KR20020095774A - A probabilistic approach to planning biped locomotion with prescibed motions - Google Patents

Abstract

PURPOSE: A method for probabilistic approach for a movement motion plan by using motion capture data is provided to find footprint sequence of a final motion and its corresponding motion sequence to generate a final motion by re-targeting the motion sequence. CONSTITUTION: A method for probabilistic approach for a movement motion plan by using motion capture data includes the steps of composing a road map for guiding movement motion of an imaginary character by randomly extracting points for the character having two feet to safely step in an imaginary environment to find out a series of motions for the character to move from a starting point to a target point, converting and searching the road map with reference to a posture transition graph which shows connected relations between motions captured in advance for obtaining target footprints and the motion sequence from the starting point to the target point, and deforming the input motion sequence for precisely following a target footprint sequence for the landing points of the input motion sequence to generate a final movement motion.

Description

A probabilistic approach to planning biped locomotion with prescibed motions}

The present invention relates to a probabilistic approach to mobile motion planning using motion capture data. More specifically, the present invention relates to a technique for utilizing pre-captured motion data to plan a motion in which a virtual character moves naturally from a starting point to a target point.

With the recent development of motion capture system technology and the widespread use of these devices, it is possible to easily obtain very realistic motions, and to produce realistic animations of 3D virtual characters appearing in movies and video games. Such motion data is also widely used. Accordingly, researches focused on manipulating, editing, and recycling motion data are actively underway. They provide an effective way of transforming an animator into a desired behavior by describing the characteristics that the behavior should satisfy. However, the study of the motion plan using motion data has not been presented yet.

An animation scenario consists of several tasks that can be performed in a series of actions. Task-level motion planning, which utilizes motion data, can provide very realistic motion in rapid motion prototyping, which allows you to quickly see animations early in the animation design phase. It can also effectively support an interactive animation system in which a user-controlled character interacts with a virtual environment. Task-level directives are required to easily create realistic motion in interactive animations, but traditional motion planning techniques are difficult to provide while efficiently utilizing pre-captured motion data.

The creation of realistic movement gestures has received a lot of attention in computer animation. Bruderlin and Calvert proposed a goal-directed dynamic approach that generates the desired walking behavior given the parameters of speed, stride length, and number of steps. Boulic et al. Used biomechanical data to exploit the intrinsic kinetics of gait. Ko and Badler also proposed a similar technique for generating dynamically-balanced gait, which then extended to walking along the curve and automatically generated footprints. Raibert and Hodgins have shown that they can create physically realistic movement behaviors with hand-crafted controllers, and Hodgins and others have applied them to more complex models such as athletes. Hodgins and Pollard also described how to automatically adapt existing controllers to new characters. van de Panne proposed an optimization-based technique for generating biped movement motions from a given footprint as an input, which has also been extended to the generation of quadruped walking motions.

In recent years, commercial animation software has also added packages for creating biped movements. For example, Boston Dynamics' "BDI Guy", CredoInteractive's Life Forms Studio ", 3D Studio Max's" Character Studio ", and Motion Factory's Motivate. They automatically create movements that move along when the user specifies the path of movement of the virtual character or the footprints to step on.

Models of characters similar to those in computer animation have about 40 high degrees of freedom (DOF), making it difficult to plan their movements immediately. Most studies to support human navigation have reduced the configuration space of the motion plan to two dimensions by simplifying the character model with a cylinder, and then applying the 2D path planning algorithm to determine the moving path in the virtual environment. Based on the existing two-group movement motion generation technique, we created a movement motion that follows it. Reynolds introduced a reactive behavior to simulate simple groups of creatures such as swarms, herds, and fish. Terzopoulos et al. Simulated the repulsive behavior of artificial meat with synthetic vision. Noser et al. Proposed a local navigation model for human-like characters using virtual vision. We also simulated a human memory model to avoid obstacles. Kuffner and Latombe solved a similar problem in a dynamically changing environment. Reich et al. Presented a real-time model for human navigation in rugged terrain and employs a simulated sensor to detect geometrical elements of obstacles. Bandi and Thalmann divided the virtual environment into a three-dimensional uniform grid to explore the paths of autonomous characters.

