JP4502733B2 - Obstacle measuring method and obstacle measuring device - Google Patents

Description

  In the present invention, a distance (inter-vehicle distance) from an own vehicle to an obstacle such as a preceding vehicle in front of the own vehicle, and a lateral width (vehicle width) of the obstacle from a captured image of the monocular camera mounted on the own vehicle in front of the own vehicle. The present invention relates to an obstacle measurement method and an obstacle measurement device that perform estimation and measurement.

  In recent years, in the automobile industry, various in-vehicle systems that apply image processing technologies such as vehicle recognition and lane recognition have been developed and commercialized, and among them, a vehicle driving support system (Adaptive Cruise Control) called ACC. ) And damage-reducing automatic braking systems, it is possible to measure at least the distance from the vehicle to the obstacle (inter-vehicle distance) for obstacles such as preceding vehicles ahead of the vehicle by image processing of the captured image. Required.

  In order to perform this obstacle measurement, in the ACC and the damage reduction automatic braking system, etc., a stereo camera has been conventionally mounted on the own vehicle, and based on the image taken in front of the stereo vehicle, the principle of triangulation is used. It has been proposed to measure distances such as the inter-vehicle distance (see Non-Patent Document 1, for example).

  However, the measurement from this stereo photographed image requires a stereo camera, is expensive, and has the disadvantage that the size of the optical system is large and large.

  Therefore, instead of a stereo camera, an inexpensive and small-sized monocular camera is mounted on the own vehicle, and a captured image of the monocular camera in front of the own vehicle (hereinafter, the captured image of the monocular camera is referred to as a monocular image) is processed. The Kalman filter process based on the detection of the peak point of the edge histogram of the edge estimates and measures the distance such as the above-mentioned inter-vehicle distance under the condition that the road gradient is constant and the vehicle width of the preceding vehicle as an obstacle is known. (For example, refer nonpatent literature 2).

  In addition, for a monocular image in front of the vehicle of a monocular video camera, first, the traveling lane of the own vehicle is detected, the horizontal edge at the bottom of the screen in the traveling lane is extracted as the shadow of the preceding vehicle, and the extracted horizontal The initial inter-vehicle distance is calculated from a proportional calculation based on the height of the image on the edge. It has also been proposed to measure and measure the inter-vehicle distance and measure the inter-vehicle distance and the vehicle width every moment (see, for example, Patent Document 1).

Junji Tsuji and Noriyuki Togawa, “Development of Stereo Image Recognition Device for Car Front Monitoring”, IEICE Technical Report (Technical Report of IEICE), Vol. 101, no. 302, pp37-42, Sept. 2001 Kenichi Yamada, Toshio Ito, “A Method for Capturing Vehicles by Density Projection of Edge Images”, IEEJ Transactions E, Vol. 118-E No6. pp 327-332, 1998 JP-A-7-334799 (paragraphs [0019]-[0023], [0029]-[0030], FIGS. 3, 4, 5, and 9)

  In the case of the conventional measurement described in Non-Patent Document 2, the area of the preceding vehicle in the image is extracted from the edge histogram of the monocular image, and the vehicle width (width) of the preceding vehicle on the image is detected. From the calculation of the following equation (a) showing the relationship between the vehicle width and the inter-vehicle distance and the shooting conditions (camera parameters) of the mounting angle and camera pitch of the monocular camera, the inter-vehicle distance is calculated and estimated. . In the equation, iwv is the detected vehicle width on the image, f is the focal length of the camera, Wv is the actual vehicle width, Ldist is the actual inter-vehicle distance, β is the camera pitch angle, and h0 is the camera mounting angle.

  iwv = f · [Wv / {Ldist · cosβ + h0 · sinβ}] (a)

  Therefore, as described above, if the actual vehicle width Wv of the preceding vehicle is not known, the inter-vehicle distance Ldist cannot be calculated and estimated from the equation (a).

  Under the actual traveling environment, it is difficult to make the actual vehicle width Wv of the preceding vehicle known, and it is difficult to measure the vehicle width Wv with high accuracy. The same applies to obstacles ahead of the host vehicle other than the preceding vehicle.

  Next, in the case of the conventional measurement described in Patent Document 1, the inter-vehicle distance Ldist and the vehicle width Wv of the preceding vehicle can be measured from a monocular image, but the measurement results are likely to fluctuate due to the influence of disturbance, and the accuracy There is a problem that high measurement cannot be performed.

