JP7456590B2 - Tracking devices and programs - Google Patents

Description

The present invention relates to a tracking device and a tracking program, and for example, to one that tracks pedestrians.

In recent years, there has been active development of robots for use in the living environment, such as hotel guide robots and cleaning robots. These robots are particularly expected to be useful in resolving labor shortages due to population decline in the future, and providing lifestyle support, such as in commercial facilities, factories, and nursing care businesses.
In order to operate within a human living environment, it is necessary to understand the surrounding environment, such as the person to be tracked and the obstacles to be avoided.
An example of such technology is "Autonomous mobile robot, autonomous mobile robot control method, and control program" in Patent Document 1.
This technology not only predicts the destination of a person being tracked, but also predicts the destination of obstacles that obstruct the view of the camera that is photographing the person. The field of view of the camera is changed so that the image becomes larger.

By the way, when a robot recognizes and tracks a walking person in this way, the human frequently changes direction and speed whimsically within close range of the robot. was an issue.

Japanese Patent Application Publication No. 2018-147337

The present invention aims to robustly track objects.

(1) The present invention includes: a particle generating means that generates particles to be used in a particle filter in a three-dimensional space based on a probability distribution of a position where an object exists; a photographing means that photographs the object; a mapping means for mapping the generated particles; an image recognition means for recognizing the photographed object as an image by setting a detection area based on the position of the mapped particle in the image; a likelihood acquisition unit that acquires the likelihood of the generated particles based on the result; a tracking unit that updates the probability distribution based on the acquired likelihood to track the location where the target exists; The tracking device is characterized in that the particle generating means sequentially generates particles based on the updated probability distribution along a plane parallel to the plane in which the object moves .
(2) The present invention also includes a particle generation function that generates particles to be used in a particle filter in a three-dimensional space based on the probability distribution of the position where the target exists, a photographing function that photographs the target, and the photographed image. a mapping function that maps the generated particles to the image; an image recognition function that recognizes the photographed object by setting a detection area based on the position of the mapped particle in the image; and an image recognition function that recognizes the photographed object as an image. a likelihood acquisition function that acquires the likelihood of the generated particles based on the result of the above; and a tracking means that updates the probability distribution based on the acquired likelihood to track the location where the target exists. is realized by a computer, and the particle generation function provides a tracking program that sequentially generates particles based on the updated probability distribution along a plane parallel to the plane in which the object moves .

According to the present invention, the target can be robustly tracked by generating particles in a three-dimensional space where the target exists and updating the probability distribution of the position of the target.

FIG. 2 is a diagram showing an example of the appearance of a tracking robot according to the first embodiment. FIG. 2 is a diagram showing the hardware configuration of a tracking device. FIG. 2 is a diagram for explaining a virtual camera that takes stereo images. FIG. 3 is a diagram for explaining a method of measuring distance and direction to a target. FIG. 2 is a diagram for explaining the superiority of the congestion stereo system. FIG. 1 is a diagram for explaining a method for generating particles. FIG. 3 is a diagram for explaining mapping of particles to a camera image. FIG. 3 is a diagram for explaining a method of tracking the position of a subject with a virtual camera. FIG. 3 is a diagram for explaining a likelihood calculation method. It is a flowchart for explaining tracking processing. FIG. 7 is a diagram illustrating an example of the appearance of a tracking robot according to a second embodiment. It is a figure for demonstrating the surveying method in 2nd Embodiment.

(1) Overview of Embodiment The tracking device 1 (FIG. 2) includes omnidirectional cameras 9a and 9b placed on the left and right sides of the tracking robot.
The tracking device 1 pastes the left omnidirectional camera image taken by the omnidirectional camera 9a onto the spherical object 30a (FIG. 3(a)), and places a virtual camera 31a inside the spherical object 30a (FIG. 3(a)). establish.
The virtual camera 31a can freely rotate in a virtual photographing space formed inside the spherical object 30a and can acquire a left camera image of the outside world.
The tracking device 1 also includes a virtual camera 31b that acquires a right camera image from a right omnidirectional camera image taken by the omnidirectional camera 9b, and the virtual cameras 31a and 31b constitute a vergence stereo camera.

The tracking device 1 uses the vergence stereo camera configured as described above to track the position of the subject 8 using a particle filter.
The tracking device 1 generates particles three-dimensionally in the space where the target person 8 exists, but since the target person 8 is assumed to be a pedestrian and moves parallel to the walking surface, the torso of the target person 8 A large number of particles are generated around a circular area 32 centered on the subject 8 on a plane parallel to the walking surface at a certain height.

The tracking device 1 then acquires left and right camera images using the virtual cameras 31a and 31b, and maps particles generated in the real space in which the subject 8 is walking in correspondence with the left and right camera images, respectively.
That is, the generated particles are projected onto the left and right camera images, respectively, and the particles mapped onto the left and right camera images are matched together so that they can be identified as being the same particle in three-dimensional space.

Next, the tracking device 1 sets a detection area in each of the left camera image and the right camera image based on the mapped corresponding particles, and performs image recognition of the subject 8 in each of the left camera image and the right camera image.
The tracking device 1 determines the likelihood of particles generated in the real space where the subject 8 exists based on the likelihood in the left camera image and the likelihood in the right camera image from the result of image recognition. For example, the tracking device 1 averages the likelihood in the left camera image and the likelihood in the right camera image to determine the likelihood of the particle generated in the real space where the subject 8 exists.

In this way, the tracking device 1 calculates the likelihood of each particle generated around the subject 8 in real space, and weights each particle based on the likelihood. From the distribution of the weights, a probability distribution of the location of the subject 8 can be obtained.
This probability distribution makes it possible to estimate the probability of the subject 8 being present in a certain space in three-dimensional real space (in this case, the space where the torso is located, since the particles are dispersed at about torso height).
This makes it possible to obtain the position of the subject 8 (place of high probability density).

Then, the tracking device 1 resamples the subject 8 and updates the probability distribution by selecting particles with large weights as resampling targets and deleting particles with small weights.
That is, around particles with large weights, many particles are randomly generated, and around particles with small weights, no particles are generated (or fewer particles are generated).
As a result, a distribution of particle density (shading) corresponding to the current probability distribution of the subject 8 is obtained.

The tracking device 1 newly acquires left and right images, calculates the likelihood of these newly generated particles, and updates the weights. This updates the probability distribution.
By repeating this process, the tracking device 1 can track the current position of the subject 8 (that is, the latest probability distribution).

In this way, the tracking device 1 uses a particle filter that repeats particle generation, likelihood observation, particle weighting, and resampling to track a position where the target person 8 is likely to exist.
Then, the tracking device 1 uses the virtual cameras 31a and 31b to survey a place where there is a high probability that the target person 8 is present, thereby determining the distance d to the target person 8 and the angle θ where the target person 8 is present. is calculated and the movement of the tracking robot is controlled based on this.
Note that the position of the target person 8 is expressed in a cylindrical coordinate system of (d, θ, height z), but since the height z of the pedestrian is considered to be constant, the position of the target person 8 is expressed by (d, θ). 8 position is shown.

In the second embodiment, the omnidirectional cameras 9a and 9b are disposed in the vertical direction, and the virtual cameras 31a and 31b are disposed in the vertical direction.
By providing the virtual cameras 31a and 31b above and below, the walking environment of the subject 8 can be photographed and surveyed in 360 degrees without any blind spots.

(2) Details of the embodiment (first embodiment)
Each diagram in FIG. 1 illustrates an example of the appearance of the tracked robot 12 according to the first embodiment.
The tracking robot 12 is an autonomous mobile tracking robot that recognizes a target and tracks it from behind.
In the following, the tracking target is mainly a pedestrian, but this is merely an example, and the tracking target may be other moving objects, such as a vehicle or an air vehicle such as a drone.

FIG. 1(a) shows an example in which a tracking robot 12a is constructed compactly using a tricycle, with the main purpose of tracking itself.
For example, you can watch over children and the elderly as they walk, follow a person in charge into a work site or disaster site to collect information, track livestock and other animals to monitor and observe them, and limit the number of people who can be targeted. It is possible to track and monitor areas to prevent them from entering the area.

The tracking robot 12a includes a cylindrical housing 15 that includes a pair of rear wheels 16 that constitute driving wheels, and one front wheel 17 that changes direction and guides the tracking direction.
In addition, these wheels may be an endless track used in a bulldozer or the like, or a leg structure like the arthropod of an insect.