Given the starting point and the target point, the research to find the path of the robot moving between them has been actively studied in the field of robotics. Recently, probabilistic path planning has emerged as a study for quickly finding a path of a robot having a high degree of freedom. Barraquand and Latombe devised a randomized path planning technique to escape the local minima in the potential field method. Later, Kavraki and Latombe, Overmars and Svestka randomly sample the construction space of the motion plan in the preprocessing phase to generate a roadmap, and in the path planning phase, navigate the roadmap to find the robot's movement path. Were individually and jointly proposed. These methods have shown good performance in the very difficult problems of moving cars with non-holonomic constraints or robots with very high degrees of freedom.

In recent years, a limited theoretical basis has been laid for this experimental success. Koga et al. Combine random path planning and inverse kinematics to create animations that automatically manipulate human arms. Kalisiak and van de Panne proposed a grasp-based motion planning algorithm in a limited two-dimensional environment given a handle or foothold. Kindel et al. Developed a path planner for robots with dynamic constraints and demonstrated their effectiveness in real and simulated environments.

As motion capturing equipment is widely used, researches for generating realistic animations with ease are being actively conducted, and tools for editing pre-captured motions are being developed. Bruderlin and Willams employ signal processing techniques to manipulate motion, and use displacement mapping to modify motion while maintaining detailed characteristics of motion data. Witkin and Popovic proposed motion warping for this purpose. Unuma et al. Interpolated and extrapolated motion data in the frequency domain using Fourier analysis. Lamouret and van de Panne discussed several issues in reusing motion data.

Rose et al. Seamlessly connect behaviors using spacetime constraints. Gleicher simplifies the space-time constraint problem for motion retargeting, which allows existing characters of one character to be suitable for characters of the same structure but different sizes, resulting in the ability to edit motion interactively. there was. To shorten the execution time in this approach, Lee and Shin proposed a hierarchical displacement mapping technique based on multilevel B-spline approximation.

Given a starting point and a target point in the virtual environment, the problem to be solved in the present invention is to find a series of motions to move the virtual character from the starting point to the target point. In this problem, existing motion planning techniques are almost impossible to obtain realistic animations of human-like characters. In addition, existing motion editing techniques do not have a high-level planning capability to produce the desired motion.

Accordingly, an object of the present invention is to provide a technique that utilizes previously captured motion data to plan a motion in which a virtual character moves naturally from a starting point to a target point.

In order to achieve the object of the present invention described above, the proposed technique combines probabilistic path planning and hierarchical motion fitting to obtain a footprint sequence of the final operation and its corresponding. After finding the motion sequence to perform a motion retargeting (motion retargeting) to generate a final motion to move along the footprint.

According to the present invention, given a starting point and a target point in the virtual environment, randomly picks points at which the biped character can safely step on the virtual environment in order to find a series of movements for moving from the starting point to the target point using predetermined motions. Extracting and constructing a road map for guiding a movement of the virtual character; Converting and searching a road map with reference to a posture transition graph representing a connection relationship between an input operation sequence for moving from the starting point to a target point and pre-captured operations to obtain target footprints; And generating a final movement motion by modifying the input motion sequence so that the landing points of the input motion sequence exactly follow the target footprint sequence. Presented.

FIG. 1 is a schematic diagram illustrating a process of checking whether the first and last footprints can be converted to coincide with the footsteps of two nodes by modifying an operation within a certain limit according to the present invention.

2A to 2C are schematic diagrams showing a road map converted using a posture transfer graph according to the present invention.

3 is a schematic diagram showing a sequence of footprints refined continuously according to the present invention.

Figure 4 is a schematic diagram showing the refining process of the footprint in accordance with the present invention.

5 is a schematic diagram showing a process of predicting the trunk path according to the present invention.

6 is a schematic diagram showing a process of a planned movement operation in the terrain consisting of the sea and the island according to the present invention.

7 is a schematic diagram illustrating that various operations can be used to plan a movement operation in a complex environment according to the present invention.

8 is a schematic diagram showing a movement operation in a room with several households considered as obstacles according to the present invention.

In the present invention, in order to quickly generate realistic motions of a virtual character through high-level directives, a roadmap is constructed by combining probabilistic path planning and hierarchical motion fitting. We propose a mobile motion planning and generation technique that consists of three steps: roadmap construction, roadmap search, and motion generation. The following briefly describes each of these techniques in order to illustrate the overall nature of this technique.