  Furthermore, it is well known that the image processing used for this type of measurement has its recognition performance deteriorated due to image quality degradation due to ambient environmental conditions such as weather conditions, and disturbances such as noise present in the image. In addition, the effects of modeling errors due to variations in individual imaging systems and vehicle movements are also considered to affect the measurement performance. It is desirable to measure the distance from the own vehicle to the front obstacle such as the inter-vehicle distance Ldist and the lateral width of the obstacle such as the vehicle width Wv of the preceding vehicle with high accuracy.

The present invention avoids the influence of the disturbance or the like as much as possible from a monocular image, in particular when the width of an obstacle ahead of the host vehicle such as the width of the preceding vehicle is unknown. The object is to accurately measure the distance from an obstacle to the front and to measure the width of the obstacle from a monocular image at the same time based on the measurement of the distance.

In order to achieve the above-described object, the obstacle measuring method of the present invention includes a monocular camera mounted on the own vehicle, and the preceding vehicle on the photographed image is processed by image processing of the monocular image in front of the own vehicle of the monocular camera. Filtering of adaptive filter of nonlinear state space model that detects the position in the height direction of the reference part of the obstacle ahead of the vehicle and the lateral width of the obstacle as the observation position and the observation width, and inputs the observation position Based on the observation position and the photographing condition of the monocular camera, the distance from the vehicle to the obstacle is estimated, and the linear state space is filtered by the adaptive filter of the linear state space model using the observation width as an input. The lateral width is estimated by adjusting the coefficient of the observation matrix of the adaptive filter of the model by the estimated value of the distance of the adaptive filter of the nonlinear state space model, and from the monocular image, It said distance harm product is characterized by measuring the width at the same time (claim 1). In this case, the obstacle reference portion is the center of the lower end of the obstacle detected from the horizontal and vertical edge histograms of the captured image of the monocular camera, and the width of the obstacle is the peak of the two-edge histogram. It is preferable to detect from (Claim 2 ).

The obstacle detection method of the present invention is characterized in that the adaptive filter is H∞ filter (claim 3).

Next , the obstacle measuring apparatus according to the present invention performs image processing on a monocular camera mounted on the own vehicle and a monocular image in front of the own vehicle of the monocular camera, and forwards the vehicle such as a preceding vehicle on the monocular image. Image processing means for measuring the position of the obstacle in the height direction and the width of the obstacle as an observation position and an observation width, and filtering of an adaptive filter of a nonlinear state space model using the observation position as an input Estimating the distance from the vehicle to the obstacle based on the observation position and the photographing condition of the monocular camera, and filtering the linear state by using an adaptive filter of a linear state space model with the observation width as an input The lateral width is estimated by adjusting the coefficient of the observation matrix of the adaptive filter of the spatial model by the estimated value of the distance of the adaptive filter of the nonlinear state space model, and the distance of the obstacle It is characterized by further comprising a filtering unit for measuring the width at the same time (claim 4). In this case, from the horizontal and vertical edge histogram of the image captured by the monocular camera, the image processing means detects the lower end center of the obstacle as the reference portion of the obstacle, and from the peak of the both edge histogram, it is preferable to detect the width of the obstacle (claim 5).

The obstacle measuring apparatus of the present invention is characterized in that the adaptive filter is H∞ filter (claim 6).

First, according to the configuration of claims 1 and 4 , the height direction of the reference portion of the obstacle ahead of the vehicle on the image is processed by image processing of the monocular image ahead of the vehicle of the monocular camera mounted on the vehicle. width of the position and the obstacle is detected as the observation position and the observation width, observation position moves up and down in perspective of the obstacle.

Then, the through filtering of the adaptive filter of the observed nonlinear state space model position is input, the observed position and based on the photographing condition of monocular camera, estimate the distance from the vehicle as much as possible eliminate the disturbance influence to the obstacle A state estimation filter is formed, and with this state estimation filter, it is possible to accurately estimate and measure a distance such as an inter-vehicle distance from the own vehicle to the obstacle. At this time, the width of the obstacle (vehicle width) is Even if the width of the obstacle is unknown (unknown), the distance from the vehicle to the obstacle can be accurately measured by the state estimation filter.

