distance measure, clustering methods can easily be given. Single linkage and complete linkage clustering need a similarity measure only, and that defines the clustering. On the other hand, k-means and k-medoids clustering requires a definition of the mean and the median, respectively, regardless of whether the data are numbers or trajectories.
1.3 Mean and Median Trajectories
The intuition behind a mean trajectory is that it averages locations, one of each trajectory, like a center of gravity. In contrast, the intuition behind a median trajectory is that it always is central with respect to the number of given trajectories. Imagine a collection of GPS tracks from hikes by different people on different days. The hikers may have followed the same route globally, but there may have been options like going left around a lake or right, or taking a detour to a viewpoint. From their tracks, we want to extract a good global route. Notice that if in such a data set, seven hikers went left around a lake and three went right, then a mean trajectory would actually go through the lake, whereas a median trajectory would go with the group of seven. Similarly, with a side path to a viewpoint and back, if the majority goes to the viewpoint, then a median trajectory should do so as well. A mean trajectory might go partially to the viewpoint, which does not make sense in this context.
1.4 Overview of Results
In Sect. 2 we discuss the idea of median trajectories. No definition has been suggested yet, so we investigate properties that a suitable median should have. We first propose a simple definition (simple median) that directly follows the definition of a level in an arrangement of lines, a standard concept in computational geometry [17]. We also propose a more refined definition (homotopic median) that uses geometric and topological concepts, and may be better suited to most applications that involve trajectories. Then we discuss the maximum combinatorial complexity of a median according to these definitions.
In Sect. 3 we prove that both median definitions have the property that it is always in the middle (in a sense to be defined later).
In Sect. 4 we present algorithms that compute median trajectories according to the two definitions. We can compute the simple median in O((nm)^{2}) time, and the homotopic median in O((nm)^{2+ϵ }) time for any ϵ>0, in the worst case. Here, m is the number of given trajectories and n is the maximum complexity of any trajectory. We improve our algorithms for practical situations. We can compute the simple median in O((nm+k)α(nm)log(nm)) time, where α is the inverse Ackermann function and k is the number of vertices of the median, i.e., the output complexity. Under certain assumptions related to the sampling of the trajectory, we can compute the homotopic median in O(nmlog^{2}(nm)+kα(nm)log(nm)) time. We note that k=O((nm)^{2}) in the worst case, as we show in Sect. 2.2. One would expect that typically, k=Ω(n) and k=O(nm).
In Sect. 5 we give results of tests that we obtained from an implementation. We use a random similar trajectory generator and analyze the length, total angular change, and description size of median trajectories according to both definitions. In Sect. 6 we discuss our results and suggest directions for further research.
2 On the Definition of a Median Trajectory
Let us consider median trajectories. Trajectories include a temporal component as well as a spatial component, but it is not clear whether a median can take the temporal component into account in a useful way. We discuss some examples. Suppose the trajectories came from a group of animals that were traveling in a herd. Then we can use the temporal component because we know that the animals were together at any point in time. Next, suppose that the animals were traveling solitary, according to a similar route. Then they traveled on different days or months, and we cannot use the temporal component. Even if the animals had the same starting location of the route, we cannot simply align the starting times of the travel, because one animal may have been held up due to a predator, which upsets the time correspondence that we assumed at the start. The same is true for trajectories of cars with the same origin and destination: an initial time correspondence may easily be upset due to traffic lights or traffic conditions.
Thus, in many situations we want a median trajectory that does not take the temporal component into account. A similar motivation was given for trajectory similarity measures: many of these are partly shape-based, like dynamic time warping or largest common subsequence [26], or fully shape-based (ignoring the temporal component), like Hausdorff distance [4] or Fréchet distance [5]. Hence, we will concentrate on medians of trajectories based on the path of the trajectory. The median that we will define and compute will therefore be the path of a median trajectory. With slight abuse of terminology, we will just write “median trajectory” for brevity.
We note that with a temporal component, some research on modeling motion and kinetic data structures is related to the median (or mean) trajectory (e.g. [1–3]).