First, the roadmap construction step randomly samples the valid configuration of a virtual character to construct a roadmap in a given virtual environment. At this time, for the efficiency of execution time, instead of the posture of the character (posture), the configuration of the foot (stance foot) as the configuration of the character is called a footboard (foothold). The road map for generating the movement movement of the virtual character is represented by a directed graph having nodes and positions of the support feet extracted as a sample. If a virtual character can move between two nodes within a range that does not lose its liveness using any given motion, a directed edge connecting the two nodes is inserted.

In the roadmap search step, the starting point and the target point are connected to the roadmap, and then the minimum-cost path having the constraint is found in the directed graph. That is, to find the least cost path where all two consecutive edges share one node of a posture transition graph. The nodes in the posture transition graph represent postures, and the two nodes are connected by directed edges representing motion. In order to ensure connectivity between the motions, the roadmap is transformed by referring to the posture transition graph, and then an input motion sequence and a target footprint sequence are obtained by applying a minimum cost path algorithm.

In the motion generation step, a realistic final movement motion is generated from the input motion sequence and the target footprint sequence found by searching the road map. The input motion sequence was obtained by simply concatenating the motions corresponding to each edge of the least cost path in the roadmap search phase. Thus, target footprints modified to match the footsteps of each node in the least cost path may be different than the footprint of the input operation sequence. This problem is solved in two steps. First, predict the final action to get a better initial guess. Then, using this initial estimate, we apply a hierarchical displacement mapping technique to create the final motion following the target footprint.

Then, in the present invention, 1. Roadmap configuration (node creation and connection, local planner, cost function, roadmap enhancement); 2. Roadmap search (pathfinding, refining footprints); 3. motion generation; 4. Experimental results will be described in detail with reference to the accompanying drawings in sequence.

1. We will look at the process of node creation and connection in constructing road map.

Foothold f = (p, q) denotes the configuration of the support foot, Wow Respectively indicate position and direction. The two-dimensional position (u, v) and direction of the support foot on the terrain By lowering the configuration space of the scaffold, the dimension of the configuration space is lowered. If is given, the height, pitch, and roll may be extracted from the geometric information of the terrain surface to restore the entire configuration f = (p, q). These parameters u, v, The nodes of the roadmap are generated by random sampling of the scaffolding, assuming that follows a uniform distribution within a defined interval of each parameter. For each extracted scaffold, make sure that the character can safely rest on the terrain. In this way, a predetermined number of valid scaffolds are extracted and stored as nodes in the roadmap.

The newly created nodes are added to the roadmap through a fast local planner, which will be described later. As the number of nodes in the roadmap increases, the time it takes to connect a newly created node with all other nodes created so far is enormous. Taking into account the fact that distant nodes rarely connect with each other, they try to connect only to the K nodes closest to the newly created node. Where K is a positive integer given in advance. Whenever a local planner can successfully connect between two nodes, add a new directed edge connecting the two nodes to the roadmap. The K-nearest neighbor problem has been actively studied in computational geometry, and finding it among a large number of points requires quite a bit of computation time.

The present invention utilizes a spatial partitioning technique, such as uniform cells, to quickly solve this problem in an easier way. For each cell, any node belonging to that cell is selected to represent all other nodes belonging to that cell. When a new node is created, the distance from the representative node of each cell is calculated to represent the distance from all nodes of each cell. By selecting K nodes in the order away from the nearest cell, the nearest K nodes are newly found from the newly created node.

Next, we will look at the local planner of the roadmap.

The local planner checks which of the two nodes in the roadmap can be connected using one operation. In other words, the behavior is modified to the extent that the first and last footprints can be converted to match the footsteps of the two nodes, respectively. This is as shown in FIG. If successful, the local planner gives a footprint sequence that connects the two nodes. If either footprint of this footprint sequence cannot be safely placed on the terrain, i.e. it is not a valid footprint, the two nodes are not connected. If all footprints are valid and the deformed action does not cause an collision with an obstacle, connect the two nodes.

Now let's explain how to convert footprints. Given an action M, the moment when the foot lands and rises interactively describes the footprint sequence f (M). The support foot is fixed on the ground from the moment the foot lands to the time the foot is lifted up, creating a footprint. Suppose M consists of n frames and m footprints. j th support foot If we were on the ground for the time interval of f (M) is written as

here, Represents the j th footprint, And Each represents the location and direction of the footprint, Is between the moment the foot lands and the moment the foot rises, Is the duration the foot is in the floor. For all j ego to be.

