JP2006276985A - Method and device for recognizing object - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method for recognizing an object and a device for recognizing the object using the method of stably recognizing the shape of the object with a small computation quantity even if data of objects to be detected is mixed. <P>SOLUTION: A method for recognizing an object to detect surface shape information of objects present around a mobile object and recognize surface shapes of the objects performs a sample extracting process #1 of extracting an arbitrary sample among a sample group constituting the surface shape information; a shape model setting process of setting a shape model on the basis of the extracted sample; a degree of matching computing process of computing a degree of matching of the sample group to the shape model; and a contour shape decision process of deciding the contour shapes on the basis of the degree of the matching. The extracting process #1 has a schematic shape determining process #14 of determining whether a prescribed shape can be formed as a shape model from the extracted sample, and if it is determined that the prescribed shape cannot be formed, the extraction process #1 extracts a new sample. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to an object recognition apparatus and an object recognition method for recognizing a contour shape of an object existing around a moving body.

  As such an apparatus, there exists an obstacle detection apparatus as described in the patent document 1 shown below. This device detects an obstacle existing around a vehicle (moving body) and issues an alarm. The present invention is made in contrast to the conventional apparatus configured to issue an alarm only when the distance between the vehicle and the obstacle is simply measured and shorter than a predetermined distance. That is, the warning based on the distance alone is made in view of the problem that it is difficult for the driver to determine which object around the vehicle is an obstacle. According to this, a plurality of obstacle detection sensors are mounted on the vehicle, and the distance to the obstacle is calculated. Whether the shape of the obstacle is a straight line (flat plate shape) or a circular shape (convex surface shape) is estimated from the obtained calculation results and displayed. By comprising in this way, it alert | reports using the distance with an obstruction, and the shape of an obstruction.

JP 2003-194938 A (page 2-3, FIG. 1-7)

  The above known technique is useful for the user in that it also estimates the shape of the obstacle. However, in actual measurement, detection data that detects objects other than the detection target object (obstacle) is often mixed. And since detection data other than a detection target object act as a noise component, it is thought that it becomes a cause of misrecognition at the time of shape estimation of a detection target object. That is, it cannot be said that the stability in detecting an object to be detected such as an obstacle is sufficient. In general, if such a noise removal function is provided, the amount of calculation increases, and accordingly, the processing time increases and the scale of the apparatus increases.

  The present invention has been made in view of the above problems, and uses an object recognition method capable of stably recognizing the shape of an object with a small amount of calculation even when data to be detected is mixed, and this method. The object is to provide an object recognition device.

The feature of the object recognition method according to the present invention for achieving the above object is as follows:
A method for detecting surface shape information of an object existing around a moving body and recognizing the contour shape of the object,
A sample extraction step of extracting an arbitrary sample from a sample group constituting the surface shape information;
A shape model setting step for determining a shape model based on the extracted sample;
A degree of coincidence calculation step of calculating the degree of coincidence of the sample group with respect to the shape model;
And a contour shape determination step for determining the contour shape based on the degree of coincidence,
The sample extraction step includes a schematic shape determination step for determining whether or not a predetermined shape can be formed as the shape model from the extracted sample. If it is determined that a predetermined shape cannot be formed, the sample extraction step is performed again. The point is to extract.

  The present invention relates to a method for detecting surface shape information of an object and recognizing the contour shape of the object based on the surface shape information. Here, the surface shape information is information indicating the shape of the surface of the object viewed from the moving body. Various means can be used for detecting the surface shape information as described below. A reflective sensor using radio waves or ultrasonic waves may be used, or an image sensor or camera (regardless of moving images or still images) that obtains image data using visible light, infrared light, or the like. Also good.

  Then, the contour shape is recognized from the sample group constituting the surface shape information. Here, the sample group refers to a collection of individual data constituting the surface shape information. The individual data is information corresponding to each location obtained by receiving a signal reflected at each location of the obstacle when, for example, a reflective sensor is used. When image data is used, data obtained by various image processing such as edge extraction and three-dimensional conversion can be used. Thus, regardless of the detection method, data representing the surface shape of the object is treated as a sample, and a collection of the samples is referred to as a sample group.

  In the sample extraction process, some samples are extracted arbitrarily (randomly) from this sample group. Based on these extracted samples, a shape model is determined in the shape model setting step. When determining this shape model, it may be calculated geometrically from the extracted sample, or a method in which a plurality of templates are prepared in advance and applied to the optimum one may be used. Then, in the coincidence degree calculation step, the degree of coincidence of how much the entire sample group matches the shape model is calculated. Based on the calculation result, it is determined whether or not the embodied shape model is suitable for the sample group. Based on the degree of coincidence, the contour shape is determined in the contour shape determination step.

  That is, when a sample with noise characteristics is included in an arbitrarily extracted sample, the degree of coincidence between the determined shape model and the sample group is low. Therefore, it can be determined that this shape model does not match the sample group. When the shape model is determined without including a noise sample, the degree of coincidence increases. Therefore, it can be determined that the shape model matches the sample group. In this way, the contour shape of the target object can be recognized by removing a noisy sample with a small amount of calculation.