Near the center of the upper surface of the casing 15, a columnar member whose height is approximately the height of a pedestrian's torso is vertically erected upward, and the photographing section 11 is provided at the tip of the columnar member.
The photographing unit 11 has two omnidirectional cameras 9a and 9b installed horizontally about 30 cm apart. Hereinafter, unless there is a particular distinction between them, they will simply be abbreviated as omnidirectional camera 9, and the same will apply to the other components.

The omnidirectional cameras 9a and 9b are each configured with a combination of fisheye lenses, and can provide a 360 degree field of view. The tracking device 1 (FIG. 2) mounted on the tracking robot 12a stereoscopically tracks the tracking target using virtual cameras 31a and 31b that cut out flat images from the respective spherical camera images taken by the spherical cameras 9a and 9b. and measure the distance and direction (angle, direction) of the tracked target using triangulation.
The tracking robot 12a moves behind the object to be tracked based on the survey results and tracks the object.

Inside the casing 15, a computer forming the tracking device 1, a communication device for communicating with a server, a mobile terminal, etc., a battery for supplying electric power, a drive device for driving wheels, and the like are housed.

FIG. 1(b) shows an example in which the tracking robot 12b is equipped with a loading function.
The tracking robot 12b includes a housing 20 whose longitudinal direction is the traveling direction. In addition to housing a computer, a communication device, a battery, a drive device, etc., the housing 20 can be equipped with, for example, a loading platform, a storage box, and a saddle-shaped seating section.
An imaging unit 11 similar to the tracking robot 12a is provided at the top end of the housing 20.
Further, the tracking robot 12b includes a pair of rear wheels 21 that constitute driving wheels, and a pair of front wheels 22 that change direction and guide the tracking direction. These wheels may be tracks or leg structures.

The tracking robot 12b can, for example, assist in transporting luggage or carry a person on the seat. Further, among the plurality of tracking robots 12b, the first tracking robot 12b is set to track the tracking target, and the other tracking robots 12b are set to track the immediately preceding tracking robot 12b. They can also be configured to be connected by software and run in tandem. This allows one guide to transport a large amount of luggage.

FIG. 1(c) shows an example in which the tracking robot 12c is mounted on a drone.
A plurality of propellers 26 for levitating the tracking device 1 are provided on the top surface of the housing 25, and the photographing unit 11 is suspended below the bottom surface. The tracking robot 12c tracks the target while floating and flying in the air.
For example, during a cold outbreak, it can track people who are not wearing masks and use an on-board loudspeaker to warn them, ``Wear a mask.''

FIG. 2 is a diagram showing the hardware configuration of the tracking device 1.
The tracking device 1 includes a CPU (Central Processing Unit) 2, a ROM (Read Only Memory) 3, a RAM (Random Access Memory) 4, a GPU (Graphics Processing Unit) 5, an imaging unit 11, and a storage unit. 10, control unit 6, drive The device 7 and the like are connected by a bus line.
The tracking device 1 three-dimensionally tracks the position of the subject 8 by image recognition using stereo camera images. Here, the target person 8 is assumed to be a pedestrian.

The CPU 2 performs image recognition of the subject 8 according to the tracking program stored in the storage unit 10, measures the position thereof, and issues a command for the tracking robot 12 to move to the control unit 6 according to the control program.
The ROM 3 is a read-only memory that stores basic programs and parameters for the CPU 2 to operate the tracking device 1 .

The RAM 4 is a readable and writable memory that provides a working memory for the CPU 2 to perform the above processing.
The image photographed by the photographing unit 11 is developed in the RAM 4 and used by the CPU 2.
The GPU 5 is an arithmetic device having a function of performing multiple calculations simultaneously and in parallel, and is used in this embodiment to perform high-speed parallel image processing for each particle based on a large number of generated particles.

The imaging unit 11 is configured using omnidirectional cameras 9a and 9b that can capture color images of 360 degrees of the surrounding area at once.
The omnidirectional cameras 9a and 9b are installed horizontally apart from each other by a predetermined distance (about 30 cm in this case), and acquire stereoscopic images of the subject 8.
When the subject 8 is in front of the tracking device 1, the omnidirectional camera 9a is located on the left side of the subject 8, and the omnidirectional camera 9b is located on the right side of the subject 8. When the subject 8 goes behind the tracking device 1, the left and right sides are reversed.

Since the omnidirectional cameras 9a and 9b are wide-angle cameras with a field of view of 360 degrees, the tracking device 1 thus acquires a left wide-angle image and a right wide-angle image from the left wide-angle camera and the right wide-angle camera, respectively. It is equipped with a wide-angle image acquisition means, and these left wide-angle cameras and right wide-angle cameras are respectively a left omnidirectional camera (if the subject 8 is located in front of the tracking robot 12, the omnidirectional camera 9a) and a right omnidirectional camera. It is composed of a spherical camera (all-celestial spherical camera 9b). Note that even if the field of view of these wide-angle cameras is 360 degrees or less, the tracking device 1 can be configured although the tracking range is limited.

Below, we will explain the case where the subject 8 is in front of the tracking device 1, and assume that the omnidirectional camera 9a photographs the subject 8 from the left side, and the omnidirectional camera 9b photographs the subject 8 from the right side. do.
When the subject 8 is located on the back side of the tracking device 1, the left and right sides of the explanation can be read interchangeably.
The drive device 7 includes a motor that drives wheels, and the control unit 6 controls the drive device 7 based on signals from the CPU 2 to adjust the traveling speed, turning direction, and the like.

Each figure in FIG. 3 is a diagram for explaining a virtual camera that captures a stereo image of the subject 8.
The omnidirectional camera 9a is configured by combining two fisheye lenses, and by pasting the left omnidirectional camera image taken with these onto the surface of the spherical object 30a shown in FIG. 3(a), Construct two fisheye camera images with one sphere.
As a result, a globe-like object whose surface is a 360-degree view around the spherical camera 9a is created.

Then, by placing a virtual camera 31a consisting of a virtual pinhole camera inside the spherical object 30a and virtually rotating it using software, it is possible to obtain a left camera image of the surrounding scenery as seen in the shooting direction of the virtual camera 31a with little distortion, similar to that of a monocular camera.

The virtual camera 31a can freely rotate continuously or discretely within the spherical object 30a to select a shooting direction.
Thereby, as shown by the arrow, the virtual camera 31a can be panned or tilted by an arbitrary amount in an arbitrary direction within the spherical object 30a.
In this way, the inside of the spherical object 30a serves as a virtual imaging space for the virtual camera 31a.

Since the virtual camera 31a is formed by software, it is not affected by the law of inertia, and the shooting direction can be controlled without using a mechanical mechanism. Therefore, the shooting direction can be switched continuously or discretely in an instant.
Note that it is also possible to provide a plurality of virtual cameras 31a in the spherical object 30a and rotate them independently to simultaneously obtain left camera images in a plurality of photographing directions.
For example, in the following, a case will be described in which a single target person 8 is tracked, but it is also possible to form virtual cameras 31a, 31a, . .

The above description has been made regarding the omnidirectional camera 9a, but the same applies to the omnidirectional camera 9b.
Although not shown, the right omnidirectional camera image can be acquired by the omnidirectional camera 9b and pasted onto the spherical object 30b, and the surrounding scenery can be photographed in a virtual photographing space by the virtual camera 31b.

The left omnidirectional camera image is composed of a fisheye lens image, and in the image of the desk shown in the example of FIG. 3(b), the straight portion of the desk is curved. For example, the left spherical camera image is composed of a fisheye lens image using an equidistant projection method or the like in which the angle is proportional to the distance from the center of the screen.
When this is photographed with the virtual camera 31a, a left camera image of the desk with little distortion is obtained, as shown in FIG. 3(c). In this way, when the virtual camera 31a is used, a two-dimensional camera image used in general image recognition can be obtained, so normal image recognition techniques can be applied.

The same applies to the right spherical camera image, and by using the virtual camera 31b, it is possible to obtain a two-dimensional camera image used for normal image recognition.
In this embodiment, the virtual cameras 31a and 31b are configured with virtual pinhole cameras, but this is just an example, and other methods of converting a fisheye lens image into a planar image may be used.
Here, the virtual cameras 31a and 31b function as a photographing means for photographing an object.

Each figure in FIG. 4 is a diagram for explaining a method of measuring the distance and direction to an object using a camera.
In order to track the subject 8, the tracking device 1 needs to measure the position of the subject 8 in a three-dimensional space (walking space) using a camera.
There are mainly three methods for such measurement:

FIG. 4(a) is a diagram showing a measurement method using geometric correction.
In the monocular geometric correction, the distance is determined based on the installation position of the monocular camera and the geometric state of the object 33 (how the object appears) in the camera image.
For example, the distance to the object 33 can be determined by the standing position of the object 33 with respect to the bottom of the camera image, and in the illustrated example, the standing positions when the distance to the object 33 is 1 m, 2 m, and 3 m are indicated by horizontal lines.
Furthermore, the direction in which the object 33 is present can be obtained from the left and right positions of the camera image on the horizontal line.