Further, by adjusting the coefficient of the observation matrix of the adaptive filter of the linear state space model with the observation width as an input according to the estimated value of the distance of the adaptive filter of the nonlinear state space model, the observation width and the nonlinear state space model The lateral width such as the vehicle width can be accurately estimated and measured based on the estimation result of the adaptive filter.

  Accordingly, the adaptive filter of the nonlinear state space model that receives the observation position in the height direction of the reference portion of the obstacle in front of the vehicle on the monocular image in front of the vehicle of the monocular camera is input to the monocular image on the monocular image. Filtering the optimum distance and width even if the width of the obstacle is unknown (unknown) by connecting the adaptive filter of the linear state space model with the observation width of the obstacle in front of the car in series And can be measured accurately at the same time.

Next, according to the configurations of claims 2 and 5 , since the reference portion of the obstacle is the central portion of the lower end of the obstacle detected from the horizontal and vertical edge histograms of the monocular image, an error due to the size of the obstacle, etc. Therefore, it is possible to detect an observation position that varies depending on the distance, and to estimate the distance and width based on the detection of the observation position with extremely high accuracy. It can be detected accurately and reliably.

According to the configuration of claims 3 and 6 , since the adaptive filter of the nonlinear state space model and the adaptive filter of the linear state space model are configured by H∞ filters, these adaptive filters are subject to disturbances and modeling errors. The distance and width can be measured with high accuracy by filtering with a practical configuration with high accuracy.

  Next, in order to describe the present invention in more detail, the embodiment will be described in detail with reference to FIGS.

<Overall configuration>
FIG. 1 is a block diagram of an obstacle measuring device mounted on a host vehicle 1, and 2 is a single-lens camera having a small and inexpensive monochrome CCD camera configuration. The vehicle is mounted so as to photograph the road surface downward at a predetermined angle. After the engine of the host vehicle 1 is started, the front of the host vehicle is continuously photographed repeatedly and a monocular image (captured image) of each frame is output. At this time, the position of the obstacle on the monocular image moves up and down according to the distance of the obstacle ahead of the host vehicle.

  Reference numeral 3 denotes an image processing unit that captures a monocular image (captured image) of each frame of the monocular camera 2, and forms image processing means, A / D-converts the captured monocular image, for example, the latest monocular image for several frames Is stored in a rewritable memory, and a monocular image in the memory is processed by a microcomputer (CPU) according to a set image processing program, and obstacles ahead of the host vehicle such as a preceding vehicle on the monocular image are processed. The position in the height direction of the reference part is detected as the observation position.

  In this embodiment, the reference portion of the obstacle is the center of the lower end of the obstacle detected from the horizontal and vertical edge histograms of the monocular image, and the detection error due to the size of the obstacle is suppressed as much as possible.

Reference numeral 4 denotes a filter calculation unit subsequent to the image processing unit 3, which forms filter calculation means and estimates and measures only the distance from the host vehicle to the obstacle, such as the distance between the preceding vehicles, by the optimum filtering calculation of the microcomputer. In this case, an adaptive filter of a later-described nonlinear state space model Γ ₂ formed by software is provided, and the distance from the vehicle to the obstacle such as the inter-vehicle distance and the lateral width such as the vehicle width of the obstacle are estimated and measured. In this case, an adaptive filter of a later-described linear state space model Γ ₁ formed by software is also provided, and the distance and the lateral width are repeatedly estimated and measured.

The adaptive filter for state estimation of the nonlinear state space model Γ ₂ and the linear state space model Γ ₁ may be a filter based on various algorithms. In this embodiment, each filter is formed by an H∞ filter. .

<Derivation of H∞ filter>
Next, the derivation of the H∞ filter will be described.

  First, the discrete-time deterministic state space model is generally given by the following Equation (1) and Equation (2).

  Here, ξk is an n-dimensional state vector, ζk is a p-dimensional observation vector, Fk, Gk, and Hk are n × n, n × r, and p × n known state transition matrices, driving matrices, observation matrices, wk, and vk is r-dimensional noise and p-dimensional noise (vector), respectively.

  However, noises wk and vk are deterministic noises whose energy is bounded, that is, noises that satisfy the following conditional expression 3 for N> 0. This is regardless of the stochastic nature of the noise present in the filter system.