  The shape model is determined from arbitrarily extracted samples having a much smaller number of samples than the sample group. Therefore, the amount of calculation required when extracting a sample or determining a shape model is small. Therefore, the calculation time is short and the apparatus is not scaled up. In addition, the degree of coincidence of the sample group with respect to the shape model can be calculated geometrically using the coordinates in the space of each sample. Therefore, the degree of coincidence can be calculated with a small amount of calculation. Furthermore, since these calculation amounts are small, an increase in the total calculation amount can be suppressed even if different shape models are repeatedly determined and the degree of coincidence is calculated. As a result, the contour shape can be recognized with high accuracy.

  Furthermore, the sample extraction step includes a schematic shape determination step for determining whether or not a predetermined shape can be formed as a shape model from the extracted sample. If it is determined that the predetermined shape cannot be formed, the sample is extracted again without going to the shape model setting step, the coincidence degree calculating step, and the contour shape determining step. Therefore, the process using the sample in which the shape model expected to have a low degree of coincidence can be terminated on the upstream side. As described above, the object recognition method according to the present invention can recognize a contour shape of a target object by removing a noisy sample with a small amount of calculation. Here, the object extraction can be realized in a smaller amount of calculation and in a shorter time by further providing a rough shape determination step in the sample extraction step.

  As described above, according to this feature, it is possible to perform object recognition capable of stably recognizing the shape of an object with a small amount of calculation even when non-detection target data (noise) is mixed.

  Here, the predetermined shape is one continuous convex shape formed by all of the extracted samples, and in the approximate shape determination step, can the predetermined shape be formed by the success or failure of the convex shape? It is preferable to determine whether or not.

  When the surface shape of an object to be recognized by applying the present invention has a gentle bulge, this can be approximated to one continuous convex shape. The continuous convex shape can be determined by simple evaluation of the extracted specimen. By providing the rough shape determination process as described above, waste of downstream processes (such as a shape model setting process, a matching degree calculation process, and a contour shape determination process) can be eliminated. Therefore, if the processing load is heavy or the processing takes time, the meaning of providing this step is reduced. However, if only the success or failure of the convex shape is determined, it can be realized by a simple evaluation, so that the success or failure of the predetermined shape can be determined at a high speed without increasing the calculation load.

  Further, it is preferable that the approximate shape determining step determines the success or failure of the convex shape by comparing the coordinate values of the extracted samples.

  The coordinate value of the sample indicates the location of each sample. And the shape model formed based on the extracted specimen is obtained by connecting each location or its vicinity. Therefore, the relative relationship between the extracted samples can be known by comparing the coordinate values indicating the location of each sample. For example, the coordinate value of the other coordinate axis is examined according to the order of the coordinates in the direction of one coordinate axis of the orthogonal coordinates. Here, it is assumed that the coordinate value of one coordinate axis advances in one direction of the axis (for example, the X axis). At this time, it is assumed that the coordinate value of the other coordinate axis (for example, the Y axis) has a change point where the coordinate value of the one axis (X axis) changes from rising to falling or falling to rising. . In this case, it can be determined that the convex shape is established with the change point as the apex portion.

  Further, it is preferable that the rough shape determination step determines the success or failure of the convex shape by comparing the slopes of straight lines obtained by connecting two adjacent samples of the extracted sample.

  The slope of the straight line obtained by connecting two adjacent samples can be obtained by simple calculation. When a convex shape can be formed by the extracted sample, this inclination changes in a certain direction. For example, the absolute value of the slope changes from a large value to a small value, and the sign is inverted to change from a small value to a large value. Considering the sign, it will change continuously to a large slope or to a small slope. The slope of a straight line connecting two points can be calculated by a simple calculation, and the magnitudes of the slopes can be compared by a simple calculation. Therefore, the success or failure of the convex shape can be easily determined.

  In addition, the sample extraction step includes a random number generation step, and the sample is extracted based on the generated random number, and the schematic shape determination step includes a rearrangement step of rearranging the random numbers in numerical order. It is preferable to determine success or failure of the convex shape by confirming the continuity of the extracted samples according to the rearranged order.

  When samples are extracted from a sample group using random numbers, each sample has a number (numerical value) associated with the random number as an attribute. This number is assigned with a certain rule such as the detection order of surface shape information and the order of coordinate positions. Random numbers are literally irregular numbers. Therefore, the samples extracted by being instructed by random numbers are not in the order along the surface shape information as they are. Therefore, when it is attempted to determine whether or not the arrangement of the samples can be a convex shape, it is necessary to rearrange the samples from the coordinate values of each sample, or to perform determination while considering the coordinate values. However, in the above method, since the random numbers are rearranged in numerical order, the samples associated with the random number numerical values are also arranged based on certain rules such as the detection order of surface shape information and the order of coordinate positions. As a result, for example, when comparing the coordinates of adjacent samples or calculating the inclination, the target sample can be identified immediately. As a result, the success or failure of the convex shape can be determined at high speed with a light calculation load.

In addition, the characteristic configuration of the object recognition device according to the present invention for achieving the above object is as follows:
Comprising object detection means for detecting surface shape information of the object and shape recognition means for recognizing the contour shape of the object, and recognizing the contour shape of the object existing around a moving body,
The shape recognition means includes
A sample extraction means for extracting an arbitrary sample from a sample group constituting the surface shape information;
Shape model setting means for determining a shape model based on the extracted sample;
A degree of coincidence calculating means for calculating the degree of coincidence of the sample group with respect to the shape model;
Contour shape determining means for determining the contour shape based on the degree of coincidence,
The sample extraction means has a schematic shape determination means for determining whether or not a predetermined shape can be formed as the shape model from the extracted sample, and when it is determined that a predetermined shape cannot be formed, the sample extraction means The point is to extract the sample.

