JP2002286416A - Obstacle-recognizing apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an obstacle-recognizing apparatus for recognizing obstacles in the all directions required for autonomously moving a mobile unit such as a robot and an automobile safely without increasing the number of sensors and without making complex a data-processing system. SOLUTION: The obstacle-recognizing apparatus has an image-outputting means that is provided at the mobile unit, an obstacle extraction means for extracting an obstacle at least in front of the mobile unit from image data in the image-outputting means, a means for calculating the relative position between the mobile unit and the obstacle, a means for extracting an amount of feature in the obstacle, and a means for storing the amount of feature corresponding to the relative position. Thereby, it is possible to grasp the presence of the obstacle existing even outside the range that can be detected by the sensor by referring to past data that are accumulated at an obstacle database even using a sensor whose detection range is limited as a simple substance as the obstacle sensor.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an obstacle recognizing device for recognizing at least an obstacle in front of a moving object from image data of image output means provided on the moving object.

[0002]

2. Description of the Related Art Ultrasonic sensors are widely used as obstacle sensors for self-propelled moving bodies such as robots and automobiles. This has the advantage that the output processing is simple and inexpensive and is suitable for use in control such as distance measurement and obstacle avoidance determination. However, the ultrasonic sensor outputs the distance information only for one point of the obstacle on which the ultrasonic wave is reflected, and the amount of information that can be output is small, and it cannot cope with a complex environment in which a plurality of objects exist, In particular, it is not good at catching obstacles three-dimensionally.

On the other hand, as obstacle sensors capable of dealing with a relatively wide range, there are known a two-lens stereoscopic system using two cameras, a two-dimensional laser radar, and the like. However, even in these cases, the detection range of a single body is limited, and if an attempt is made to detect an obstacle over the entire periphery of the moving body, the system becomes large and has to be expensive.

[0004]

SUMMARY OF THE INVENTION The present invention has been devised to solve such problems of the prior art.
Its main purpose is to remove obstacles around the surroundings necessary for a mobile object such as a robot or car to move safely and autonomously.
An object of the present invention is to provide an obstacle recognition device capable of performing recognition without increasing the number of sensors and without complicating a data processing system.

[0005]

In order to achieve the above object, according to the present invention, there is provided an image output means provided on a moving body, and at least an obstacle in front of the moving body is obtained from image data of the image output means. Obstacle extraction means for extracting
An obstacle, comprising: a relative position calculation unit for a moving object and an obstacle; a feature amount extraction unit for extracting a feature amount of the obstacle; and a storage unit for storing the feature amount in association with the relative position. A recognition device was provided.

In particular, it is preferable that the obstacle extracting means performs a cluster process on the detected data.

[0007] In this way, as an obstacle sensor,
The history of previously detected obstacle data, that is, stored in the obstacle database, even when using a single object with a limited detection range, such as a binocular stereo vision system, two-dimensional laser radar, and a small number of ultrasonic sensors By referring to the past data obtained, for example, it is possible to grasp the existence of an obstacle existing outside the detectable range of the sensor.

[0008]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention applied to an autonomously moving robot will be described below in detail with reference to the accompanying drawings.

As shown in FIG. 1, an obstacle recognition device according to the present invention includes an image output means 1 such as a binocular stereoscopic vision system for capturing images of two cameras, and at least a forward object of a moving object from image data of the cameras. Obstacle extracting means 2 for extracting an obstacle, and relative position calculating means 3 for a moving body and an obstacle
And a storage unit 5 for storing the feature amount of the obstacle in association with the relative position information.

The processing steps in this device will be described below with reference to FIG. First, the image output means 1 obtains brightness images of the obstacles 10a to 10e as shown in FIG. 3, and obtains three-dimensional distance data for each pixel (step 1). The camera image obtained in this case is represented by a perspective view in which one point at infinity is a vanishing point. As the robot moves forward, the image shown in FIG. The obstacles 10a to 10c that are continuously changed from the field of view of the camera are not displayed on the screen.

Next, self-position data of the robot is input (step 2). The position data includes, for example, the output of a moving distance sensor based on the number of rotations of a wheel, the output of a traveling direction sensor using a gyro compass,
It can be obtained by using the output of the PS system together.

The latest three-dimensional distance data obtained from the image data is subjected to, for example, histogram processing to extract a cluster of obstacles (hereinafter referred to as a current cluster) (step 3). Further, each feature amount is extracted for each current cluster (step 4). The feature amount is an average position of an obstacle in the real space, circumscribed rectangular data of a cluster on the brightness / distance image, a height, a size, a width, and a distribution (a covariance matrix or the like) in the real space. is there.

Next, using the displacement data of the robot position, the data relating to the position of the cluster data stored in the obstacle database is corrected, and the data which is far away from the robot and need not be referred to is deleted. (Step 5).

The cluster data stored in the obstacle database is as shown in FIG. In FIG. 4, reference numeral 1a denotes a list header, and reference numerals 1b to 1d denote structure data in which feature amounts of respective clusters are stored. The definition of each numerical value is shown below.