  In general, the H∞ norm for the transfer matrix T is defined by Equation (3) in Equation 5 using the following 2 norm of Equation 4 in time series u.

  Deriving a filter that minimizes Equation (3) with input as unknown disturbance (noise) ξk, wk, vk, output as estimated error ef, k is known as an optimal H∞ filtering problem. However, in general, it is difficult to obtain the optimum H∞ filter, and the derivation of the filter is treated as a finite-time suboptimal H∞ filtering problem.

That is, the optimal H∞ filtering problem is treated as obtaining a suboptimal H∞ filter that gives a groove of ‖Tk (fk) ‖∞ <γf when a certain level of control parameter γf> 0 is given. This means that the obtained suboptimal H∞ filter is a filter whose purpose is to make the maximum value of the ratio of disturbance energy and estimated error energy equal to or less than a given value γ ² f.

  The H∞ filter of the present application means this sub-optimal H∞ filter. When a level γf> 0 is given, the H∞ filter of the level γf that grooves ‖Tk (fk) ‖∞ <γf. This is expressed by the following equation (4) to equation (7).

  Here, the matrix Re, k is expressed by the following Equation 10 (8), and the error covariance matrix P ^ k | k−1 is an inequality (existence condition) shown in Equation 11 (9). Must satisfy.

  Next, the H∞ filter is given as an optimum filter in the sense of minimizing the H∞ norm of the transfer function Tk (Ff) from the disturbance (initial state, noise) to the estimation error, and this filter is the worst disturbance. In general, it is said to be robust against noise characteristic changes and modeling errors.

  There are an infinite number of H∞ filters, and various methods have been proposed for the derivation of the algorithms. For example, derivation of a discrete-time H∞ filter is attempted as a minimax estimation problem, or indefinite. Since the Kalman filter in the Klein space, which is a kind of metric space, is an H∞ filter, a unified interpretation is given to both, and it is formulated by the H∞ filter algorithm shown in the above equations (4) to (9). Is well known.

  On the other hand, when this H∞ filter is formed and filtering is performed, the determination of the parameter γf and the large amount of calculation are problematic.

[H∞ filter of linear state space model Γ ₁ ]
In this embodiment, as the adaptive filter of the linear state space model Γ ₁ , the following parameter γf can be easily determined and the amount of calculation is small, and the following Expression (10) to Expression 16 (14) are used. Use the H∞ filter of the filter algorithm shown.

  Here, the matrices λi, (i = 1,..., N) and Vk are obtained from the eigenvalues and eigenvectors of the matrix Mk as shown in the following equation (15).

  The matrix Mk is a matrix represented by the following equation (16).

The matrix Vk is a noise matrix of Vk = (v1,..., Vi,..., Vn), and the noise vi is obtained as an eigenvector of the matrix Mk corresponding to the matrix λi, where | vi | = 1. Furthermore, Vk ^T Vk = VkVk ^T = I.

[H∞ filter of nonlinear state space model Γ ₂ ]
The H∞ filter of the equations (12) to (16) is derived as an adaptive filter of the linear state space model Γ _1. Next, as an adaptive filter of the linear state space model Γ ₂ , An H∞ filter is derived for the non-formal state space model (system) Γ in which the observation matrix Hk is a nonlinear function hk (ξk).

  That is, the nonlinear function hk (ξk) in the nonlinear system Γ is changed around the estimated values (values) ξ ^ k | k, ξ ^ k | k−1 at the time of k (present) and k−1 (previous), Taylor expansion is performed as shown in the following equation (17).

  Here, the observation matrix Hk is expressed by the following equation (18).

  Therefore, the nonlinear state space model Γ is expressed by the following equation (19) to equation (22).

  Then, the parameter mk in Expression (22) is obtained for each step (for each frame) from the previous estimated amount (value) ξ ^ k | k−1, and the nonlinear state space model Γ can be handled as a linear state space model. It becomes.

  Therefore, from the H∞ filter for the linear state space model Γ1 and the above equations (19) to (22), in this embodiment, as an adaptive filter for the linear state space model Γ2, the following equations (23) to (23) The H∞ filter of the filter algorithm shown in Equation (28) of 30 is used.