  Provided is an object recognition device capable of stably recognizing the shape of an object even when data for non-detection is mixed with a small amount of calculation and a short calculation time, as in the object recognition method according to the present invention. Can do. Further, since the amount of calculation is small, the scale of an arithmetic device such as a microcomputer or an electronic circuit can be suppressed. Naturally, this object recognition apparatus can be provided with the functions and effects described regarding the object recognition method, as well as all the additional features and effects.

[System configuration example]
DESCRIPTION OF EXEMPLARY EMBODIMENTS Hereinafter, a preferred embodiment of the invention will be described with reference to the drawings, taking as an example a case where a vehicle recognizes another vehicle. As shown in FIG. 1, a distance sensor 1 (object detection means) directed to the side is mounted on a vehicle 10 as a moving body. The distance sensor 1 is, for example, a point sensor, that is, a single beam sensor or a sonar using ultrasonic waves. The vehicle 10 measures the distance to the parked vehicle 20 by the distance sensor 1 when passing by another vehicle 20 parked and stopped (hereinafter referred to as a parked vehicle) in the X direction in the figure. The parked vehicle 20 corresponds to the object of the present invention. In FIG. 1, for the sake of simplicity, the distance sensor 1 is provided only on the left side of the vehicle 10, but it may be provided on both sides as a matter of course.

  The distance sensor 1 measures the distance from the parked vehicle 20 according to the movement of the vehicle 10. The surface shape information of the parked vehicle 20 obtained in this way is discrete data corresponding to the moving distance of the vehicle 10. Note that “according to the moving distance” of the vehicle 10 also includes the meaning of “according to a predetermined time interval”. For example, when the vehicle 10 moves at a constant speed, if it is measured according to a predetermined time interval, it is measured according to the moving distance. The moving speed, moving distance, and moving time of the moving body 10 are determined linearly. Therefore, any method may be used as long as the surface shape information can be obtained almost uniformly as a result. In this way, the vehicle 10 acquires surface shape information of the object (corresponding to an “object detection step” described later).

  The distance sensor 1 may include an accompanying sensor such as a timer that measures the movement time, an encoder that measures the movement distance, and a rotation sensor that measures the movement speed. In addition, these sensors may be provided separately to obtain information.

  FIG. 2 is a schematic block diagram of the object recognition apparatus according to the embodiment of the present invention. In FIG. 2, the shape recognition unit 2 (shape recognition means) is configured by an electronic circuit such as a microcomputer 2A. Each processing unit in the shape recognition unit 2 does not necessarily indicate a physically different electronic circuit, but indicates a processing unit as a function. For example, the case where different functions are obtained by executing different programs by the same CPU is also included.

  As shown in FIG. 2, the surface shape information measured by the distance sensor 1 is input to the shape recognition unit 2. The input surface shape information is mapped to a two-dimensional plane with the X and Y directions shown in FIG. 1 as axes, and is stored in the specimen storage unit 2a. The sample storage unit 2a is composed of a memory. In the present embodiment, a form incorporated in the microcomputer 2A is shown. Of course, a so-called external form may be used by using a memory separate from the microcomputer 2A. Also, other storage media such as a register and a hard disk may be used regardless of whether they are internal or external.

  Hereinafter, the surface shape information of the object (parked vehicle 20) existing around the moving body (vehicle 10) is detected (object detection step), and the contour shape of the object is recognized (shape recognition step). explain.

[Object detection process]
First, the object detection process will be described. As shown in FIG. 3, the surface shape information S on the parked vehicle 20 is measured by the distance sensor 1. The surface shape information is measurement data obtained discretely in a form along the outer shape of the bumper portion of the parked vehicle 20 in the present embodiment. Here, a group of these discretely obtained data is referred to as a sample group S (Large S). The sample group S is a data set that is a recognition target of the contour shape. Moreover, the data of each point which comprises a data set is called the sample s (small es).

  The sample group S is mapped on XY two-dimensional orthogonal coordinates in the sample storage unit 2a as shown in FIG. For ease of explanation, not all samples s are shown in the figure. Among the samples shown in FIG. 4, a sample s indicated by a black dot is referred to as an inlier, and a sample s indicated by a white dot is referred to as an outlier. In the figure, samples s1, s13, etc. are inliers, and samples s2, s7, s10 are outliers. Although details will be described later, the inlier is a sample constituting the contour shape of the parked vehicle 20. The outlier is a so-called noise sample that deviates from the contour shape of the parked vehicle 20.

[Shape recognition process (1)]
Hereinafter, the procedure (shape recognition process) for recognizing the contour shape of the parked vehicle 20 from the obtained sample group S will be described using the flowcharts shown in FIGS. 7 and 8 in addition to the block diagram of FIG.

  The sample extraction unit 2b extracts several arbitrary samples si (i is a sample number) from the sample group S (samples s1 to s13) (sample extraction process, FIG. 7 # 1). Which sample s is extracted is determined randomly. Preferably random numbers are used. For example, a random number generator (not shown) is provided in the microcomputer 2A to generate a random number (random number generation step, FIG. 8 # 11). Alternatively, the sample number may be determined by a random number generation program executed by the microcomputer 2A. Then, a sample si having the generated random number as a sample number is extracted (FIG. 8 # 13).