2.1 Requirements for a Median Trajectory
Let a set T={τ _{1},…,τ _{ m }} of m trajectories be given, each containing n vertices. We assume for convenience that all trajectories start at the same point s and end at the same point t; this is a strong and unrealistic assumption, but we are interested in a clean definition of the median where behavior at the ends is not considered important. We also assume that no trajectory passes through s or t a second time and that s and t are incident to the unbounded face of the arrangement of curves corresponding to the trajectories. We need the latter assumption to decide which trajectory is in the middle at the start and end. Note that we also assume a direction on the trajectories, namely from s to t, and for convenience we assume that m is odd. Finally, we assume that the curves do not touch or coincide with each other or themselves at any point, unless they cross, and no three trajectories pass through a common point. Several of these assumptions can be removed, but they make the description easier.
- 1.
The median trajectory is a polygonal curve from s to t.
- 2.
Any point on the median trajectory lies on some trajectory of the input.
- 3.
Any edge of the median trajectory has the same direction as the edge of the input trajectory on which it lies.
- 4.
For any point p on the median trajectory, the minimum number of distinct trajectories that p must cross to reach the unbounded face (including the one(s) on which p lies) is (m+1)/2.
Besides these requirements, a number of desirable properties of the median trajectory can be given: Its length, total angular change, and number of vertices should be about the same as in the input trajectories. Finally, the median trajectory should be robust with respect to outliers: if ten trajectories follow the same route but one or two are completely different, the presence of these two outliers should not influence the median trajectory much. In particular, in the presence of outlier trajectories, the last property should be restated with m defined as the number of non-outlying trajectories instead of the total number of trajectories.
Let \({\mathcal{A}}\) be the arrangement formed by the (paths of the) trajectories in T. It is composed of O(nm) line segments and therefore it may have complexity up to Θ((nm)^{2}). The median trajectory is a path that follows edges of this arrangement. In the immediate neighborhood of s, it is clear how the median trajectory leaves s: Since we assume that s is on the outer face, exactly one face incident to the start point is the unbounded face. Then we can order the m edges adjacent to s with the first and last edge adjacent to the outer face. Then the edge the median starts on is simply the ⌈m/2⌉-nd edge in the order.
2.1.1 Simple Definition
Inspired by the median level in an arrangement of lines, we can give a very simple definition of the median: It is the trajectory obtained after leaving s in the only possible way while satisfying the last property, and then switching the trajectory at every intersection point, following the next trajectory in the forward direction (see Fig. 1 (right)). Note that if the trajectories are x-monotone, then this definition gives the same result as the ⌈m/2⌉-level or median function given before. We refer to a median by this definition as a simple median.
Lemma 2.1
The simple median satisfies the four required properties for a median.
Proof
Properties 1, 2 and 3 are clearly true. Property 4 follows from Lemma 3.1, which we prove in the next section together with the proof of this property for the other median trajectory definition. □
2.1.2 Homotopy Definition
To obtain a more natural median trajectory definition, we identify parts of the plane with respect to which the median should behave the same as most of the input trajectories. In both examples of Fig. 2, we have a region in the plane that is a bounded face of the arrangement \({\mathcal{A}}\) that is relatively large. We propose placing poles in such large faces and require that the median goes in the same way around the poles as the input trajectories, using the concept of homotopy.
We make this more precise. Let T={τ _{1},…,τ _{ m }} be the input trajectories and let P={p _{1},…,p _{ h }} be a set of h poles which are assumed to not lie on any trajectory. Since the trajectories all go from s to t, we can use deformability of the trajectories into each other in the punctured plane [23]. Two trajectories τ _{ i } and τ _{ j } are homotopic if one can be deformed continuously into the other without passing over any pole, and while keeping s and t fixed. In Fig. 2 (right), τ _{1} and τ _{2} are homotopic to each other, while τ _{3} is not homotopic to τ _{1} or τ _{2}. Homotopy is an equivalence relation.
We first discuss how we define the median when all trajectories in T are homotopic with respect to P. We use a variation of the trajectory switching approach: follow the median over the correct edge at s. Assume we have followed the median and we are on an edge of trajectory τ _{ i }, and we encounter an intersection v with a trajectory τ _{ j }, with 1≤i,j≤m. If the median so far, concatenated with trajectory τ _{ j } from v until t, has the same homotopy type as the input trajectories, then we switch to τ _{ j }, otherwise we ignore the intersection and stay on τ _{ i }. This approach maintains the invariant that if we would simply stay on the current trajectory until t, the homotopy type of the median is correct. In fact, this method is equivalent to computing the simple median, but in a universal cover space based on the poles, where only the intersections with the correct homotopy remain; details about this are given in Sect. 3.2. A median by this definition is referred to as a homotopic median. In Fig. 2 (left) the dotted loop would be included in the homotopic median.