Suppose we want to connect two nodes in the roadmap with action M. Wow Let be called the scaffold configuration of each node. To modify the footprint sequence f (M), first adjust the position of the footprint and then adjust the direction. This is as shown in FIG. 1 is a schematic diagram showing a process of checking whether the motion can be transformed to match the footsteps of the two nodes by modifying the motion within the limits.

To adjust the position difference, first for all j The position of the first footprint by moving it and To match. Now, the location of the last footprint and For all j to be as close as possible Is rotated about an axis perpendicular to the surface, then the direction vector Rotate around the axis orthogonal to. Where the two axes of rotation Goes through. Footprint direction for later adjustment These rotations are also applied. Generally, this position shift and rotation and Cannot match exactly. Just both Wow It just lies on a straight line connecting. For all j To

By scaling at the rate and Can be matched. This also scales movement motion parameters such as stride length, movement speed, and turning angle.

Now let's adjust the difference in direction. Logarithmic function of unit quaternion algebra And exponential functions Difference between directions at two nodes using Wow Propagates to the intermediate footprints of the first and last footprint. new direction of the j'th footprint Of the footprints to gain The linear difference is linearly interpolated using the chord-length parameterization. That is, interpolation parameters for all rules Linear interpolation using

To maintain the quality of operation M, we need to check whether the last two non-rigid transformations are within a given threshold. About position difference in Percentage to measure how long or shorter cord lengths are Use Where s is given in equation (2). For the direction difference, the tight upper bound of the direction difference in linear interpolation given by Equation 3 is used. In other words, Use If the difference in position and direction is within each threshold, the transformation connecting the two nodes is successful.

Next, the cost function of constructing the road map will be described.

Each directed edge of the roadmap is charged with the amount of work required for the virtual character to move from its head node to its tail node in motion of that edge, which costs the route map by traversing the roadmap. It helps guide you in finding. If the cost function is well-designed, we can find a minimum-cost path that goes from the starting node to the target node in a series of actions intended by the user. The present invention uses a mixture of factors that measure the cost of the operation of each edge in various aspects.

One of the most intuitive and important factors that can be included in the cost function is the distance the virtual character should move. The distance to move when using that gesture by adding the distance between all adjacent footprints of that gesture. Get In other words, If is the position of the j's footprint, to be. Similar to, slightly different number of frames of operation You can think of Is a measure of distance in space Is a time scale. Thus, the weights for both If is given, we mix them into one cost function as follows:

A local planner was applied to transform any given action to fit the footings of two nodes. The local planner adjusted the difference in position and direction of each footstep with the first and last footsteps of the movement. This difference is indicative of how the motion has to be modified to satisfy the constraint that the footing and footprint must match. It is necessary to minimize the degree of deformation while satisfying these constraints in order to maintain the liveness of a given original operation. Wow In terms of the difference between the position and the direction, the present invention uses the weighted sum of the two as follows to measure the degree of deformation.

here, Is the weight according to the metric difference of position and direction.

In addition to the cost of distance and deformation, the user's preference for a particular action can be considered. For example, in general, a walking operation is preferable to a broad jump operation. Preferred cost inversely proportional to user's preference to account for this preference Is included in the cost function term. Finally, the cost function for the edge is defined by multiplying the last cost by the weighted sum of the two costs mentioned above.

here Wow Is a weight adjustable by the user. In addition to the factors mentioned so far, factors such as distance to obstacles near the edge may be included in the cost function.

Next, the roadmap enhancement of the roadmap configuration will be described.

If enough nodes are extracted, the roadmap covers most of the configuration space evenly. However, when the proper number of nodes are evenly extracted, the roadmap is not well connected in difficult parts of the construction space such as a narrow passage. In the present invention, a heuristic method is adopted to encourage better road maps. The probability density function for node v belonging to node set V of the road map is defined as follows.

here Is the number of edges coming into v Is the number of edges leaving v. This probability density function stems from the fundamental idea that nodes that are not well connected will lie in difficult parts of the construction space with high probability. Therefore, we select node v from V according to probability density function P (v | V) and add new nodes around v so that the roadmap is better connected in the difficult part of the construction space. It is experimentally known that the number of additional nodes shows the best performance when it is between 1/3 and 1/2 of the number of nodes created according to the even distribution.

Next, the road map search will be described.

Here, we describe how to find the least cost path in the roadmap that gives an input motion sequence and a target footprint sequence given a starting point and a target point. In addition, the refinement of the target footprint using the local planner will be described.