  The number of samples to be extracted varies depending on the target shape to be recognized. For example, it is 2 points when recognizing a straight line, and 5 points if it is a quadratic curve. In the present embodiment, the bumper shape of the parked vehicle 20 is approximated to a quadratic curve, and five points are extracted (see FIG. 8 # 11). The collection of individual data and the sample s extracted in this way is a subset as a concept corresponding to the data set.

  Subsequently, the shape model setting unit 2c determines a shape model based on this subset (a collection of randomly extracted samples s) (shape model setting step, FIG. 7 # 2).

  FIG. 5 is an explanatory diagram for calculating the degree of coincidence between the shape model L (first shape model L1) and the sample group S determined from the sample si arbitrarily extracted from the sample group S shown in the scatter diagram of FIG. . This first shape model L1 is determined based on the five samples s1, s1, s5, s8, s11, and s13. The shape model L can be easily obtained by linear calculation with a light calculation load. Alternatively, several types of template shapes may be prepared in advance, and an optimum one may be selected from these template shapes.

  Further, as shown in FIG. 5, points separated by a predetermined distance in both directions orthogonal to the tangent line of the shape model L are connected along the shape model L to define dotted lines B1 and B2. A portion sandwiched between the dotted lines B1 and B2 is an effective range W.

  Next, the coincidence calculation unit 2d calculates the coincidence between the determined shape model L and the sample group S. Specifically, the degree of coincidence is calculated depending on how much each sample si constituting the sample group S is included in the effective range W determined as described above (the degree of coincidence calculation step, FIG. 7 # 3). ).

  The effective range W for the first curve model L1 shown in FIG. 5 includes all the samples s except for the outliers of the samples s2, s7, and s10. Therefore, the degree of coincidence of the first shape model L1 with the sample group S is 77% (10/13). That is, it can be said that the first shape model L1 obtained an agreement (consensus) with a high support rate (77%) by each sample s constituting the sample group S.

  Next, the main arithmetic unit 2e determines whether or not the degree of coincidence exceeds a predetermined threshold (determination step, FIG. 7 # 4). When the threshold value is exceeded, the shape model (first shape model L1) determined from the extracted subset is recognized as a recognition result (authorization process, FIG. 7 # 5). That is, the first shape model L1 is the contour shape. For example, when the threshold is set to 75%, the first shape model L1 is set as the contour shape. When the threshold value is not exceeded, the process returns to the process # 1 in the flowchart of FIG. 7, and another sample s is extracted again to form a new subset, and the same process is performed. If the threshold value is not exceeded even if the processes # 1 to # 4 are repeated a plurality of times, it is determined that there is no target object (such as the parked vehicle 20). This number may be defined in advance.

  In the present embodiment, the total number of samples s constituting the sample group S is 13 in order to facilitate understanding. The threshold value (75%) is also a value for facilitating the description of the present embodiment. Accordingly, neither the number of samples nor the determination threshold value for the degree of coincidence is a value that limits the present invention. For example, if the number of samples is large, the number of inliers relative to the outlier is relatively large, and a threshold value higher than the above example can be set.

  In the shape model L (second shape model L2) shown in FIG. 6, samples s2, s4, s7, s10, and s13 are extracted as subsets. As described above, the samples s2, s7, and s10 are so-called noise samples that deviate from the contour shape of the parked vehicle 20. Therefore, when viewed from the contour shape of the parked vehicle 20, it is a specimen that should be an outlier. Therefore, as shown in FIG. 6, there are many samples s that are out of the effective range W for the second shape model L2 determined based on the subset including these samples s2, s7, and s10. When the degree of coincidence is calculated by the same method as that for the first shape model L1, the degree of coincidence is 38% (5/13). That is, the second shape model L2 does not have an agreement (consensus) with a high support rate by the samples s constituting the sample group S.

  When the two shape models L1 and L2 are extracted, the contour shape that is the recognition result is the first shape model L1. In determining the first shape model L1, the samples s2, s7, and s10, which are noise samples s, are unused. These noise samples are treated as outliers and removed. That is, with the small amount of calculation as described above, even if non-detection target data (outlier) is mixed, it can be removed and the shape of the object can be recognized stably.

[Comparison with conventional shape recognition methods]
Various methods for calculating the contour shape from the specimen S without using such a method have been conventionally proposed. One of them is the least square method. In the least square method, the shape is calculated by using all the samples s in the data set, with each sample s having the same weight. As a result, a contour shape different from the original shape is recognized under the influence of the above-described outlier (sample s2 and the like). It is also possible to reconfirm the degree of coincidence with the entire data set after recognizing the contour shape. However, the calculation load of the least square method itself is relatively heavy, and if the shape recognition by the repeated least square method is repeatedly performed based on the result of this reconfirmation, the calculation load is further increased.

As another method, there is a method of using a Hough transform suitable for straight line recognition. As is well known, the Hough transform utilizes the property that a straight line existing on an orthogonal coordinate (for example, an XY plane) intersects at one point on a polar coordinate (ρ-θ space). The conversion formula is
ρ = X · cosθ + Y · sinθ
It is. As can be understood from the above formula, if the range of ρ and θ is expanded or a fine resolution is obtained in the polar coordinate space, the amount of calculation increases accordingly. That is, the memory as the primary storage means is required to have a large capacity and the number of calculations is increased.