Lemma 2.2
The homotopic median satisfies the required properties for a median.
Proof
Properties 1, 2, and 3 are clearly true. Property 4 follows from Lemma 3.2, proved in the next section. □
The remaining question is how to place a set of poles P such that ideally, the trajectories in T are homotopic with respect to it. A number of different strategies for this are conceivable. We choose to use a simple approach that places a pole in a face of \(\mathcal {A}\) whenever it is larger than r, i.e., a disk of size r fits in the face, for some value of r to be determined later. This is motivated by the fact that large faces are more likely to be important. This choice gives no guarantee that the trajectories will be homotopic, however. To solve this we could either increase r or allow some of the trajectories to have a different homotopy, and instead compute the median for a subset T′⊂T, i.e., effectively treating the remaining trajectories as outliers.
We note that homotopy has been used recently to define the similarity of two curves. In an environment without obstacles, the Fréchet distance is a good measure for the dissimilarity of two curves, but when there are obstacles between the curves, an extension is needed. This has led to the definition of homotopic Fréchet distance [10], which also uses the punctured plane.
2.2 The Complexity of Median Trajectories
Since the arrangement \({\mathcal{A}}\) formed by the m trajectories has complexity O((nm)^{2}), the median trajectory cannot have more edges than that. We show that it is actually possible for the median to have complexity Ω((nm)^{2}) for both definitions. We show that this lower bound holds even for m non-self-intersecting trajectories.
We first give an example that shows that even a single trajectory can give rise to a median that has Θ(n ^{2}) complexity. Consider the simple definition or assume that r is so large that there are no poles. Then at every intersection point, we switch the trajectory.
Consider the resulting median trajectory. First, the trajectory zigzags using slightly more than half of the long vertical line segments, then it goes back and zigzags again using the other slightly less than half of the long vertical line segments. After having used all long vertical line segments, the trajectory zigzags from bottom to top using all long horizontal line segments, the bottommost one directed from right to left. The trajectory and the start of the median are indicated in the figure. It is easy to see that the median of one trajectory with n edges has complexity Ω(n ^{2}).
Essentially the same construction can be used to show that m trajectories with n edges each can give rise to a median of complexity Ω((nm)^{2}). We start with similar sets of long directed vertical and horizontal line segments, Ω(nm) of each type, see Fig. 3 (right). We use ⌈m/2⌉ trajectories to connect the long vertical line segments into trajectories. We use pairs of trajectories that together essentially look like the first half of the trajectory in the single-trajectory example. They are placed next to each other. We use the other ⌊m/2⌋ trajectories to connect the long horizontal line segments. Using suitable intersections between the trajectories to make the median proceed its structure of zigzagging up and zigzagging down alternatingly, we obtain the lower bound of Ω((nm)^{2}) for the median complexity. Notice that, in contrast to the single trajectory construction, the input trajectories do not self-intersect.
Finally, we also note that for x-monotone trajectories, using the median level lower bound for arrangements of lines immediately leads to an Ω(nmlogm) lower bound on the complexity of the median in this case [31].
3 Staying in the Middle
In this section, we prove that our definitions and algorithms give medians that “stay in the middle”, as required, that is, we prove Property 4. This completes the proofs of Lemmas 2.1 and 2.2. We will first consider the simple setting without poles, corresponding to Sect. 2.1.1; after that we will show how to deal with poles.
3.1 Simple Setting
The vertices of \(\mathcal {A}\) arising from intersections all have degree 4, having 2 incoming and 2 outgoing edges, with the two incoming edges being adjacent and the two outgoing edges as well. Vertex s has indegree 0 and outdegree m, and vertex t has indegree m and outdegree 0.
Let f be a face of \(\mathcal {A}\), and let p be a point in the interior of f. Consider a ray starting at p and going to infinity that does not go through any vertex of \(\mathcal {A}\) and is not tangent to any edges of \(\mathcal {A}\). Let l be the number of crossings of this ray with the edges of \(\mathcal {A}\) where the edge crosses the ray from right to left, and let r be the number of such crossings where it crosses from left to right. We define the order of f to be (r−l) mod m. Note that there can be m different orders. Rotating the ray continuously does not change the order. Thus, the order is unique no matter in which direction we shoot the ray. The outer face has order 0. Also note that there are m faces adjacent to s (or t), and that they all have different orders. We will say that the outgoing edge from s between the face of order 0 and the face of order 1 is the first edge, the edge between the faces of order 1 and 2 is the second edge, etc., until the m-th edge between the faces of order m−1 and order 0.