Before explaining roadmap navigation, let's take a closer look at the posture transition graph, which plays a major role in ensuring connectivity between motions. The nodes in the posture transition graph represent valid postures of the virtual character and the edges between the nodes represent motion. If the operation of one edge is connectable to the operation of the other edge, the head node of the front edge and the tail node of the rear edge are the same. This is as shown in FIG. 2A to 2C are schematic diagrams showing a road map converted by using the posture transition graph, wherein 2a is a posture transition graph, 2b is a road map, and 2c is an enlarged road map.

Each edge is also accompanied by a footprint sequence obtained from the motion corresponding to that edge. To transform the motions of the posture transition graph into a form that can be easily transitioned, techniques for generating transition motions between motions can be used.

next, The path finding of the road map search will be described.

Scaffold indicating starting point and target point Wow Given a path search, it finds the input sequence of operations and the target footprint that connects them, which is processed in two steps. first, Wow Add nodes x and y to the roadmap. We then search the roadmap to find the least cost path between x and y. If either node is not connected to the roadmap, it randomly walks from that node and connects to the roadmap.

Careful attention is required in the roadmap search because not all of the predefined actions can be linked together. Suppose we have reached v through the edge e adjacent to node v while searching the roadmap. To move forward in v, the edge is adjacent to v and has an action that can be linked to that of edge e. Should go through The node v visited during the path search must remember which of the edges adjacent to v has reached v. Also, referring to the posture transition graph, the motion of e You should check if it can be connected by the operation of. Therefore, the least cost path algorithm cannot be directly applied to the roadmap search.

The directed graph representing the road map is called G (V, E). In order to remember which edge visited the current node while looking for a path and avoid referencing a posture transition graph, the present invention draws a new directed graph G (V ', E') from G (V, E). Take it and give it a sequence of actions that can be linked to any path on G '. K nodes representing different postures Suppose it is made of. Then, using the nodes in the attitude transition graph, set all nodes v of G (V, E) to k nodes belonging to G '. Split into Each node Is a node in the posture transition graph In addition to the posture at, we have a scaffold at node v of roadmap G (see FIG. 2). In G and The behavior of the edges connecting the posture transition graph Wow If the behavior of the edges connecting Wow Insert a directed edge connecting G 'to G'. Finally, remove the nodes in G 'that are not connected to any other nodes.

Since the node of G 'is associated with the footsteps on the terrain as well as the virtual character's pose, any pair of edges adjacent to any node of G' is the edge whose motion starts at that node. Ensures that it can transition to the operation of. Thus, any path of G 'gives a linkable sequence of operations. As a result, the least cost path algorithm can be directly applied to the graph G '. Since G 'has up to k | V | nodes and | E | edges, the time complexity of finding the target footprint and the sequence of motion to move from the starting point to the target point is to be.

If path planning fails frequently, the roadmap can be gradually extended by extracting the configuration space more densely and adding new nodes. If only part of the terrain has changed, you can gradually update the roadmap by deleting only the nodes that were in that part of the terrain before the change, and extracting new nodes only in that part of the terrain after the change.

Subsequently, the step refinement of the roadmap search will be described.

The target footprint sequence is obtained at the edge of the path found in the roadmap, and this target footprint sequence may be within a reasonable range from the original footprint sequence of each operation as described above. In addition, this difference may be different for each edge. The present invention proposes a local refining scheme in which each footprint of the target footprint sequence is effectively adjusted in turn with reference to the footprints of the original operation to smooth these differences in the edges of the path. By applying local refinement forward and backward repeatedly for all the footprints of the target footprint sequence, the difference between the footprint sequence of each edge of the path and the footprint sequence of the original motion at that edge propagates more evenly along the path. This is as shown in FIG. FIG. 3 is a schematic diagram showing a sequence of continuously refined footprints. The number of refining is 0 (left), 4 (center), and 8 (right).

The present invention uses the above local planner to locally refine the footprint sequence. For each footprint f in the target footprint sequence, the corresponding footprint in the original operation. Take footprints adjacent to them. This is as shown in FIG. Figure 4 shows the process of refining the footprint. Local planner Hypothetical nodes in the roadmap are plotted with two footprints adjacent to footprint f to apply to them and their adjacent footprints. and Regarded as a springboard. Apply a local planner to a sequence of three footprints, including two adjacent footprints and Adjust to the footsteps of to get a new footprint f '.