  Compared to these conventional calculations, the method of “calculating the degree of coincidence of the sample group S with the shape model L determined based on the sample s arbitrarily extracted from the sample group S constituting the surface shape information” of the present invention is The amount is small and the required memory capacity is also small.

  In the above description, the degree of coincidence between the shape model L and the sample group S is examined. If the degree of coincidence exceeds a predetermined threshold value, the shape model L is used as a recognition result. That is, the shape model L that first exceeds the threshold value becomes the recognition result as it is. However, the shape model L may not be immediately recognized as a recognition result when the threshold value is simply exceeded, and a plurality of shape models L may be evaluated. A specific procedure will be described below.

[Shape recognition process (2)]
FIG. 9 is a flowchart for explaining another example of the method for recognizing the contour shape from the sample group shown in the scatter diagram of FIG. In this method, a subset is extracted a plurality of times to define a shape model L, and the shape model L having the highest degree of coincidence among them is set as a recognition result. Hereinafter, this method will be described with reference to FIG. However, processes # 1 to # 4 are the same as those in the flowchart shown in FIG.

  In the method shown in FIG. 9, since the subset is extracted repeatedly a plurality of times, the number of repetitions is temporarily stored. At the start of the shape recognition process, first, the number of repetitions temporarily stored is cleared (initialization process, FIG. 8 # 0). Thereafter, similarly to the method shown in FIG. 7, the sample s is randomly extracted from the sample group S in the sample extraction step (# 1) to create a subset. Next, in the shape model setting step (# 2), the shape model L is determined based on this subset. Then, the coincidence degree between the shape model L and the sample group S is calculated in the coincidence degree calculating step (# 3), and whether or not the coincidence degree exceeds a predetermined threshold value in the determining step (# 4). Determine whether.

  As a result of the determination, if the threshold value is exceeded, the previously determined shape model L and the degree of coincidence with the shape model L are stored in a temporary storage unit (not shown) (storage step, # 41). Next, it is determined whether or not the degree of coincidence exceeds a second threshold set to a higher value (second determination step, # 44). The second threshold value is a high degree of coincidence that is almost close to a perfect score, for example. This step is provided because it is not necessary to repeatedly extract the subset when the degree of coincidence is very high. If it is determined in the second determination step (# 44) that the second threshold value is exceeded, the shape model L determined in the shape model setting step # 2 (= the shape stored in the storage step (# 41)) Model L) is the recognition result (certification process, # 5).

  If it is determined that the threshold value (second threshold value) is not exceeded in the determination step (# 4) or the second determination step (# 44), the process proceeds to the counting step (# 42). That is, since the evaluation for one shape model L is completed, the number of repetitions is incremented.

  Next, it is determined whether or not the number of repetitions has reached a predetermined number (may be exceeded or not) (leave determination step, # 43). If the predetermined number of times has not been reached, the process returns to the sample extraction process (# 1), and the following determination process (# 4) is performed to evaluate a new shape model L. If the predetermined number of times has been reached, the shape model L having the highest degree of coincidence is selected from the stored shape models L, and this is used as the contour shape as the recognition result (certification process, # 51). ). Here, in the determination step (# 4), when there is no thing exceeding the coincidence threshold, it is determined that there is no corresponding in the certification step (# 51).

  As described above, in both the method shown in FIG. 7 and the method shown in FIG. 8, the shape model L determined based on the subset is recognized as the contour shape. In general, it is considered that the shape model L determined based on a small number of samples does not reproduce an accurate contour shape. However, in the present invention, the degree of coincidence between the shape model L and all samples in the sample group S is evaluated. Therefore, it can be considered that the shape model L can reproduce (recognize) the contour shape almost accurately. Thus, the ability of the shape model L determined from the small number of samples constituting the subset to reproduce the contour shape greatly contributes to a reduction in the amount of calculation.

  As described above, acknowledging the shape model L as a contour shape as a recognition result as it is greatly contributes to a reduction in the amount of calculation. However, the present invention is not limited to this. In the case where the computing means has a sufficient capacity such as the microcomputer 2A, the contour shape may be recalculated.

  For example, if the shape model L whose coincidence exceeds a threshold value is used as a reference, the samples s constituting the sample group S can be defined as inliers and outliers. In the certification process, this inlier and outlier are certified. Then, the shape is recalculated using the least square method or the like for all samples s certified as inliers (recalculation step). As described above, the least square method may not be able to reproduce the shape correctly under the influence of the noisy sample s. However, in this recalculation process, since the noisy sample s is removed as an outlier, an accurate contour shape can be reproduced.

[Improved sampling process]
FIG. 10 is a flowchart for explaining an example in which the sample extraction process (# 1) shown in FIG. 8 is improved. As shown in FIG. 10, in the random number generation step (# 11), five random numbers corresponding to the sample number of the sample s of the sample group S are generated. Next, these five random numbers are rearranged in the order of sample numbers (rearrangement process, # 12). The generated random numbers appear in an arbitrary order due to the nature of the random numbers. Therefore, before extracting the sample s of the sample number corresponding to the random number, the random number is first rearranged in this rearrangement step (# 12). Then, five samples s having sample numbers corresponding to the random numbers rearranged in ascending order or descending order are extracted (# 13).