Now we can state formally what we want to show.
Lemma 3.1
Let E, s, t, and T be as described above. A path π leaving s on the k-th edge, switching in the forward direction at every intersection, will end at t arriving on the k-th edge. Furthermore, at any point on π, we have the property that leaving on the left side of π and without crossing π again, at least k−1 distinct trajectories from T must be crossed to reach the outer face, and leaving on the right side of π, at least m−k−1 distinct trajectories must be crossed.
Proof
A path that starts on the k-th edge is an edge between a face of order k−1 and a face of order k. Each vertex of \(\mathcal {A}\) other than s and t has one face between its two incoming edges and one face between its two outgoing edges, and these have the same order. Whenever we reach a vertex (intersection) of \(\mathcal {A}\), according to our procedure, we “switch to the other trajectory”. Suppose that the other trajectory is coming from the left. This means that we switch to the right, stay incident to the face of order k−1 we were incident to before, and switch to another face of order k. We will always stay between a face of order k−1 and one of order k, therefore we end on the k-th edge. Furthermore, since the order difference between adjacent faces is always 1, at least k or m−k trajectories must be crossed to get from a face of order k to the outer face (which has order 0). □
We can, in fact, define a new set T′ of m trajectories going from s to t to be the m paths that the simple switching algorithm produces when leaving s on its m different incident edges. We remove every vertex of \(\mathcal {A}\) (except for s and t) by connecting each incoming edge to its adjacent outgoing edge, joining two pairs of edges. We shortcut slightly in an ε-neighborhood of the vertex, merging one pair of faces. After this, the arrangement has been decomposed into m simple trajectories that go from s to t and possibly a number of isolated closed loops, none of which intersect each other. Figure 4 (right) shows an example.
Finally, observe that if s and t are incident to the outer face, then the ⌈m/2⌉-th face is what we would intuitively call the “middle”.
3.2 Homotopic Curves Amidst Poles
Let P={p _{1},…,p _{ h }} be a set of h poles in the plane; the poles act as point obstacles. We will consider the domain D=ℝ^{2}−P, which is a punctured plane, and see how to define medians of trajectories that live in D.
It is well-known that a closed curve in D has a homotopy class which can be described as a reduced sequence of elements of the generating set of its fundamental group and their inverses (see, e.g., [7]). We can choose as generating set the set of h counterclockwise curves around each of the poles in P rooted at some base point (not in P). In other words, each closed curve can be decomposed into a number of these cycles and their inverses.
Two trajectories from a starting point s to an endpoint t are homotopic if the closed curve obtained by gluing them together can be contracted to a point, or equivalently, is in the homotopy class 1. One of the consequences of this definition is that the points s and t play no special role. In particular, it does not make a difference if a trajectory makes an extra loop around one of them or not.
Now, let T be a set of m trajectories from s to t that are all homotopic to each other in D. We will show that similar properties as in the simple setting still hold for T. We follow the description in [9] and argue based on a universal cover U of D.
A universal cover is a topological space that locally looks like the original space but is simply-connected (see e.g. [30] for a formal description). In our setting, it can be described as follows: we cut up D along vertical rays starting at each point of P, resulting in vertical slabs. We now glue together copies of these vertical slabs at the boundary rays, in such a way that no cycle is formed. Thus, if one stands in a point of one of the copies and starts walking, and walks around one of the poles in P, one will end up in a different copy of D, even though the projection will be the same.
Lemma 3.2
Let s, t, T and P be as defined above. A path π starting on the k-th edge leaving from s and switching at every intersection with the correct homotopy type intersections will end on the k-th edge of t. Furthermore, at any point on π, we have the property that leaving on the left side of π and without crossing π again, at least k−1 distinct trajectories from T must be crossed to reach the outer face, and leaving on the right side of π, at least m−k−1 distinct trajectories must be crossed.