FIG. 4(b) is a diagram showing a measurement method using parallax stereo (compound eyes).
In the parallax stereo system, a pair of cameras 35a (left camera) and 35b (right camera) facing the front are fixed at a predetermined distance on the left and right, and the object 33 is viewed stereoscopically by the parallax from the cameras 35a, 35b with respect to the object 33.・Triangulate.
As shown in the figure, in the parallax stereo method, the larger triangle shown by the thick line is made up of the object 33 and the base line, and the smaller triangle shown by the thick line is made up of the base and the center of the lens due to the parallax formed on the imaging surface. The distance and direction of the object 33 can be determined from the similarity relationship between the two triangles.
For example, when the distance to the object is Z, the base line length is B, the focal length is F, and the parallax length is D, Z is expressed by equation (1). Directions can also be determined from similarity relationships.

FIG. 4(c) is a diagram showing a measurement method using the convergence stereo method.
Convergence refers to the so-called cross-eyed action, and a pair of cameras 36a (left camera) and camera 36b (right camera) arranged at a predetermined distance on the left and right side converge to view the object 33, thereby creating a stereoscopic view of the object 33.・Survey.
As shown in the figure, in the vergence stereo method, the shooting directions of the right camera and the left camera are directed toward the object 33, the baseline length is B, the distance from the left camera to the object 33 is dL, and the optical axis of the left camera lens is If the angle with the front is θL, the angle between the optical axis of the right camera lens and the front is θR, the orientation of the object 33 with respect to the vergence stereo camera is θ, and the distance from the vergence stereo camera to the object 33 is d, then the geometric From the relationship, dL is expressed by equation (2), and d can be calculated from equation (3). The angle θ corresponding to the azimuth can also be determined from the geometric relationship.
Note that to prevent erroneous conversion of character codes (so-called garbled characters), the subscripts and superscripts shown in the diagram are written in normal characters. The same applies to other formulas described below.

Any of the three measurement methods mentioned above can be used, but as described below, among these measurement methods, the vergence stereo method is superior in pedestrian tracking and exhibits outstanding ability, so this method is not suitable for this purpose. In the embodiment, a congestion stereo method is adopted.

FIG. 5 is a diagram for explaining the superiority of the congestion stereo system.
It is clear that the parallax stereo method and the vergence stereo method are superior to the monocular method, so a description of the monocular method will be omitted.
As shown in FIG. 5A, in the parallax stereo system, the shooting directions of the cameras 35a and 35b are fixed to the front. Therefore, the photographing area 37a by the camera 35a and the photographing area 37b by the camera 35b are also fixed, and the common photographing area 37c becomes an area that can be surveyed.

On the other hand, in the convergence stereo system, the shooting directions of the left and right cameras can be freely set individually by rotating the cameras 36a and 36b independently, so it is possible to stereoscopically view and survey a wide area other than the common shooting area 37c. be.
For example, as shown in FIG. 5B, even if the object 33 is located at a short distance in front of the camera and outside the photographing area 37c, the left and right virtual cameras 33 By viewing the target 33 convergently, the position and direction can be measured.

Further, as shown in FIG. 5(c), even if the object 33 is located to the left and is included in the imaging area 37a but not included in the imaging area 37b, the arrow As shown by the line, it can be surveyed using convergence vision. The same applies when the object 33 is located on the right side.

As shown in FIG. 5(d), even if the target 33 is located further to the left and is not included in the imaging area 37a, the survey can be performed using convergence vision as shown by the arrow. can do. The same applies when the object 33 is located on the right side.
In this way, the vergence stereo method can measure a wider area than the parallax stereo method, and is suitable for tracking pedestrians who move freely and whose walking status frequently changes from a short distance.
Therefore, in this embodiment, virtual cameras 31a and 31b are formed on the omnidirectional cameras 9a and 9b, and the subject 8 is viewed convergently by these.

In this way, the photographing means included in the tracking device 1 photographs the object using a vergence stereo camera using a left camera and a right camera.
The photographing means constitutes a left camera with a virtual camera (virtual camera 31a) that acquires a left camera image in an arbitrary direction from a left wide-angle image (left omnidirectional camera image), and a right wide-angle image ( The right camera is constituted by a virtual camera (virtual camera 31b) that acquires a right camera image in an arbitrary direction from the right spherical camera image).
Furthermore, the tracking device 1 shoots in a virtual shooting space (a shooting space using spherical objects 30a and 30b) where the left camera and the right camera acquire a left camera image and a right camera image from the left wide-angle image and the right wide-angle image, respectively. Can move direction.

The tracking device 1 uses a particle filter to track the location of the target person 8. Here, an overview of general particle filtering will be explained.
First, in particle filtering, a large number of particles are generated in locations where an observation target may exist.
Then, the likelihood of each particle is observed using some method, and each particle is weighted according to the observed likelihood. The likelihood corresponds to the degree of certainty that the observed object is the observed object when observed based on the particle.

After observing the likelihood of each particle, each particle is weighted such that the greater the likelihood, the greater the weight. As a result, particles are weighted more heavily in locations where the observation target exists to a higher degree, so that the weighted particle distribution corresponds to a probability distribution representing the existence of the observation target.

Furthermore, resampling is performed to track time-series changes in the probability distribution as the tracking target moves.
In resampling, for example, particles with low weighting are thinned out, particles with high weighting are left behind, new particles are generated near the remaining particles, and the likelihood of each generated particle is observed at the current moment. and weight it. Thereby, the probability distribution is updated, and it is possible to update a location with a high probability density, that is, a location where the observation target is likely to exist.
Thereafter, by repeating resampling, it is possible to track time-series changes in the position of the observation target.

Each figure in FIG. 6 is a diagram for explaining a method of generating particles.
The tracking device 1 estimates the probability distribution of the position where the target person 8 exists using a particle filter.
In general image recognition using a particle filter, particles are generated in a two-dimensional camera image, but in contrast, the tracking device 1 generates particles in a three-dimensional space where the subject 8 exists. By mapping and projecting these three-dimensional particles onto the left and right camera images, the image of the subject 8 including the three-dimensional information is recognized.

When recognizing images without including stereoscopic information, it is necessary to generate particles independently in the right and left camera images. In this case, the left and right cameras may observe different positions, which may affect surveying accuracy and result in mistracking.
On the other hand, the tracking device 1 performs image recognition using left and right camera images taken by pointing the left and right cameras at the same particle in three-dimensional space, so that the same area can be observed by both the left and right cameras, thereby enabling effective searching for the subject 8.

In this way, the tracking device 1 generates particles around the target person 8, but in this embodiment, the tracking target is a pedestrian walking in front of the tracking device 1, and particles are generated in a two-dimensional manner parallel to the floor surface. Since the robot moves in a consistent manner, we decided to scatter the particles on a plane parallel to the walking surface.
Note that if the target to be tracked, such as a drone or bird, moves in the height direction and moves in a three-dimensional manner, it can be tracked by scattering particles three-dimensionally.

FIG. 6A shows the subject 8 walking in an xyz space with the tracking device 1 set as the origin.
An xy coordinate system is set on the plane (walking surface) on which the subject 8 walks, and the height direction is set as the z axis. The imaging unit 11 is located at the height of the torso of the subject 8 (about 1 m).

As shown in the figure, the tracking device 1 generates noise centered around the subject 8 so that particles are scattered in a circular area 32 parallel to the xy plane at a height approximately near the torso. A predetermined number of particles centered on particle 8 are generated.
In this embodiment, 500 particles were generated. According to experiments, it is possible to track the number of particles starting from about 50.
Although particles are generated here on a plane including the circular region 32, they may be distributed in a thick space with width in the height direction (z-axis direction).

The position of the torso is a place with a high probability density that the target person 8 exists, and since the particles are resampled according to the weights (according to the probability distribution) after weighting, the tracking device 1 calculates the probability of the position where the target person exists. The apparatus includes particle generation means for generating particles to be used in the particle filter in a three-dimensional space based on the distribution.
Further, the particle generating means generates particles along a plane parallel to the plane in which the object moves.
Furthermore, in order to follow time-series changes in the probability distribution as the subject 8 moves through resampling, the particle generation means sequentially generates the current particles based on the previously updated probability distribution.