  Next, it is determined whether or not the shape model L formed by the extracted five samples s can have a predetermined shape (outline shape determination step, # 14). As described above, when recognizing the contour shape of the bumper portion of the parked vehicle 20, the bumper shape is approximated to a quadratic curve. The quadratic curve is, for example, an ellipse or a parabola, and the ellipse or parabola has a continuous convex shape. Therefore, when a continuous convex shape can be formed by the extracted five samples s, the probability that a quadratic curve such as an ellipse or a parabola can be formed by these samples s becomes high. In the approximate shape determination step (# 14), as shown in FIG. 10, predetermined preprocessing is performed in the preprocessing step (# 14a) to determine the possibility of forming a convex shape (# 14b). A series of sample extraction steps (# 1) including the outline shape determination step (# 14) will be described in detail based on the specific examples shown in FIGS.

  FIG. 11 is a diagram showing the sample group S constituting the surface shape information and the coordinates of each sample s of the sample group S. FIG. 11A shows an example in which a sample group S is configured by eight samples s for easy explanation. The samples s101 to s108 are specified on the XY plane, and each is assigned an index as a sample number. In this example, Index is given as an integer of 1 to 8 for the samples s101 to s108. Each sample s has a coordinate value (x, y) on the XY plane. The coordinate values (x, y) corresponding to the sample number (Index) are as shown in FIG.

  FIG. 12 is an explanatory diagram showing an example in which the sample s is extracted from the sample group S shown in FIG. 11 by the sample extraction step (# 1) shown in FIG. FIG. 12A shows the random number generated by the random number generation step (# 11). As described above, since random numbers are arbitrarily generated, they are arranged as 4, 6, 3, 1, 7 regardless of ascending order / descending order as shown in the figure. FIG. 12B shows random numbers after being sorted (sorted) by the sorting process (# 12). Here, they are rearranged in ascending order of 1, 3, 4, 6, 7 in correspondence with the arrangement from the left side to the right side (X-axis direction) in FIG. This also corresponds to the order in which the surface shape information is detected, as described with reference to FIG. FIG. 12C shows the result of extracting the sample s of the sample number (Index) corresponding to these random numbers based on the rearranged random numbers.

  Hereinafter, a specific method for determining the success or failure of the convex shape by confirming the continuity of the extracted sample s through the rearrangement step (# 12) according to the rearrangement order will be described.

[Schematic shape determination step (1)]
By rearranging the random numbers in ascending order, the five samples s designated by the random numbers are extracted in the order of the sample numbers. Here, when the X coordinate of each sample s is examined,
s101 → s103 → s104 → s106 → s107
2 ≦ 5 ≦ 9 ≦ 13 ≦ 16,
And in ascending order. Therefore, they are arranged in order in the X direction. When examining the Y coordinate,
12 ≧ 5 ≧ 3 ≦ 5 ≦ 12,
Once in descending order, then in ascending order. The evaluation of the coordinates of the sample s corresponds to the preprocessing step (# 14a) described above. From this evaluation result, it is estimated that a convex shape can be formed by s101, s103, s104, s106, and s107 with the sample s104 as a vertex (# 14b). Thus, the success or failure of the convex shape is determined by comparing the coordinate values of the extracted samples s.

[Schematic shape determination step (2)]
Further, the approximate shape of the shape model formed by the extracted specimen s can be determined by the method as described below. As shown in FIG. 13, a straight line is formed by connecting two adjacent samples s. Since the coordinate value of each sample s is known, the slope of the obtained straight line can be calculated by simple linear calculation. For example, the straight line obtained by connecting s101 and s103 is
s101 = (2,12), s103 = (5,5)
Is a straight line passing through
y = ax + b (a: slope, b: intercept)
Substituting into
12 = 2a + b, 5 = 5a + b
It becomes. Solving these simultaneous linear equations,
a = -7 / 3, b = 50/3
It becomes. This inclination a corresponds to a1 shown in FIG. Similarly, the inclinations a2, a3, and a4 are calculated between the other two points. Although intermediate calculations are omitted,
a2 = -2, a3 = 2, a4 = 7/3
It becomes. This calculation of the inclination corresponds to the preprocessing step (# 14a) described above. In the above calculation example, the solution is also shown for the intercept b. However, in the approximate shape determination step (# 14) as described later, since only the inclination is used, the intercept need not be obtained. Therefore, the pretreatment process (# 14a) is further simplified.

As described above, when the slope of a straight line connecting adjacent samples s in the X-axis direction is obtained by simple linear calculation (simultaneous linear equations), the slope is evaluated.
a1 ≤ a2 ≤ a3 ≤ a4
And the slope gradually increases. Also, when evaluating the absolute value of the slope,
| A1 | ≧ | a2 | = | a3 | ≦ | a4 |
The absolute value of the slope gradually decreases and turns in the larger direction on the way. That is, the slope gradually becomes gentler, the direction of the slope is changed before and after the sample s104, and thereafter the slope becomes gradually larger. Thereby, it can be determined that the shape is a convex shape with the vicinity of the sample s104 as a vertex (# 14b).

  In this way, if it is determined in the approximate shape determination step (# 14) that the five samples s are suitable as a subset for generating the shape model L, the sample extraction step (# 1) is skipped and the shape model is determined. The process proceeds to the setting process (# 2). FIG. 14 shows an example in which the shape model L3 is generated based on the evaluated subset through the approximate shape determination step (# 14).