Proof
Let \(\mathcal {A}\) be the arrangement of the trajectories in D, and let \(\mathcal {B}\) be the arrangement of the same trajectories lifted into U. All vertices of \(\mathcal {B}\) are also vertices of \(\mathcal {A}\). On the other hand, a vertex v of \(\mathcal {A}\) of degree 4 is a vertex of \(\mathcal {B}\) of degree 4 only if the two partial trajectories from s to v involved in the crossing have the same homotopy in D (or equivalently, if their continuations from v to t have the same homotopy). The homotopy definition for a median switches only at a vertices that are present in \(\mathcal {B}\). Hence, the lemma follows from the same argumentation as used for Lemma 3.1 (but now in \(\mathcal {B}\)). □
As before, we can define a set T′ of the m trajectories that the strategy of switching at every intersection with the correct homotopy type produces. Figure 5 (right) shows an example. Note that the crossings among the trajectories in T′ correspond exactly to those vertices of \(\mathcal {A}\) that are not in \(\mathcal {B}\).
Also observe that, again, if s and t are incident to the outer face of the arrangement \(\mathcal {A}\) of the trajectories in D, then they are also incident to the outer face of \(\mathcal {B}\), so the “middle” is again what we intuitively expect it to be.
4 Algorithms to Compute a Median Trajectory
In this section we show that both simple medians and homotopic medians can be computed efficiently. The simple median can be computed in O((nm)^{2}) time in the worst case, while the homotopic median can be computed in O((nm)^{2+ϵ }) time in the worst case for any ϵ>0. This is (close to) quadratic in the input size. In practice, running times will be faster because they depend on complexities of intermediate results that should be much less than the worst-case situations. In particular, let A be the complexity of the arrangement formed by the nm edges of the trajectories. Although A=O((nm)^{2}) and this is tight in the worst case, for trajectories we typically expect it to be much smaller. Therefore, making the time bound depend on A instead of (nm)^{2} is desirable. Similarly, let h be the number of poles. Then h=O(A)=O((nm)^{2}), but only large enough faces give rise to a pole so h is typically much smaller than A. Finally, the complexity of the median itself, the output size k, is O(A)=O((nm)^{2}), but typically we expect it to be smaller. Notice that h or k can be large when the other is small. We will show that the simple median can be computed in an output-sensitive manner as well, and the homotopic median can be computed more efficiently when the number of poles and the output size are not very large. With a natural sampling assumption, we can remove the dependency on the number of poles.
4.1 Computing the Simple Median
A simple algorithm to compute the median with the simple definition is via the construction of the arrangement \({\mathcal{A}}\). The arrangement can be constructed in O(nmlog(nm)+A) time, where A is the complexity of the arrangement [21]. Then we simply follow the median trajectory through this arrangement, taking O(1) time at every intersection point or trajectory vertex. Every intersection point and trajectory vertex will be a vertex of the median we compute. Hence, this algorithm takes O(nmlog(nm)+A)=O((nm)^{2}) time.
The complexity A of the arrangement is not directly related to the complexity of the median. For instance, it can happen that A=Θ((nm)^{2}) even when the median never switches trajectories. For an output-sensitive algorithm, we use Har-Peled’s randomized algorithm for an on-line walk in a planar arrangement [22]. This algorithm allows one to compute the part of an arrangement that is found when “walking” on it. In other words, we can compute the part intersected by a curve not known in advance, with the walking direction potentially changing every time an edge is crossed. In our case, the walk is precisely the median trajectory we want to compute. The expected runtime of this algorithm is O((nm+I)α(nm+I)log(nm)) in an arrangement of nm line segments, where I is the number of intersections between the walk and the arrangement, and α denotes the inverse Ackermann function. In our case, I=k is the complexity of the median, because the median switches trajectories at every intersection. Furthermore, I=O((nm)^{2}) and then α(nm+I)=O(α(nm)).
Theorem 4.1
The simple median of m trajectories with n edges can be computed in O((nm)^{2}) time or in O((nm+k)α(nm)log(nm)) expected time, where k is the size of the output.
4.2 Computing the Homotopic Median
- 1.
Compute poles, one for each face in which a disk of radius r fits.
- 2.
Compute the homotopy type of each trajectory, and determine the type that occurs most often. Remove all trajectories that do not have this type.
- 3.
Follow the median from s: at every intersection, determine whether the continuation on the new trajectory yields the correct homotopy type (when the new trajectory is followed to t).
4.2.1 Step 1
Step 1 can be performed in O(nmlog(nm)+A) time by constructing the arrangement [21] and then computing the medial axis in each face in linear time [11]. The medial axis of a polygon P contains all points p such that there is a circle centered at p that lies inside P but touches P in at least two points. The vertices of the medial axis correspond to circles that touch P in at least three points. The number of vertices of P is O(|P|). Since there must be a largest circle that fits inside P that touches it in at least three points, we can find this by checking all vertices of the medial axis.