Here, the generated noise is white noise (normal white noise) that follows a Gaussian distribution centered around the subject 8, and by following the noise, particles can be generated around the subject 8 according to a normal distribution. can. The circular region 32 in the figure is a range of, for example, about 3σ of the generated particles.
Note that other generation methods may be adopted, such as generating particles uniformly in the circular region 32.

In addition, as will be described later, at the start of tracking, the tracking device 1 measures the position of the target person 8 using normal image recognition, and based on this, generates particles centered on the target person 8. If the position of the target person 8 is unknown, the probability distribution of the existence of the target person 8 is uniform in space, so it is sufficient to uniformly generate particles in the xy plane including the circular area 32.
Since the likelihood of particles at a location where the target person 8 is present becomes higher, by resampling this, a probability distribution according to the position of the target person 8 can be obtained.
The tracking device 1 tracks the subject 8 by resampling the particles generated as described above.

FIG. 6(b) is a diagram schematically showing the circular area 32 viewed from above.
As shown by the black dots in the figure, particles are generated in a circular area 32 centered on the subject 8, but since these z-coordinate values are constant, the tracking device 1 generates particles in a circular area 32 for convenience. The position of the subject 8 is expressed in polar coordinates using (d, θ) coordinates. Note that it may be expressed in xy coordinates.

Furthermore, if the direction in which the subject 8 is walking is known, the distribution of particles can be generated in a circular region 32a whose longitudinal direction is the walking direction, as shown in FIG. 6(c). . By generating particles along the walking direction, it is possible to suppress unnecessary calculations caused by generating particles in places where the probability that the subject 8 is present is low.

Furthermore, in order to scatter particles also in the depth direction of the camera image, which is the shooting direction, for example, when the tracking device 1 is moving in a hallway in a building, the layout of the floor plan is obtained from the plan view inside the building, By referring to this, it is possible to prevent particles from being generated in a place where there is no possibility that the subject 8 exists, such as inside a wall or a room where entry is prohibited.
In this way, since the tracking device 1 generates particles also in the depth direction of imaging in the three-dimensional space in which the subject 8 moves, the tracking device 1 generates particles in an arbitrary distribution that takes into account the state of motion of the tracked subject and the surrounding environment. It is possible to do so.

FIG. 7 is a diagram for explaining mapping of particles to a camera image.
The tracking device 1 captures the particles generated as above in a camera image 71a (left camera) using functions g(d, θ) and f(d, θ) as shown in FIG. image) and camera image 71b (right camera image) to the camera image coordinate system.
The camera image coordinate system is, for example, a two-dimensional coordinate system with the upper left corner of the image as the origin, the horizontal right direction as the x axis, and the vertically downward direction as the y axis.

In this way, the tracking device 1 is provided with a mapping means for mapping particles generated in the real space in which the subject 8 exists onto the captured image.
The mapping means acquires the positions of the generated particles in the left camera image and the right camera image by calculating them using a predetermined mapping function.

As a result, for example, particles 41 scattered in space are mapped to particles 51a on the camera image 71a by the function g(d, θ), and mapped to particles 51b on the camera image 71b by the function f(d, θ). be done.
Note that these mapping functions can be derived by calculating the relational expression of convergence stereo vision and the angle for each pixel of the camera image acquired by the virtual camera 31.
In this way, the mapping means maps the particles generated in real space to the left camera image and the right camera image taken by the left camera and the right camera, respectively, in association with each other.

Incidentally, the particles 41 are accompanied by state parameters that are parameters for setting the detection area in the camera image, such as the position of the detection area for image recognition and the size of the detection area, and the tracking device 1 Based on this, a detection area 61a and a detection area 61b are set in the camera image 71a and the camera image 71b, respectively.
In this way, the particles 41, 42, 43, . . . are represented by state vectors having state parameters as components.

The detection areas 61a and 61b have a rectangular shape, and images within the detection areas 61a and 61b become partial area images to be subjected to image recognition. The tracking device 1 performs image recognition of the subject 8 using each partial region image divided by the detection regions 61a and 61b.
Here, the detection areas 61a and 61b are set so that the particles 51a and 51b after mapping become the center of gravity of the rectangle. This is just an example, and the position of the detection area 61 may be offset from the position of the particle 51 using a fixed value or a function.
In this way, the tracking device 1 is equipped with an image recognition unit that sets a detection area based on the position of the mapped particle in the camera image and recognizes the photographed object as an image.

Further, since the tracking device 1 tracks a pedestrian at a predetermined distance, the sizes of the detection areas 61a and 61b rarely change greatly.
Therefore, in the tracking device 1, the size of the detection area 61 is set according to the height of the subject 8 before tracking, and the detection areas 61a and 61b of fixed size are used.

Note that this is just an example, and the size of the detection area 61 can also be used as a parameter for particle filtering.
In this case, particles are generated in a state vector space of (x coordinate value, y coordinate value, size).
That is, even if the x and y coordinate values are the same, if the sizes are different, the particles are different, and the likelihood is observed for each particle. This increases the likelihood of particles whose size is suitable for image recognition, and thereby allows the optimum size of the detection region 61 to be determined.

In this way, by generating particles in the state vector space that defines the particles 41 without being limited to the real space, more expanded operation becomes possible. If there are n parameters, particles will be generated in an n-dimensional space.
For example, there is likelihood 1, which is calculated by the first method, and likelihood 2, which is calculated by the second method, and the former is α and the latter is combined at a ratio of (α-1) (for example, 0 <α<1) If you want to calculate the likelihood by combining both, let the state vector be (x coordinate value, y coordinate value, size, α).

When the particles 41 are generated in such a state vector space, the likelihood can be calculated for different α by particle filtering, and the optimal (x coordinate value, y coordinate value) for image recognition of the subject 8 value, size, α) and the likelihood in that case.
Regarding the combination of likelihoods using α, an example of combining the likelihood based on the HOG feature amount and the likelihood based on the color distribution feature will be described later.

The tracking device 1 generates particles according to such a procedure, and as shown in FIG. 7(b), maps the unillustrated particles 41, 42, . . . onto the particles 51a, 52a, . . . in the camera image 71a, Based on this, detection areas 61a, 62a, . . . are set.
Also for the camera image 71b, particles 41, 42, . . . are mapped onto particles 51b, 52b, . . . and detection regions 61b, 62b, .

Then, the tracking device 1 recognizes the target person 8 in the detection area 61a of the camera image 71a, thereby determining the likelihood of the particle 51a (the likelihood of the mapped particle in the left camera image, hereinafter referred to as the left likelihood). ), and by image-recognizing the subject 8 in the detection area 61b of the camera image 71b, the likelihood of the particle 51b (the likelihood in the right camera image of the mapped particle, hereinafter referred to as right likelihood) is calculated. The likelihood of the mapping source particle 41 is calculated by averaging the left likelihood and the right likelihood.

The tracking device 1 similarly calculates the likelihood of the particles 42, 43, . . . generated in the three-dimensional space.
In this way, the tracking device 1 maps particles generated in the three-dimensional space in which the subject 8 is walking onto a pair of left and right stereo camera images, and calculates the left likelihood of the particles mapped onto the two-dimensional camera image. and the right likelihood to calculate the likelihood of the mapping source particle.

The tracking device 1 integrated the left likelihood and the right likelihood to observe the likelihood of the mapping source particle in the three-dimensional space, but this is just an example, and it could be integrated by other calculation methods. It's okay.
Further, it is sufficient to obtain an integrated likelihood using at least one of the left likelihood and the right likelihood, such as by using the higher likelihood of the right likelihood and the left likelihood as the likelihood of the mapping source.

In this way, the image recognition means included in the tracking device 1 recognizes the left camera image and the right camera image, respectively.
The tracking device 1 is equipped with a likelihood acquisition unit that acquires the likelihood of the generated particles based on the result of image recognition, and the likelihood acquisition unit includes a first The likelihood is obtained using at least one of the likelihood (left likelihood) and the second likelihood (right likelihood) based on image recognition of the right camera image.

In the above example, the particles 41, 42, 43, ... were mapped to a pair of left and right stereo camera images by calculating with the functions g (d, θ) and f (d, θ), but the virtual camera 31a , 31b, the virtual camera 31a and the virtual camera 31b are directed toward each of the generated particles 41, 42, . . . to obtain left and right camera images for each particle. It is also possible to map the particles 41, 42, . . . to the center of the image for each set.