Numobj: Number of obstacles around the robot maxid: Maximum value of identification values assigned to all obstacles around the current position gmypos: X, Y, Z are positions of the robot in the real space, VX, VY are The speed of the robot calculated from the displacement from the previous time time: The updated time id: The identification value of the obstacle assigned to each obstacle sphist []: sphist [0] is the identification value of the recognized specific structure, s
phist [1 ..] is the past identification result x, y: average position on the local map sig [3]: distribution (covariance matrix) on the local map, σxx, σxy, σyy * gim: brightness image of cluster Data start address pointer (stores brightness data in a rectangle circumscribing the cluster on the image) * x, y, z: Real space data of the cluster elements constituting the cluster, X, Y, Z start address pointer ( X, Y, Z: average position of obstacles in real space SIG [5]: distribution (covariance matrix) in real space, σxx, σx in order
y, σyy, σxz, σzz relation: Identification value of whether or not the cluster is within the visual field range (determined from the camera posture of the robot and the position of the cluster) rect [4]: The circumscribed rectangle of the cluster on the brightness / range image Data, upper left x, y lower right x, y RECT [6]: circumscribed rectangular data of cluster in real space, m in order
in (X), min (Y), min (Z), max (X), max (Y), max (Z) (min (X) is
History: derect is the number of times detected in the past, count is the number of image inputs since this cluster was first detected, prob
Is the existence probability calculated by derect / count

Next, among the clusters in the obstacle database, for a cluster whose relative position to the robot is within a certain distance, a matching evaluation value with each current cluster is calculated (step 6), and the matching evaluation is performed. A current cluster having a value smaller than a predetermined value is assigned a new identification value as a new cluster and additionally registered in the obstacle database, and a current cluster having a predetermined value or more is updated as the same cluster (step 7). As the matching evaluation value, for example, the sum of the absolute values of the differences for each block that is normally used in binocular stereopsis may be used.

Next, the existence probability (prob) is calculated for each cluster in the obstacle database (step 8),
If the existence probability is less than the predetermined value, the data is deleted from the database (step 9).

The robot detects an obstacle ahead by referring to a global map which is always optimized by repeating the above flow periodically with a cycle time appropriately set according to the moving ability of the robot. Only with this, it is possible to accurately grasp the existence of obstacles around the entire area (see FIG. 5).

In the above processing flow, the data of the center position (average position X, Y in the real space) of the cluster of each obstacle and the distribution (σxx, σxy, σy) of the cluster in the real space are used.
y), an ellipse is created, a rectangle circumscribing the ellipse is created, and this is used as the base of the cube, and zmin and zmax are extracted from the pixel data (x, y, z) constituting the cluster. Is the height of the cube, so that the size of the obstacle as a solid can be easily recognized.

Next, a description will be given of a method for obtaining an example of the extent to which the data of an obstacle is to be stored as a history based on the distance accuracy of a binocular stereoscopic system.

The distance accuracy Δr of the two-lens stereoscopic vision system is generally obtained by the following equation.

Δr = Δd (ur ² / bf)

Here, Δd: parallax resolution (for example, 0.1 when calculating parallax up to 1/10 subpixel of one pixel), u: size of one pixel, r: distance from viewpoint, b: 2 The distance between the two cameras (the lines of sight of the two cameras are parallel to each other) and f: focal length.

For example, u is 9.9 μm, b is 70 mm,
If f is 6 mm and Δd is 1, the relationship between r and Δr is as shown in FIG.

Therefore, for example, when the allowable error is assumed to be 50 cm, r is about 4.6 m. Therefore, it is practically set to a radius of about 4.6 m around the robot as a history range. Become. Then, based on this value, the history range may be determined in consideration of the traveling speed and acceleration ability of the robot and the processing speed of the CPU.

[0026]

As described above, according to the present invention, it is possible to grasp an obstacle outside the detection range of the sensor only by detecting an obstacle in front of a moving object such as a robot, and to detect the moving object. A great effect can be achieved in simplifying the obstacle avoidance control at the time of autonomous movement.

[Brief description of the drawings]

FIG. 1 is a block diagram of a device configuration to which the present invention is applied.

FIG. 2 is a processing flowchart of the present invention.

FIG. 3 is a conceptual diagram of a screen displaying an image captured by a camera.

FIG. 4 is a block diagram showing contents of an obstacle database;

FIG. 5 is a conceptual diagram showing how an obstacle is grasped according to the movement of the robot.

FIG. 6 is a diagram showing the relationship between the distance accuracy Δr of the binocular stereoscopic vision system and the distance from the lens.

[Description of Signs] 1 Image output unit 2 Obstacle extraction unit 3 Relative position calculation unit 4 Feature extraction unit 5 Storage unit

Claims (2)

[Claims] 1. An image output means provided on a moving body, an obstacle extracting means for extracting at least an obstacle in front of the moving body from image data of the image output means, and a relative position between the moving body and the obstacle An obstacle recognition apparatus, comprising: a calculation unit; a feature amount extraction unit that extracts a feature amount of an obstacle; and a storage unit that stores the feature amount in association with the relative position. 2. The obstacle recognition apparatus according to claim 1, wherein the obstacle extraction unit performs a cluster process on the detected data.