4.2.2 Step 2
Step 2 can be performed using an algorithm of Cabello et al. [9]. They showed that deciding whether two paths are homotopic takes \(O(n\sqrt{h}\log h)\) time, assuming the two paths have n edges and there are h poles. We show how to use this algorithm in our setting.
The algorithm of Cabello et al. [9] lifts the paths to a universal cover of the plane without the poles. A possible way to construct a universal cover is the following: Take a spanning tree of the poles, and extend the tree by a ray to the left from the left-most pole, and to the right from the right-most pole. This subdivides the plane into two parts B _{0} and B _{1}. Now glue together copies of B _{0} and B _{1} in such a way that if paths cross a different edge of the extended spanning tree, they move into different copies of B _{0} or B _{1}.
Two paths are homotopic in the original space if and only if they start and end in the same points when lifted to the universal cover. To check for homotopy we construct the portion of the universal cover in which the paths live. More specifically, we construct the corresponding part of the dual tree of the universal cover. In the dual tree we have a node for each copy of B _{0} and B _{1}, and an arc between two nodes if they are glued together, i.e., a path goes from one to the other through an edge of the extended spanning tree of the poles. We label this arc by the corresponding edge of the spanning tree. The dual tree is infinite, but we use only the relevant parts, depending on the trajectories.
Instead of two infinite rays, the algorithm in [9] uses a bounding box and connects the spanning tree to a point on the left and on the right boundary. Furthermore, the extended spanning tree is replaced by a simple path that traces the spanning tree using perturbation. We will nonetheless refer to it as the spanning tree of the poles. As a result, B _{0} and B _{1} are simple polygons which are preprocessed for efficient ray shooting. The algorithm uses a spanning tree on the h poles with stabbing number \(O(\sqrt{h})\) for any line, which can be constructed in O(h ^{1+ϵ }) time for any ϵ>0. The algorithm proceeds by concatenating two paths to a loop, traverses the loop to construct the part of the universal cover of the plane without the poles in which the loop lives, and checks whether the loop starts and ends at the same point in the universal cover (refer to [9, Lemma 11]).
For our algorithm, we need to check the homotopy types of more trajectories in order to find the largest homotopy class. We do not concatenate every pair of trajectories, but trace each from the same starting point in the dual tree and find the node where it ends. The node with the highest count gives the largest homotopic subset. Tracing one trajectory takes \(O(n \sqrt{h} \log{h} )\) time, where the \(O(\sqrt{h})\) factor comes from the fact that each line segment intersects \(O(\sqrt{h})\) spanning tree edges, and the O(logh) factor comes from ray shooting in B _{0} and B _{1} [24]. The total running time of Step 2 is \(O(mn \sqrt{h}\log{h} +h^{1+\epsilon})\) for any ϵ>0. With each node in the dual tree of the universal cover we also store the parts of trajectories in the corresponding part of the universal cover, which is needed for Step 3.
4.2.3 Step 3
Step 3 can be performed by explicitly computing the arrangement of the trajectories. We trace the median, switching between two intersecting trajectories only if this intersection also occurs in the universal cover, or equivalently, if the two trajectories are at the same node of the dual tree. Since the intersection lies in the same node of the dual tree, the parts before the intersection are homotopic and therefore the parts after the intersection are homotopic as well, since the full trajectories are homotopic. The running time of this step would then consist of O(nmlog(nm)+A) time to compute the arrangement and \(O(mn \sqrt{h}+A)\) time to trace the median. To obtain an output-sensitive algorithm we again use Har-Peled’s algorithm, but now we walk only in the arrangements of the trajectories restricted to the corresponding part of the universal cover. More specifically, we trace the lifted median in the universal cover using the dual tree. With each node of the tree we have stored the corresponding parts of the trajectories, and at every node of the tree we will use only these parts. While we trace the lifted median, it may return to the same node of the dual tree several times and therefore go through the same arrangement several times. But since we leave/enter the corresponding part of the universal cover through the same edge of the spanning tree of the poles, we can simply concatenate the walks without adding intersections. If N _{ i } denotes the number of line segments in the i-th node and k _{ i } denotes the complexity of the median restricted to the i-th node, the expected running time of the walk is O(∑_{ i }(N _{ i }+k _{ i })α(N _{ i }+k _{ i })log(N _{ i })) which is bounded by \(O((nm\sqrt{h}+k)\alpha (nm) \log (nm))\) since \(\sum_{i} N_{i} = O(nm\sqrt{h})\) and \(\sum_{i} k_{i} = k+O(nm\sqrt{h})\).