[Other Embodiments of Object Recognition Apparatus]
Hereinafter, an object recognition apparatus using the above-described object recognition method will be described. FIG. 15 is a schematic block diagram showing another embodiment of the object recognition apparatus according to the present invention. It is an example in case a vehicle recognizes another vehicle similarly to the structural example shown in FIG. As shown in FIG. 15, the microcomputer 2A includes a relative arrangement calculation unit 3 (relative arrangement calculation means). As described above, the shape recognition unit 2 recognizes the contour shape of the parked vehicle 20 as viewed from the vehicle 10, that is, the bumper shape. In this recognition, since the surface shape information of the parked vehicle 20 is acquired using the distance sensor 1, distance information between the vehicle 10 and the parked vehicle 20 can also be obtained at the same time. The relative arrangement calculation unit 3 calculates the relative arrangement of the vehicle 10 and the parked vehicle 20 using the distance information and the contour shape.

  Here, the relative arrangement is a relative position between each part of the vehicle 10 and each part of the parked vehicle 20. The external shape of the vehicle 10 is known because it is its own shape. The contour shape of the parked vehicle 20 viewed from the vehicle 10 can be recognized as described above. As a result, the relative arrangement calculation unit 3 calculates the relative arrangement of the vehicle 10 and the parked vehicle 20 as shown in FIG. In FIG. 16, for easy understanding, the entire parked vehicle 20 is indicated by a dotted line. However, in actuality, a relative arrangement between the recognized contour shape E and the vehicle 10 is calculated. Of course, when the contour shape E is recognized including other places, all relative arrangements can be calculated.

  Moreover, you may display this relative arrangement | positioning on alerting means, such as a display. When a navigation system or the like is mounted on the vehicle 10, the monitor may also be used. In displaying (notifying), the outer shape of the vehicle 10 and the recognized contour shape E are displayed. Alternatively, the entire parked vehicle 20 may be represented as an illustration based on the contour shape E, and the relative arrangement relationship between the vehicle 10 and the parked vehicle 20 may be displayed.

  Further, the notification is not limited to visual display as described above, but may be notified by sound (including sound) using a buzzer, a chime, or the like. Some navigation systems have a voice guide function, and this voice guide function may also be used in the same manner as the use of a monitor.

  As shown in FIG. 15, in this example, there is shown a form in which a moving state detecting means 4 for detecting a moving state of the vehicle 10 such as a wheel speed sensor 4 a and a rudder angle sensor 4 b is provided. If the input information from the movement state detection means 4 is taken into account, the near future layout can be calculated. That is, it is possible to estimate (predict) the future relative arrangement relationship as well as know the current relative arrangement in which the contour shape E is recognized.

  The wheel speed sensor 4a is a sensor provided in each wheel portion of the vehicle 10 (front right FR, front left FL, rear right RR, rear left RL). This is, for example, a rotation sensor using a Hall IC. The steering angle sensor 4b is a sensor that detects the rotation angle of the steering of the vehicle 10 and the rotation angle of the tire. Alternatively, it may be an arithmetic device that calculates the steering angle based on the measurement results (the difference between the rotation speeds and rotation speeds of the left and right wheels) at each wheel portion of the wheel speed sensor 4a described above.

  Taking into account the movement state detected by these sensors, the relative positional relationship between the vehicle 10 and the contour E of the parked vehicle 20 is calculated. The traveling direction is estimated by the steering angle sensor 4b, and the traveling speed is estimated by the wheel speed sensor 4a. And the relative locus | trajectory of several seconds after the estimated locus | trajectory of the vehicle 10 and the outline shape E of the parked vehicle 20 and the vehicle 10 is calculated. FIG. 17 shows an example of the relative arrangement relationship between the vehicle 10 and the contour shape E of the parked vehicle 20 calculated in this way. Reference numeral 10A denotes a near future position of the vehicle 10, that is, an estimated (predicted) position.

[Other Embodiments]
In the above description, the distance sensor 1 that detects the surface shape information of the parked vehicle 20 as the vehicle 10 moves as shown in FIG. However, the object detection means according to the present invention is not limited to this. The distance sensor 1 outputs surface shape information regardless of the movement of the vehicle 10, and can be selected for each movement distance and every elapsed time in the subsequent information processing. Further, a scanning unit that scans a wide-angle area with respect to the parked vehicle 20 regardless of the movement of the vehicle 10 may be provided, and surface shape information may be detected based on the obtained scanning information. That is, not only a point sensor such as the distance sensor 1, but also a sensor that can obtain a signal (surface shape information) reflecting the shape of an object, such as a one-dimensional sensor, a two-dimensional sensor, or a three-dimensional sensor.

  FIG. 18 shows an example in which a one-dimensional sensor is used as the object detection means according to the present invention. Here, a scanning laser sensor is used as an example of a one-dimensional sensor. As shown in FIG. 18, the object (parked vehicle 20) is scanned radially from the sensor position (position of the scanning means 1a). The distance distribution can be measured by the reflection of the laser wave from each position of the object. If the azimuth angle θ when the laser wave is emitted is detected by an encoder or the like, surface shape information can be obtained in the same manner as shown in FIG. And it can map to XY orthogonal coordinate.

  As another example of the one-dimensional sensor, an ultrasonic radar, an optical radar, a radio radar, a triangulation distance meter, or the like may be used.