In the present invention, instead of simply replacing the original footstep f with the newly obtained footstep f ', the two footprints are linearly interpolated using a predefined weight. If the footprint obtained through linear interpolation is not valid, the weight is lowered to obtain a new footprint closer to the original footprint f. You can repeat this process several times to get a valid footprint, and if you don't get a valid footprint after a given number of attempts, use f as is.

Next, the operation generation will be described.

In the present invention, a virtual character moves from a starting point to a target point from an input motion sequence and a target footprint sequence obtained by searching a road map. Considering each footprint of the target footprint sequence as a variational constraint over time, the problem to be solved in the present invention may be formulated as motion retargeting. Initial body trajectory is known to be very important in this problem for numerical optimization to converge quickly and produce good results. Trunk trajectories are usually expressed in terms of the position and direction of the pelvis over time. By analyzing the input motion sequence and the target footprint and predicting the torso trajectory of the target motion, an initial estimate of the target motion in each frame with each joint angle of the input motion sequence can be obtained. This initial estimate is used in conjunction with a hierarchical motion fitting technique to retarget the input motion sequence to the footprint.

For motion retargeting, we will only describe how to derive the initial estimate by predicting the body trajectory of the target motion. Let f (A) and b (A) represent the footstep sequence and body trajectory for the input motion sequence A, respectively. Let f (B) and b (B) be defined similarly for the target operation sequence B which is not known yet. The purpose of motion generation then is to predict the target trunk trajectory b (B). At this time, B is implicitly designated by f (B) and A. Suppose A and f (A) consist of n frames and m footprints, respectively. By tracking the position and orientation of the pelvis in each motion of the motion sequence, it is easy to get b (A) in every frame of A.

To predict b (B), we use the relationship between b (A) and f (A). This is as shown in FIG. 5 shows a prediction process of the trunk path. First, a reference trajectory r (A) is derived from the footprint f (A) of the operation sequence A, and then a displacement d (A) with respect to b (A) is obtained.

Where r (A) and the operator Will be defined later. Note that b (A), d (A), and r (A) are all composed of components of position and direction. Assuming that A and B are partially very similar, we predict b (B) as follows:

To determine b (B), the reference trajectories r (A) and r (B) are needed in addition to b (A) derived directly from the original operation. Given this, the center of the body (the center of the pelvis) is expected to lie above the mid-posture of successive pairs of footprints. In Wow Let's assume that the middle pose of

For all adjacent pairs of footprints in f (A), this intermediate pose is piecewise linearly interpolated Obtain a continuous reference trajectory of A from Resample the reference trajectory on every frame of Get Similarly, r (B) can be obtained from the target footprint sequence f (B).

Now displacement map I will describe how to obtain. Ordered pair of position and direction components of r (A) silver Points u belonging to Specifies the rigid body transform to point u 'belonging to. In other words, to be. Where vector Is a purely imaginary unit quaternion Is considered. | Wow If is given, the displacement maps of the two measured in the local coordinate of r (A) Is defined as

Finally, In this case, the body trajectory b (B) is obtained as follows.

b (B) is Wow Only defined between frames. Interval to extend b (B) to all frames Taking part of b (A) in We transform the rigid body to match the position and direction at and connect it with b (B). section B (B) in Eq.

Next, the results of the experiment will be described.

The present invention has been developed as an Alias | Wavefront MAYA plug-in for Microsoft Windows 2000 and implemented in C ++. All experiments were performed using Viewpoint motion data on a personal computer with an Intel Pentium III 800 MHz processor and 512 MB memory. The human model used in the experiment has 43 degrees of freedom. Six degrees of freedom for the position and direction of the pelvis, three degrees of freedom for the spine, seven degrees of freedom for each limb, and three degrees of freedom for the neck and head, respectively. All motion data was sampled 24 times a second.

The first experiment is to plan the walking behavior of a human-like character. 6 shows the planned movement behavior in a terrain consisting of sea and islands. Here, the virtual character cannot put his feet on the sea. In this experiment, the motion data for walking and the posture transfer graph described in FIG. 2A were used. The terrain on the sea is modeled as a NURBS surface that shakes the y-coordinates of the control points on a regular grid up and down above sea level. Only footprints that are positive in height by setting sea level to zero are valid. In order to avoid obstacles, heuristic method was used to check the collision by checking whether the swinging foot is above the terrain in every frame.