Theorem 4.2
The homotopic median of m trajectories with n edges can be computed in O((nm)^{2+ϵ }) time or in \(O((nm\sqrt{h}+k)\alpha (nm) \log (nm)+ h^{1+\epsilon}+A)\) expected time for any ϵ>0, where h is the number of poles, A is the arrangement size and k is the output size.
4.2.4 Sampling Assumption
It seems reasonable to assume that the size of faces that are relevant and receive a pole, represented by r, is not much less than the length s of the longest edge in any trajectory. Suppose for instance that r≤s/6. Then the trajectories are sampled so sparsely that it can happen that two trajectories have exactly the same edge of length 2r, whereas one traversed in that time unit a distance 6r using three sides of a square of side length 2r instead of one side; note that such a square contains a disk of radius r and therefore should contain a pole. However, we cannot know that the trajectories were very different, although the choice of r suggests that this is relevant. This makes the assumption r=Ω(s) reasonable; we refer to it as the sampling assumption. Under this assumption, we can improve the running times of our algorithms.
For Step 1, we determine all faces that receive a pole without constructing the arrangement. Let D be a disk of radius r centered at the origin. Take the Minkowski sum of every edge of every trajectory and D, and compute the union of these O(nm) “race tracks”. The complement of the union contains parts of all faces that are large enough, although the same face of \({\mathcal{A}}\) may appear as several faces in the complement of the union. Notice that homotopic equivalence is not influenced if any face of \({\mathcal{A}}\) has more than one pole. The sampling assumption implies that all race tracks are fat objects of similar size, and hence the union complexity is bounded by O(nm) [32]. We use the algorithm of Kedem et al. [25] to construct it in O(nmlog^{2}(nm)) time. This gives us a set of h=O(nm) poles.
For efficiency reasons in Steps 2 and 3, we must avoid having two poles closer than 2r. Two such poles would lie in the same face of \({\mathcal{A}}\), so we can remove either one. We determine a subset of the poles such that every big face of \({\mathcal{A}}\) has at least one pole in the subset, and any two poles in the subset are at least 2r apart. We do this by computing the Delaunay triangulation of the poles. We then remove all edges that have length more than 2r, and choose one pole per connected component in our subset. Let P be this subset of poles. The idea is to construct a spanning tree on P with stabbing number O(1) for line segments of length at most s.
We do this as follows: take a set of vertical lines that are exactly r apart. Within each vertical slab, connect the poles by y-coordinate. Across vertical slabs, consider every two consecutive non-empty slabs; there may be empty slabs in between. We connect the rightmost point of the left slab with the leftmost point of the right slab. Any segment of length at most s crosses O(1) slabs, and due to vertical spacing within a slab, it intersects O(1) spanning tree edges per slab.
Lemma 4.3
The spanning tree on P has O(1) stabbing number for line segments of length at most s, in particular, for the edges of the trajectories.
Now we again use the same algorithm as above, but replace the spanning tree with stabbing number \(O(\sqrt{h})\) by this spanning tree with O(1) stabbing number. The resulting expected running time for Step 3 is O((nm+k)α(nm)log(nm)).
Theorem 4.4
The homotopic median of m trajectories with n edges can be computed in O(nmlog^{2}(nm)+kα(nm)log(nm)) expected time, where k is the size of the output, if the sampling assumption is satisfied.
5 Experimental Results for Median Trajectories
In this section we present experimental results that aim at analyzing the quality of the medians generated by our two definitions. The experiments compare the definitions quantitatively, with respect to the desirable properties mentioned in Sect. 2—number of vertices, total length, and total turning angle—and also qualitatively, by analyzing visually in which cases one or the other method produces counterintuitive results.
5.1 Experimental Set-up
To test our two definitions and compare them, we implemented a random trajectory generator that generates sets of “similar trajectories”. For each of these sets of trajectories, the medians for both definitions were computed and analyzed.