In this modified example, virtual cameras 31a and 31b are pointed at particle 41 to obtain camera image 81a (left camera image) and camera image 81b (right camera image) as shown in FIG. 7(c), then virtual cameras 31a and 31b are pointed at particle 42 to obtain camera image 82a (left camera image) and camera image 82b (right camera image)... and so on, acquiring stereo camera images with the camera direction pointed at each particle. However, only the left camera image is shown in the figure, and the right camera image is omitted.

The pinhole camera constituting the virtual camera 31 has a single focus, and even if the virtual camera 31 is aimed at particles 41, 42, etc. within the spherical object 30, the image of the subject 8 will remain in focus. It can be obtained with.
Further, since the virtual camera 31 is formed by software, it does not require mechanical driving and can quickly switch the shooting direction to take photos of the particles 41, 42, . . . .
Alternatively, it is also possible to set a plurality of virtual cameras 31, 31, . . . and drive them in parallel to obtain a plurality of stereo camera images at once.

As shown in FIG. 7C, when the virtual camera 31a is directed toward the particle 41 and photographed, a camera image 81a is obtained in which the particle 41 is mapped onto the particle 51a at the center of the image.
Although not shown, similarly, when the virtual camera 31b is directed toward the particle 41 to take an image, a camera image 81b in which the particle 41 is mapped onto the particle 51b at the center of the image is obtained.
The tracking device 1 performs image recognition using the camera images 81a and 81b to obtain the left likelihood and right likelihood of the particles 51a and 51b, and averages these to obtain the likelihood of the particle 41.

Thereafter, in the same way, the virtual cameras 31a and 31b are directed toward the particles 42 and photographed to obtain camera images 82a and 82b (the camera image 82b is not shown). The likelihood of the particle 42 is calculated from the left likelihood and right likelihood of 52b.
The tracking device 1 repeats this process to calculate the likelihood of the particles 41, 42, 43, . . . .

In this way, the photographing means in this example photographs each generated particle by pointing the left camera and the right camera, and the mapping means positions the left camera image and the right camera image at positions corresponding to the photographing directions (for example, the image center) is obtained as the particle position.
Two methods of mapping particles generated in the three-dimensional space in which the subject 8 walks to the left and right camera images have been described above, and below, a case in which the particles are mapped using the former method will be described. Note that mapping may be performed using the latter method.

Each figure in FIG. 8 is a diagram for explaining a method of tracking the position of the subject 8 with the virtual camera 31.
As explained above, the tracking device 1 performs image recognition using the detection area 61a in the camera image 71a, as shown in FIG. 8(a), and thereby calculates the left likelihood of the particle 51a. Then, in the camera image 71b (not shown), image recognition is performed using the detection area 61b, thereby calculating the right likelihood of the particle 51b.

Furthermore, the tracking device 1 calculates the likelihood of the particle 41, which is the mapping source of the particles 51a and 51b, by averaging the left likelihood and the right likelihood.
The tracking device 1 repeats this calculation and calculates the likelihood of the particles 42, 43, . . . scattered three-dimensionally around the subject 8.

Then, the tracking device 1 weights each particle generated in the three-dimensional space according to the calculated likelihood so that the larger the likelihood, the larger the weight.
FIG. 8(b) shows the particles 41, 42, 43, . . . after weighting, and the larger the weighting, the larger the size of the black spot.
In the illustrated example, the weight of the particle 41 is the largest, and the weights of the surrounding particles are also large.

In this way, a weighted particle distribution in real space is obtained, and this weight distribution corresponds to the probability distribution of the position where the subject 8 is present. Therefore, in the illustrated example, it can be inferred that the subject 8 is near the particle 41.
Various estimation methods are possible, such as estimating that the tracked object exists at the position of the peak weight, or estimating that the tracked object exists within the top 5% of the weights.

By updating the probability distribution using resampling, the location of the target person 8 can be tracked.
In this way, the tracking device 1 includes tracking means that tracks the location of the target by updating the probability distribution based on the acquired likelihood.

Then, the tracking device 1 directs the virtual cameras 31a, 31b to a location where the probability distribution is large (that is, a location where the subject 8 is likely to be present), thereby changing the imaging direction of the virtual cameras 31a, 31b toward the subject 8. can be directed to.
In the example of FIG. 8C, the virtual cameras 31a and 31b are directed toward the particle 41 that has the highest likelihood.
In this way, the tracking device 1 includes a photographing direction moving means that moves the photographing directions of the left camera and the right camera in the direction of the object based on the updated probability distribution.

Here, the virtual camera 31 is directed to the particle with the highest likelihood, but this is just an example, and the virtual camera 31 may be directed to a location with a high probability distribution according to some algorithm.
In this way, by pointing the virtual cameras 31a and 31b at a location with high probability density, the subject 8 can be captured in front of the camera.

Furthermore, since the position (d, θ) of the subject 8 can be measured from the angle at which the virtual cameras 31a, 31b converge, a command can be issued to the control unit 6 based on the output value of the position (d, θ) to control the tracking device 1 to move to a specified position behind the subject 8.
In this way, the tracking device 1 is equipped with a surveying means that surveys the position of the target based on the shooting directions of the left camera and right camera that are moved based on the probability distribution, and an output means that outputs the surveying results, and further includes a moving means that drives the drive unit 7 based on the outputted surveying results, thereby moving together with the target.

After weighting the particles as shown in Figure 8(b), resampling is performed to update the probability distribution in accordance with the movement of the subject 8. This is performed by generating the next particle (or generating a large number of particles) in the vicinity of a particle with high likelihood such as particle 41 according to white noise, and not generating the next particle in the vicinity of a particle with low likelihood (or generating a small number of particles), and then calculating the likelihood of the new particles generated in this way using new left and right camera images and weighting them.

In this way, the probability distribution is updated by sequentially repeating the process of resampling those with a high likelihood and reducing those with a low likelihood, and successively selects locations with a high probability that the target person 8 is present. can be tracked.
In the present embodiment, as an example, the state is changed based on equation (4) in FIG. 8(d), which takes into account the speed information of the subject 8 (particles for resampling are generated).
Here, xt represents the position of the particle at time t, and xt-1 represents the position of the particle at time t-1.

vt-1 is velocity information of the subject 8, and is obtained by subtracting the position at time t-1 from the position at time t, as shown in equation (6).
N(0, σ2) is a noise term and represents a normal distribution of variance σ2 at the particle position.
As expressed by Equation (5), σ2 was set so that the larger the speed, the larger the amount of movement of the subject 8, and therefore the variance was set accordingly.

FIG. 9 is a diagram for explaining the likelihood calculation method.
Although any method can be used to calculate the likelihood, an example using HOG features will be described here. This calculation method can be used to calculate right likelihood and left likelihood.
The HOG feature is an image feature using a brightness gradient distribution, and is a technique for detecting edges of an object. For example, objects are recognized by silhouettes created by edges.

The HOG feature amount is extracted from the image by the following procedure.
An image 101 shown in the left diagram of FIG. 9(a) shows an image extracted from a camera image using a detection area.
First, the image 101 is divided into rectangular cells 102a, 102b, . . . .
Next, as shown in the right diagram of FIG. 9A, the brightness gradient direction (direction from low brightness to high brightness) of each pixel is quantized into, for example, eight directions for each cell 102.

Next, as shown in FIG. 9B, by generating a histogram in which the direction of the quantized brightness gradient is the class and the number of appearances is the frequency, the histogram 106 of the brightness gradient included in the cell 102 is Create every 102.
Then, the histogram 106 is normalized so that the total frequency of the histogram 106 is 1 for each block in which several cells 102 are collected.

In the example shown in the left diagram of FIG. 9A, one block is formed from cells 102a, 102b, 102c, and 102d.
A histogram 107 obtained by arranging the thus normalized histograms 106a, 106b, . . . (not shown) in a line as shown in FIG. 9C is the HOG feature amount of the image 101.

The degree of similarity between images using HOG features is determined as follows.
First, consider a vector φ(x) whose components are the frequencies (assuming there are M) of HOG features. Here, x is a vector representing the image 101, and x=(brightness of the first pixel, brightness of the second pixel, . . . ).
Note that vectors are expressed in bold, etc., but in order to prevent erroneous conversion of character codes, vectors are expressed in normal letters below.

FIG. 9D shows the HOG feature space, and the HOG feature of the image 101 is mapped to a vector φ(x) in the M-dimensional space.
Note that in the figure, the HOG feature space is represented in a two-dimensional space for simplicity.
On the other hand, F is a weight vector obtained by learning human images, and is a vector obtained by averaging HOG feature amounts of a large number of human images.