The second experiment shows that various movements can be used to plan movement behaviors in very complex environments. The terrain of FIG. 7 has very complex features such as canyons, rivers, and the like. Actions for broad jump, running, etc. have been added to the motion repertoire, and if the height difference is small in local terrain, the jumping motion is preferred over the walking motion. Experimental results show that the techniques presented in the present invention can be suitably used in moving through complex terrain.

6 7 8 Number of actions 13 29 13 Roadmap Configuration N (initial number of nodes) 2000 3000 2000 M (number of hardened nodes) 1000 1500 1000 K (number of connection candidates) 100 100 100 E (number of edges connected) 88703 266419 87617 E / (N + M) 28.81 59.29 29.21 Time 34.250 99.050 38.880 Roadmap navigation Find route time 0.020 0.250 0.060 Move action count 5 28 14 Footprint Refining Repeat Count 4 4 4 Footprint Refining Time 0.020 0.120 0.080 Create action Total frames 197 1139 467 Initial Torso Prediction Time 0.010 0.020 0.010 Motion re aim time 0.090 0.670 0.250 Motion re aim time per frame 0.001 0.001 0.001 Total time excluding preprocessing 0.140 1.060 0.400 Time Spent Per Frame Excluding Preprocessing 0.001 0.001 0.001

In this case, N and M are the number of nodes initially extracted and the number of nodes generated through roadmap enhancement. E is the number of edges connected. The time data indicates how many seconds the CPU has been used. Table 1 summarizes the input data and the experimental results. In each experiment, the route planning of the present invention is largely carried out in three stages: roadmap construction, roadmap search, and motion generation. It can be seen that the execution time of the roadmap search is proportional to the number of nodes and edges of the roadmap, and as previously foreseen, the roadmap construction time overwhelms the entire time in each experiment. The roadmap construction time depends on the complexity of the motion data and the terrain, and the motion generation time is proportional to the number of frames of the generated motion. In all experiments, once the roadmap was created, we were able to produce approximately 1000 frames or more of motion per second. Since the roadmap configuration is regarded as a preprocessing step, the proposed motion planning technique shows the ability to interactively produce motion in the experiment of the present invention.

As described above, according to the probabilistic approach to the mobile motion plan using motion capture data according to the present invention, the following effects are obtained.

In the present invention, in order to effectively support task-level motion generation in rapid motion prototyping or interactive application, a bipedal character such as a person is a target point at a starting point. Until now, we propose a new technique for planning movements using a given set of motions. That is, by combining probabilistic path planning and hierarchical motion alignment, the motion sequence from the starting point to the target point and the target footprint sequence are found, and then the motion is retargeted to follow the target footprint sequence.

Thus, given various movements, the movement planning scheme of the present invention can move rugged terrain with various movements, such as a character such as a person running in the plains, crossing a leg, and jumping over a canyon. This has the effect of creating realistic animations very quickly and efficiently.

Claims (5)

Given a starting point and a target point in the virtual environment, the virtual character is randomly extracted from the points where the biped character can safely step on the foot in the virtual environment to find a series of movements for moving from the starting point to the target point using predetermined motions. Constructing a road map for guiding a movement operation of the vehicle; Converting and searching a road map with reference to a posture transition graph representing a connection relationship between an input operation sequence for moving from the starting point to a target point and pre-captured operations to obtain target footprints; And modifying the input motion sequence to generate a final motion motion such that the landing points of the input motion sequence exactly follow a target footprint sequence. The method of claim 1, wherein in the constructing the roadmap, a sample is generated by randomly sampling a scaffold of a support foot, and the generated nodes are added to the roadmap through a local planner. Probabilistic approach to action planning. The method of claim 1 or 2, wherein pre-captured motions are connected to all edges of the roadmap. The method of claim 1, wherein in the step of transforming and searching the roadmap, local refining is repeatedly applied back and forth to all the footprints of the target footprint sequence, and the difference between the footprint sequence and the footprint sequence of the original operation is equalized along the path. A probabilistic approach to mobile motion planning using motion capture data, characterized in that it propagates automatically. 2. The method of claim 1, wherein the generating of the final movement by modifying the input motion sequence comprises predicting a body trajectory of the target motion according to the analysis of the input motion sequence and the target footprint to obtain an initial estimate of the target motion. Probabilistic approach to mobile motion planning using motion capture data.