5.2 Data Set
We generated many sets of 8 waypoints and then 9 trajectories for each of them. A set was accepted if at least 6 out of 9 trajectories were homotopic for a fixed value of r. For each accepted set, the medians according to both definitions were computed, giving the length, angular change, and number of vertices for both. We distinguished four types of waypoint sets, depending on two properties of the polygonal line implied by the sequence of 8 waypoints: we compare no self-intersections (1) to self-intersections (2), and low angular change (a) to high angular change (b) (angles below or above 4π). We repeatedly generated sets until we had 100 sets in each of the four classes. Note that the classes refer to the properties of the waypoints, not the trajectories themselves (trajectories may self-intersect even if the waypoint sequence does not).
5.3 Results
Average length, average angular change, and average number of vertices for simple (S) and homotopic (H) median trajectories over 100 tests. The average of these measures for the input trajectories was normalized to 1. The standard deviation is also shown
Length |
Angular change |
No. of vertices | ||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|
μ _{ S } |
σ _{ S } |
μ _{ H } |
σ _{ H } |
μ _{ S } |
σ _{ S } |
μ _{ H } |
σ _{ H } |
μ _{ S } |
σ _{ S } |
μ _{ H } |
σ _{ H } | |
1a |
0.961 |
0.17 |
0.995 |
0.02 |
5.743 |
1.02 |
5.052 |
0.88 |
3.660 |
0.71 |
3.364 |
0.53 |
1b |
0.940 |
0.19 |
0.995 |
0.03 |
6.058 |
1.33 |
5.397 |
0.84 |
3.684 |
0.82 |
3.477 |
0.54 |
2a |
0.506 |
0.26 |
0.955 |
0.11 |
4.107 |
2.26 |
4.992 |
0.84 |
2.139 |
1.14 |
3.625 |
0.63 |
2b |
0.493 |
0.24 |
0.923 |
0.13 |
4.213 |
2.13 |
4.912 |
1.19 |
2.118 |
0.99 |
3.610 |
0.84 |
Visual inspection showed that the simple median occasionally made “errors” (against intuition) even for inputs without self-intersections, and nearly always for inputs with self-intersections, see Fig. 6. The homotopic median nearly always gave intuitive results, although an occasional “error” could be observed, due to the absence of poles in small or narrow regions. In the bottom figure in Fig. 6, the homotopic median misses the loop at waypoints 5 and 6. In the top figure, the homotopic median is mostly correct but it makes extra zigzags, best visible near waypoints 6 and 7. This explains the high angular change of the median.
We also tested how the number of vertices of the median is influenced by the number of trajectories. There seems to be a linear dependence, suggesting that the output size k=Θ(mn), but this observation is highly dependent on our random trajectory generator. Hence we did not test this further.
In summary, except for the high angular change and occasional missing parts, the homotopic median appears to be a good definition for sets of trajectories generated by our generator, even for intersecting trajectories. It gives much better results than simple medians.
6 Discussion and Future Research
We discussed the fundamental—but up to now missing—concept of the median of a set of trajectories. We make a first step in this direction by proposing necessary and desirable conditions that, intuitively, a trajectory median should satisfy. Based on them, we presented two definitions of the path of a median trajectory of a set of trajectories, together with efficient methods to compute them. We also proved properties of the resulting medians and analyzed them experimentally.
Given the importance of the concept of a median trajectory and its novelty, we believe this paper opens up many venues of further research.
We made several restrictions in this paper that may be unrealistic. We assumed the start and end points of all trajectories to coincide and to lie in the unbounded face. We also assumed that most trajectories are similar enough; the homotopy method can deal with some trajectories that are outliers, but it cannot deal properly with the situation where parts of many trajectories are outliers. This can result in a situation where the largest homotopy class has only one trajectory, whereas an intuitively correct median may still exist. We also assumed the parameter r to be given; it would be desirable to choose it automatically in an efficient manner.
We have not addressed the question how to assign time stamps to the median or how to use the time stamps of the input to guide the computation. Of course it may be possible to define a median that has good properties in a completely different way, with or without using the time stamps. Finally, it would be interesting to test the definitions of medians for various types of real-world data, instead of generated data.
Acknowledgements
This research has been supported by the Netherlands Organisation for Scientific Research (NWO) under BRICKS/FOCUS grant number 642.065.503, under the project GOGO, and under project no. 639.022.707. M.B. is supported by the German Research Foundation (DFG) under grant number BU 2419/1-1. M.L. is further supported by the U.S. Office of Naval Research under grant N00014-08-1-1015. R.I.S. is also supported by the Netherlands Organisation for Scientific Research (NWO). C.W. is supported by the National Science Foundation grant NSF CCF-0643597.
Open Access
This article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source are credited.