When the image 101 is similar to the learned image, φ(x) is distributed around F as in the vector 109, and when it is not similar, it is distributed in a direction different from F as in the vectors 110 and 111.
F and φ(x) are standardized, and the correlation coefficient defined by the inner product of F and φ(x) approaches 1 as the image 101 resembles the training image, and decreases as the degree of similarity decreases. approaches 1.
In this way, by mapping the images to be subjected to similarity determination into the HOG feature space, it is possible to separate images that are similar to the learning image and images that are not similar based on the brightness gradient distribution.
This correlation coefficient can be used as a likelihood.

In addition to this, it is also possible to evaluate the likelihood using color distribution features.
For example, the image 101 is composed of pixels having various color components (color 1, color 2, . . . ).
When a histogram is created from the frequencies of appearance of these color components, a vector q whose components are the frequencies is obtained.
On the other hand, a similar histogram is created for a tracking target model prepared in advance using the subject 8, and a vector p whose components are the frequencies of the histogram is created.
When the image 101 is similar to the tracked model, q is distributed around p, and when it is not similar, q is distributed in a direction different from p.

q and p are standardized, and the correlation coefficient defined by the inner product of q and p approaches 1 as the image 101 resembles the tracked model, and approaches -1 as the degree of similarity decreases.
In this way, by mapping the image that is the target of similarity determination into the color feature space, images that are similar to the tracking target model and images that are not similar can be separated based on the color feature distribution.
This correlation coefficient can also be used as a likelihood.

Furthermore, for example, it is also possible to combine the similarity based on the HOG feature amount and the similarity based on the color distribution feature.
HOG feature amounts and color distribution features have scenes that are good at recognition and scenes that are bad for recognition, and by combining these, the robustness of image recognition can be improved.
In this case, using the parameter α explained earlier (0.25<α<0.75 by experiment), α x (similarity based on HOG feature) + (1 - α) x (based on color distribution feature α that maximizes the likelihood can also be found by defining the likelihood in terms of similarity (similarity) and generating particles in a state vector space that includes α.
According to this equation, the larger α is, the larger the contribution of the HOG feature is, and the smaller α is, the larger the contribution of the color distribution feature is.
Therefore, by appropriately selecting α, a value suitable for the scene can be obtained, and robustness is improved.

FIG. 10 is a flowchart for explaining the tracking process performed by the tracking device 1.
The following processing is performed by the CPU 2 according to the tracking program stored in the storage unit 10.
First, the CPU 2 asks the user to input the height of the subject 8, etc., sets the size of the left and right detection areas based on this, and stores this in the RAM 4.
Next, the subject 8 is asked to stand at a predetermined position in front of the tracking device 1, and the CPU 2 photographs this with the virtual cameras 31a and 31b, obtains a left camera image and a right camera image, and stores them in the RAM 4. (Step 5).

More specifically, the CPU 2 stores in the RAM 4 the left omnidirectional camera image and the right omnidirectional camera image taken by the omnidirectional cameras 9a and 9b, and pastes these onto the spherical objects 30a and 30b, respectively, by calculation. Ru.
Then, a left camera image and a right camera image photographed from the inside using the virtual cameras 31a and 31b are obtained by calculation and stored in the RAM 4.

Next, the CPU 2 recognizes the subject 8 from the left and right camera images (step 10).
This image recognition uses a commonly used method, such as searching for the target person 8 by scanning a detection area of a size stored in the RAM 4 using left and right camera images, respectively.
Then, the CPU 2 directs the virtual cameras 31a and 31b toward the subject 8 whose image has been recognized.

Next, the CPU 2 acquires the location of the target person 8 by measuring the position of the target person 8 from the angles of the virtual cameras 31a and 31b using the distance d and the angle θ to the target person 8, and stores it in the RAM 4. Remember.
Then, the CPU 2 calculates the position and direction of the subject 8 with respect to the tracking robot 12 from the obtained position (d, θ) of the subject 8, the front direction of the tracking robot 12, and the angle with respect to the virtual cameras 31a, 31b, A command is issued to the controller 6 to move the tracking robot 12 so that the subject 8 is located at a predetermined position in front of the tracking robot 12. At this time, the CPU 2 adjusts the angles of the virtual cameras 31a and 31b so that the subject 8 is captured in front of the camera.

Next, the CPU 2 generates white noise on a horizontal plane at a predetermined height (around the torso) where the subject 8 is present, and generates a predetermined number of particles accordingly (step 15). Then, the CPU 2 stores the position (d, θ) of each particle in the RAM 4.
The processing for each particle in steps 20 and 25 below is performed in parallel by the GPU 5, but here, for the sake of simplicity, it is assumed that the processing is performed by the CPU 2.

Next, the CPU 2 selects one of the generated particles and converts it into a left camera image and a right camera image using functions g(d, θ) and f(d, θ), respectively. The particles are mapped onto a camera image, and the image coordinate values of these mapped particles are stored in the RAM 4 (step 20).
Next, the CPU 2 calculates the left camera image likelihood and right camera image likelihood based on the mapped particles for each of the left camera image and the right camera image, and calculates the likelihood of the mapping source particle by averaging these. and stores it in the RAM 4 (step 25).

Next, the CPU 2 determines whether the likelihood has been calculated for all of the generated mapping source particles (step 30).
If there are still particles that have not been calculated (step 30; N), the process returns to step 20 to calculate the likelihood of the next particle.
On the other hand, when the likelihoods of all particles have been calculated (step 30; Y), the CPU 2 weights each particle based on the likelihood of the particles, and stores the weight for each particle in the RAM 4.

Next, the CPU 2 estimates the position of the subject 8 with respect to the imaging unit 11 based on the distribution of particle weights, and directs the virtual cameras 31a and 31b to the estimated position of the subject 8.
Then, the CPU 2 calculates the position of the subject 8 from the angles of the virtual cameras 31a and 31b, and stores the calculated coordinates (d, θ) of the subject 8 in the RAM 4 (step 35).

Further, the CPU 2 determines the target person relative to the tracking robot 12 based on the coordinates (d, θ) of the target person 8 stored in the RAM 4 in step 35 and the angle formed by the front direction of the tracking robot 12 and the shooting direction of the virtual cameras 31a and 31b. The coordinates of the position of the subject 8 are calculated, and using the coordinates, a command is issued to the control unit 6 to control the movement of the tracking robot 12 so that it moves to a predetermined tracking position behind the subject 8 (step 40).
In response to this, the control unit 6 drives the drive device 7 to move the tracking robot 12 to track the subject 8 after the subject 8 .

Next, the CPU 2 determines whether to end the tracking process (step 45). If it is determined to continue the process (step 45; N), the CPU 2 returns to step 15 and generates the next particle, and if it is determined to end the process (step 45; Y), it ends the process.
This judgment can be made, for example, by asking the subject 8 to say something like "I have arrived" when he or she has reached the destination, and by having the subject 8 recognize this by voice, or by having the subject make a specific gesture. and do it.

Although the tracking device 1 of this embodiment has been described above, various modifications are possible.
For example, the tracking robot 12 is equipped with the photographing unit 11, the control unit 6, and the drive device 7, and the other components are provided in the server with the tracking device 1, and the tracking robot 12 and the server are connected through a communication line. The robot 12 can also be controlled remotely.

Further, in addition to the virtual cameras 31a and 31b, the photographing unit 11 may be provided with a virtual camera for external observation, and the image taken by the camera may be transmitted to the server.
Furthermore, the tracking device 1 is equipped with a microphone and a speaker, so that a third party can have a conversation with the person to be tracked while observing the image of the virtual camera for external observation via a mobile terminal or the like.
In this case, for example, the tracking robot 12 accompanies the elderly person for a walk, and the caregiver observes the surroundings of the tracking robot 12 from a mobile terminal and tells the elderly person, ``Please be careful because a car is coming.'' It becomes possible to call out.

(Second embodiment)
In the imaging unit 11 of the tracking device 1 of the first embodiment, the omnidirectional cameras 9a and 9b are arranged in the left and right directions, but in the imaging unit 11b of the tracking device 1b of the second embodiment, these are arranged in the vertical direction. to be placed.
Although not shown, the configuration of the tracking device 1b is the same as the tracking device 1 shown in FIG. 2, except that omnidirectional cameras 9a and 9b are arranged in the vertical direction.

FIG. 11 is a diagram showing an example of the appearance of the tracking robot 12 according to the second embodiment.
The tracking robot 12d shown in FIG. 11(a) is the tracking robot 12a (FIG. 1(a)) in which omnidirectional cameras 9a and 9b are installed in the vertical direction.
The photographing unit 11b is arranged at the tip of the columnar member, and the omnidirectional camera 9a is disposed on the upper side in the vertical direction, and the omnidirectional camera 9b is disposed on the lower side in the vertical direction.