As the two-dimensional sensor, there is a scanning radar that can scan in the horizontal and vertical directions. By using this scan type radar, it is possible to obtain information related to the shape in the horizontal direction and the shape in the vertical direction of the target object.
As well-known two-dimensional sensors, there are image input means such as a CCD (Charge Coupled Device) and a camera using a CIS (CMOS Image Sensor). Various feature quantities such as contour information and intersection information may be extracted from image data obtained from this camera to obtain information on the surface shape.

  The same applies to the three-dimensional sensor, and for example, information regarding the shape may be obtained using image data taken in stereo.

[Other forms of use]
As described above, the embodiment of the present invention has been described with respect to the method and apparatus for recognizing the contour shape using the parked vehicle 20 as an object, and these additional features. The “object” is not limited to an obstacle such as a parked vehicle or a building, but includes various objects such as a road lane, a stop line, and a parking frame. That is, the recognition target is not limited to the contour shape of a three-dimensional object, and can be applied to shape recognition of a planar pattern.

Explanatory drawing which shows the example in case the vehicle carrying the object recognition apparatus which concerns on this invention recognizes another vehicle Schematic block diagram of an object recognition apparatus according to an embodiment of the present invention The figure which shows the result of having measured the surface shape information of the parked vehicle of FIG. Scatter chart in which the measurement results shown in FIG. 3 are mapped onto two-dimensional orthogonal coordinates FIG. 4 is an explanatory diagram for calculating the degree of coincidence between a first shape model determined from a sample arbitrarily extracted from the sample group shown in the scatter diagram of FIG. 4 and the sample group. Explanatory drawing which calculates the coincidence of the second shape model and the sample group determined from the sample arbitrarily extracted from the sample group shown in the scatter diagram of FIG. The flowchart explaining an example of the method of recognizing a contour shape from the sample group shown in the scatter diagram of FIG. FIG. 7 is a flowchart for explaining details of a process of extracting a sample from the sample group in FIG. 7 (sample extraction process). The flowchart explaining the other example of the method of recognizing a contour shape from the sample group shown in the scatter diagram of FIG. The flowchart explaining the example which improved the sample extraction process shown in FIG. The figure which shows the sample group which constitutes surface shape information, and the coordinates of each sample of the sample group Explanatory drawing which shows the example which extracts a sample from the sample group shown in FIG. 11 by the sample extraction process shown in FIG. The figure explaining the method to determine the approximate shape of the extracted sample shown in FIG. The figure which shows the example which produced | generated the shape model using the subset which determined rough shape with the method shown in FIG. Schematic block diagram showing another configuration of the object recognition apparatus according to the present invention The figure which shows an example of relative arrangement | positioning relationship between the vehicle carrying the object recognition apparatus computed by the relative arrangement | positioning calculating part of FIG. 15, and the outline shape of another vehicle The figure which shows the other example of relative arrangement | positioning relationship between the vehicle carrying the object recognition apparatus computed by the relative arrangement | positioning calculating part of FIG. 15, and the outline shape of another vehicle. The figure which shows the example (other embodiment) at the time of using a one-dimensional sensor as an object detection means based on this invention.

Explanation of symbols

# 1 Sample extraction process # 14 Outline shape determination process

Claims (6)

An object recognition method for detecting surface shape information of an object existing around a moving body and recognizing the contour shape of the object,
A sample extraction step of extracting an arbitrary sample from a sample group constituting the surface shape information;
A shape model setting step for determining a shape model based on the extracted sample;
A degree of coincidence calculation step of calculating the degree of coincidence of the sample group with respect to the shape model;
And a contour shape determination step for determining the contour shape based on the degree of coincidence,
The sample extraction step includes a schematic shape determination step for determining whether or not a predetermined shape can be formed as the shape model from the extracted sample. If it is determined that a predetermined shape cannot be formed, the sample extraction step is performed again. Recognizing object.   The predetermined shape is one continuous convex shape formed by all of the extracted samples, and the approximate shape determination step determines whether the predetermined shape can be formed by the success or failure of the convex shape. The object recognition method according to claim 1 for determining.   The object recognition method according to claim 2, wherein the approximate shape determination step determines the success or failure of the convex shape by comparing the coordinate values of the extracted samples.   The object recognition method according to claim 2, wherein the approximate shape determination step determines the success or failure of the convex shape by comparing slopes of straight lines obtained by connecting two adjacent samples extracted. The sample extraction step includes a random number generation step, and extracts the sample based on the generated random number,
The schematic shape determination step includes a rearrangement step of rearranging the random numbers in numerical order, and confirms the continuity of the extracted samples according to the order of rearrangement, thereby determining the convex shape. The object recognition method as described in any one of Claims 2-4 which determines success or failure. An object recognition device comprising: object detection means for detecting surface shape information of an object; and shape recognition means for recognizing the contour shape of the object, and recognizing the contour shape of the object existing around a moving body,
The shape recognition means includes
A sample extraction means for extracting an arbitrary sample from a sample group constituting the surface shape information;
Shape model setting means for determining a shape model based on the extracted sample;
A degree of coincidence calculating means for calculating the degree of coincidence of the sample group with respect to the shape model;
Contour shape determining means for determining the contour shape based on the degree of coincidence,
The sample extraction means has a schematic shape determination means for determining whether or not a predetermined shape can be formed as the shape model from the extracted sample, and when it is determined that a predetermined shape cannot be formed, the sample extraction means An object recognition device that extracts a sample.

Images