In this way, in the first embodiment, the photographing section 11 was installed so that the longitudinal direction was the horizontal direction, but in the second embodiment, the photographing section 11b was installed so that the longitudinal direction was the vertical direction.
It is also possible to arrange the omnidirectional camera 9a to be located diagonally above the omnidirectional camera 9b, or to arrange the omnidirectional camera 9a to be located above the horizontal plane where the omnidirectional camera 9b is located. It may be arranged so that it is located on the lower side.
In this way, the tracking device 1b includes a photographing means for photographing an object using a vergence stereo camera using an upper camera disposed above a predetermined horizontal plane and a lower camera disposed below.

In the case of the photographing section 11, since the omnidirectional cameras 9a and 9b are installed in the horizontal direction (lateral direction), the horizontal direction becomes a blind spot.However, in the photographing section 11b, the omnidirectional cameras 9a and 9b are installed vertically. Because it is installed in the vertical direction, there is no blind spot over the entire 360-degree circumference, and images of the target person 8 are acquired no matter where the target person 8 is located around the tracking robot 12. be able to.
The tracking robots 12e and 12f shown in FIGS. 11(b) and 12f correspond to the tracking robots 12b and 12c shown in FIGS. 1(b) and 12c, respectively. 9b are arranged one above the other.

FIG. 11(d) is an example in which a pillar is erected on the road surface and the photographing unit 11b is attached to the tip of the pillar. Passersby walking on the street can be tracked.
In Figure 11(e), two pillars with different heights are erected on the road surface, a spherical camera 9b is attached to the tip of the lower pillar, and a spherical camera 9a is attached to the tip of the higher pillar. This is an example in which the photographing section 11b is configured as follows.
In this way, the omnidirectional cameras 9a and 9b may be attached to separate support members, or may be installed diagonally up and down.
FIG. 11F shows an example in which the photographing section 11b is installed under the eaves of a building such as a house or a building.

FIG. 11(g) is an example in which a photographing section 11b is provided at the tip of a flag raised by a leader of a group tour. By facial recognition of the faces of group guests, it is possible to track the locations of individual participants.
FIG. 11(h) is an example in which the photographing unit 11b is installed on the roof of the vehicle. It is possible to obtain the positions of surrounding environmental objects, such as the position of the vehicle ahead.
FIG. 11(i) is an example in which the photographing section 11b is installed on a tripod. It can be used in the civil engineering field.

FIG. 12 is a diagram for explaining the surveying method in the second embodiment.
The method of generating particles is the same as in the first embodiment.
As shown in FIG. 12A, the tracking device 1b converges virtual cameras 31a and 31b (not shown) provided on the omnidirectional cameras 9a and 9b in a plane including the z-axis and the subject 8, and It rotates around the z-axis (rotation angle is φ) and directs the photographing direction toward the subject 8.
As shown in FIG. 12(b), the tracking device 1b determines the position of the subject 8 using coordinates (d, φ).

Regarding each means of the tracking device 1b other than the photographing means, a particle generating means for generating particles, a tracking means for tracking the position of an object, an output means for outputting survey results, and a movement for moving based on the survey results. The means are the same as the tracking device 1.

The tracking device 1b also includes a mapping means for mapping particles, an image recognition means for image recognition, a likelihood acquisition means for acquiring the likelihood of particles, a photographing direction moving means for moving the photographing direction, and a photographing direction moving means for moving the photographing direction. Regarding the surveying means for surveying and the wide-angle image acquisition means for acquiring wide-angle images, left and right cameras are arranged vertically, and left camera, right camera, left camera image, right camera image, left wide-angle camera, right wide-angle camera, and left wide-angle camera are used. image, right wide-angle image, left spherical camera, right spherical camera, upper camera, lower camera, upper camera image, lower camera image, upper wide-angle camera, lower wide-angle camera, upper wide-angle image, lower wide-angle image, respectively. , an upper omnidirectional camera, and a lower omnidirectional camera.

As described above, the following configuration can be obtained in the first embodiment and the second embodiment.
(1) Configuration of the first embodiment (101st configuration) Particle generating means that generates particles to be used in a particle filter in a three-dimensional space based on the probability distribution of the position where the target exists, and a photographing unit that photographs the target. a mapping means for mapping the generated particles onto the photographed image; and image recognition for recognizing the photographed object as an image by setting a detection area based on the position of the mapped particle in the image. means for acquiring the likelihood of the generated particles based on the result of the image recognition, and updating the probability distribution based on the acquired likelihood to determine the location where the target exists. a tracking device for tracking, wherein the particle generating device sequentially generates particles based on the updated probability distribution.
(102nd Configuration) The tracking device according to the 101st configuration, wherein the particle generating means generates the particles along a plane parallel to a plane in which the object moves.
(103rd configuration) The photographing means photographs the object with a vergence stereo camera using a left camera and a right camera, and the mapping means photographs a left camera image and a right camera image photographed by the left camera and the right camera, respectively. The generated particles are mapped to an image in association with each other, the image recognition means performs image recognition on the left camera image and the right camera image, and the likelihood acquisition means performs image recognition on the left camera image. the likelihood is obtained using at least one of a first likelihood based on image recognition of the right camera image and a second likelihood based on image recognition of the right camera image, and further, based on the updated probability distribution, the left camera and the The tracking device according to the 101st or 102nd configuration, further comprising a photographing direction moving means for moving the photographing direction of the right camera in the direction of the object.
(104th configuration) The present invention is characterized by comprising a surveying means for surveying the position where the object exists based on the photographing directions of the moved left camera and right camera, and an output means for outputting the surveyed result. The tracking device according to the 103rd configuration.
(105th configuration) Wide-angle image acquisition means is provided for acquiring a left wide-angle image and a right wide-angle image from a left wide-angle camera and a right wide-angle camera, respectively, and the photographing means is configured to capture images in any direction from the acquired left wide-angle image. The left camera is configured with a virtual camera that acquires a left camera image, and the right camera is configured with a virtual camera that acquires a right camera image in an arbitrary direction from the acquired right wide-angle image, and the right camera is configured with a virtual camera that acquires a right camera image in an arbitrary direction from the acquired right wide-angle image. The direction moving means moves the photographing direction in a virtual photographing space in which the left wide-angle camera and the right wide-angle camera acquire a left camera image and a right camera image from the left wide-angle image and the right wide-angle image, respectively. The tracking device according to the 104th configuration.
(106th Configuration) The tracking device according to the 105th configuration, wherein the left wide-angle camera and the right wide-angle camera are a left omnidirectional camera and a right omnidirectional camera, respectively.
(107th configuration) The 103rd configuration is characterized in that the mapping means calculates and obtains the position of the generated particle in the left camera image and the right camera image using a predetermined mapping function. 106. The tracking device according to any one of up to 106 configurations.
(108th configuration) The photographing means points and photographs the left camera and the right camera for each of the generated particles, and the mapping means corresponds to the photographing direction of the left camera image and the right camera image. The tracking device according to any one of the 103rd to 106th configurations, wherein the tracking device acquires the position of the particle as the position of the particle.
(109th configuration) The tracking device according to the 104th configuration, further comprising a moving means that moves together with the object based on the outputted survey result.
(110th configuration) A particle generation function that generates particles to be used in a particle filter in a three-dimensional space based on a probability distribution of a position where a target exists, a photographing function that photographs the target, and a particle generating function that generates particles in the photographed image. a mapping function that maps the captured particles; an image recognition function that recognizes the photographed object as an image by setting a detection area based on the position of the mapped particle in the image; a likelihood acquisition function that acquires the likelihood of the generated particles based on the acquired likelihood; and a tracking function that updates the probability distribution based on the acquired likelihood to track the location where the target exists. The particle generation function is a tracking program that sequentially generates particles based on the updated probability distribution.

(2) Configuration of the second embodiment (201st configuration) A detection device installed on a traveling object, a building, etc. to detect a predetermined object, which includes an upper camera disposed above a predetermined horizontal plane. , a photographing means for photographing the object at a wide angle using a lower camera disposed below the horizontal plane, and an image of the photographed object as an upper camera image of the upper camera and a lower camera image of the lower camera, respectively. 1. A detection device comprising: a detection means for detecting by recognition.
(202nd configuration) Particle generation means for generating particles to be used in a particle filter in a three-dimensional space based on a probability distribution of a position where a target exists, a detection device according to claim 1, and a likelihood acquisition means. , a tracking device, wherein the imaging device of the detection device is a vergence stereo system using an upper camera disposed above a predetermined horizontal plane and a lower camera disposed below a predetermined horizontal plane. The object is photographed by a camera, and the detection means of the detection device is a mapping means for mapping the generated particles in association with an upper camera image and a lower camera image respectively photographed by the upper camera and the lower camera. Then, a detection area is set in the upper camera image and the lower camera image based on the respective positions of the mapped particles in the upper camera image and the lower camera image, and the captured object is detected in the upper camera image. and an image recognition unit that recognizes the images using the lower camera image, and the likelihood acquisition unit includes a first likelihood based on image recognition of the upper camera image and a first likelihood based on image recognition of the lower camera image. 2, and the tracking means updates the probability distribution based on the acquired likelihood, thereby determining the position where the target exists. A tracking device characterized in that the particle generating means sequentially generates particles based on the updated probability distribution.
(203rd Configuration) The tracking device according to claim 2, wherein the particle generating means generates the particles along a plane parallel to a plane in which the object moves.
(204th configuration) The apparatus further comprises a photographing direction moving means for moving the photographing directions of the upper camera and the lower camera toward the object based on the updated probability distribution. The tracking device according to item 3.
(205th configuration) The present invention is characterized by comprising a surveying means for surveying the position where the object exists based on the photographing directions of the moved upper camera and lower camera, and an output means for outputting the surveyed result. The tracking device according to claim 4.
(206th configuration) Wide-angle image acquisition means for acquiring an upper wide-angle image and a lower wide-angle image from an upper wide-angle camera disposed above a predetermined horizontal plane and a lower wide-angle camera disposed below, respectively. , the photographing means configures the upper camera with a virtual camera that acquires an upper camera image in an arbitrary direction from the acquired upper wide-angle image, and also configures the upper camera with a virtual camera that acquires an upper camera image in an arbitrary direction from the acquired lower wide-angle image. The lower camera is configured with a virtual camera that acquires the images, and the photographing direction moving means is configured to cause the upper camera and the lower camera to obtain upper camera images and lower camera images from the upper wide-angle image and the lower wide-angle image, respectively. The tracking device according to any one of claims 2 to 5, wherein the tracking device moves in the shooting direction in a virtual shooting space to be acquired.
(207th Configuration) The tracking device according to claim 6, wherein the upper wide-angle camera and the lower wide-angle camera are an upper omnidirectional camera and a lower omnidirectional camera, respectively.
(208th configuration) The mapping means calculates and obtains the position of the generated particle in the upper camera image and the lower camera image using a predetermined mapping function. A tracking device according to any one of claims up to claim 7.
(209th configuration) The photographing means photographs each of the generated particles by pointing the upper camera and the lower camera, and the mapping means corresponds to the photographing direction of the upper camera image and the lower camera image. The tracking device according to any one of claims 2 to 7, characterized in that the position of the particle is acquired as the position of the particle.
(210th configuration) The tracking according to any one of claims 2 to 9, further comprising a moving means that moves together with the object based on the outputted survey result. Device.
(211th Configuration) The tracking device according to any one of claims 2 to 10, wherein the upper camera and the lower camera are arranged on a vertical line.
(212th configuration) A detection program that causes a computer to function as a detection device that is installed on a traveling object, a building, etc. and detects a predetermined target, and includes an upper camera disposed above a predetermined horizontal plane; A shooting function that takes a wide-angle picture of the object using a lower camera disposed below the horizontal plane, and image recognition of the photographed object using an upper camera image of the upper camera and a lower camera image of the lower camera, respectively. A detection program that is characterized by a detection function that performs detection by detecting objects, and a detection program that is realized by a computer.
(213th configuration) A particle generation function that generates particles to be used in a particle filter in a three-dimensional space based on the probability distribution of the position where a target exists, an upper camera disposed above a predetermined horizontal plane, and a lower camera. a photographing function for photographing the object with a convergence stereo camera using a lower camera disposed in the camera; and a photographing function for photographing the object with a convergence stereo camera using a lower camera disposed in A detection area is set in the upper camera image and the lower camera image based on a mapping function for mapping and the respective positions of the mapped particle in the upper camera image and the lower camera image, and the photographed object is detected. an image recognition function that recognizes images in the upper camera image and the lower camera image, a first likelihood based on image recognition of the upper camera image, and a second likelihood based on image recognition of the lower camera image; a likelihood acquisition function that acquires the likelihood of the generated particles using at least one of them; and a tracking function that updates the probability distribution based on the acquired likelihood to track the location where the target exists. , is realized by a computer, and the particle generation function sequentially generates particles based on the updated probability distribution.

1 Tracking device 2 CPU
3 ROM
4 RAM
5 GPU
6 Control unit 7 Drive device 8 Subject 9 Spherical camera 10 Storage unit 11 Photography unit 12 Tracking robot 15 Housing 16 Rear wheel 17 Front wheel 20 Housing 21 Rear wheel 22 Front wheel 25 Housing 26 Propeller 30 Spherical object 31 Virtual camera 32 Circular area 33 Target 35, 36 Camera 37 Photographing area 41, 42, 43 Particles 51, 52 Particles 61, 62 Detection area 71, 81, 82 Camera image 101 Image 102 Cell 106, 107 Histogram 109, 110, 111 Vector

Claims (9)

Particle generating means for generating particles to be used in a particle filter in a three-dimensional space based on a probability distribution of a position where an object exists;
Photographing means for photographing the object;
Mapping means for mapping the generated particles onto the photographed image;
image recognition means for image recognition of the photographed object by setting a detection area based on the position of the mapped particle in the image;
likelihood acquisition means for acquiring the likelihood of the generated particles based on the result of the image recognition;
tracking means for tracking the location of the object by updating the probability distribution based on the acquired likelihood;
Equipped with
The tracking device is characterized in that the particle generating means sequentially generates particles based on the updated probability distribution along a plane parallel to the plane in which the object moves . The photographing means photographs the object with a vergence stereo camera using a left camera and a right camera,
The mapping means maps the generated particles in association with a left camera image and a right camera image taken by the left camera and the right camera, respectively,
The image recognition means performs image recognition on the left camera image and the right camera image, respectively,
The likelihood obtaining means obtains the likelihood using at least one of a first likelihood based on image recognition of the left camera image and a second likelihood based on image recognition of the right camera image,
Furthermore,
Photographing direction moving means for moving the photographing directions of the left camera and the right camera in the direction of the object based on the updated probability distribution;
The tracking device according to claim 1, further comprising: a surveying means for surveying a position where the target exists based on the photographing directions of the moved left camera and right camera;
an output means for outputting the survey results;
3. The tracking device according to claim 2, further comprising: comprising wide-angle image acquisition means for acquiring a left wide-angle image and a right wide-angle image from the left wide-angle camera and the right wide-angle camera, respectively;
The photographing means configures the left camera with a virtual camera that acquires a left camera image in an arbitrary direction from the acquired left wide-angle image, and also configures the left camera as a virtual camera that acquires a left camera image in an arbitrary direction from the acquired right wide-angle image. Configure the right camera with the virtual camera to be acquired,
The photographing direction moving means moves the photographing direction in a virtual photographing space in which the left wide-angle camera and the right wide-angle camera acquire a left camera image and a right camera image from the left wide-angle image and the right wide-angle image, respectively. ,
The tracking device according to claim 3 , characterized in that: The tracking device according to claim 4, wherein the left wide-angle camera and the right wide-angle camera are a left omnidirectional camera and a right omnidirectional camera, respectively. The mapping means calculates and obtains the positions of the generated particles in the left camera image and the right camera image using a predetermined mapping function. A tracking device according to any one of claims. The photographing means photographs each of the generated particles by pointing the left camera and the right camera,
The method according to any one of claims 2 to 5 , wherein the mapping means acquires positions of the left camera image and the right camera image corresponding to the photographing direction as the positions of the particles. A tracking device according to the claims. A moving means that moves together with the object based on the outputted survey results,
The tracking device according to claim 3, further comprising a tracking device. a particle generation function that generates particles to be used in a particle filter in a three-dimensional space based on the probability distribution of the position where the target exists;
a photographing function for photographing the object;
a mapping function that maps the generated particles to the captured image;
an image recognition function that recognizes the photographed object as an image by setting a detection area based on the position of the mapped particle in the image;
a likelihood acquisition function that acquires the likelihood of the generated particles based on the result of the image recognition;
a tracking function that tracks the location of the target by updating the probability distribution based on the acquired likelihood;
realized by computer,
The particle generation function is a tracking program that sequentially generates particles based on the updated probability distribution along a plane parallel to the plane in which the object moves .

Images