  Here, the matrices λi, (i = 1,..., N) and Vk are obtained from the eigenvalues and eigenvectors of the matrix Mk as shown in the following equation (29).

  The matrix Mk is a matrix represented by the following equation (30).

The matrix Vk is a noise matrix of Vk = (v1,..., Vi,..., Vn), and the noise vi is obtained as an eigenvector of the matrix Mk corresponding to the matrix λi, where ‖vi‖ = 1. Furthermore, Vk ^T Vk = VkVk ^T = I.

<Measurement process of FIG. 1>
Next, a specific measurement process of FIG. 1 using the above H∞ filter will be described.

[Coordinate system]
First, the world coordinate system of the driving environment, the camera coordinate system of the monocular camera 2, and the imaging coordinate system thereof will be described with reference to the perspective view for explaining the coordinate system in FIG. 2 and its side view in FIG.

  The world coordinate system expresses position information of obstacles ahead of the host vehicle 1 such as a preceding vehicle with respect to the host vehicle 1, and the three orthogonal axes are centered on the lens 2a of the in-vehicle monocular camera 2 (origin). 2 and 3 are represented by the wX axis, the wY axis, and the wZ axis.

  The three axes of the camera coordinate system are represented by the cX axis, cY axis, and cZ axis in FIGS. 2 and 3 with the lens 2a as the center, and the road surface 5 in front of the host vehicle with the photographing optical axis facing obliquely downward. In the state where the cX axis coincides with the wX axis, the coordinate system is set to a state rotated by an angle β counterclockwise about the wX axis as a central axis, and the cY axis and cZ axis are the wY axis, wZ The angle is inclined from the axis by β.

  Furthermore, the image coordinate system is a two-dimensional coordinate system having the center of the captured image plane 2b of the monocular camera 2 as the origin, and as shown in FIG. 2, the y axis is parallel to the cY axis, and the x axis is the cX axis. Parallel to

  2 and 3, h0 is the height (camera height) from the road surface 5 of the monocular camera 2, f is the focal length of the monocular camera 2, and β is the camera pitch angle. In this case, as is well known, the camera coordinate system and the image coordinate system have the relationship of the following Expression (31):

[Measurement distance, width]
Next, with respect to the distance (inter-vehicle distance) and the lateral width (vehicle width) of the object to be measured in front of the host vehicle such as the preceding vehicle based on the coordinate relationship described above, FIG. This will be described with reference to explanatory diagrams of distance and width in the world coordinate system.

  First, assuming that the gradient of the road plane 5 of the traveling environment that joins obstacles such as the white car 1 and the preceding vehicle is constant, from the relationship of each coordinate system, any point (wX1, wY1 in the world coordinate system) , WZ1) and an arbitrary point (x1, y1) in the image coordinate system, there is a relationship expressed by the following Expression 34 (32) and Expression 35 (33).

  In order to simplify the description, the obstacle is the preceding vehicle 6, the coordinates of the left and right ends of the preceding vehicle 6 in the world coordinate system are (wX1, wZl), (wXr, wZr), and the preceding vehicle in the image coordinate system. 6, the coordinates of the left and right ends of the vehicle 6 are xl and xr, the vehicle width (width) measured by the preceding vehicle 6 in the world coordinate system is Wv in FIG. 5, and the distance between the vehicle 1 and the preceding vehicle 6 is measured. 5, the captured lateral width (vehicle width) of the preceding vehicle 6 in the image coordinate system is iwv in FIG. 4, and the vehicle area of the rectangular frame in FIG. 4 detected as the lower center portion on the image plane 2b. Assuming that the underline midpoint coordinates (centroid point coordinates) are (Ox, Oy), since iwv = xr−xl, the width iwv of the preceding vehicle 6 on the image plane 2b from the above equations (32) to (33), Midpoint coordinates (Ox, Oy) and car Distance is Ldist, a relationship of Expression 36 wherein the following (34), the number 37 formula (35).

  The midpoint coordinates (Ox, Oy) are the reference portion of the preceding vehicle 6, the coordinate value Oy is the observation position in the height direction, and the equation (35) It shows that the inter-vehicle distance Ldist can be measured based on the camera height h0, the focal length f, and the angle β, which are photographing conditions.

  Further, when the horizontal width iwv is detected as the observation width, the equation (34) indicates that the vehicle width (horizontal width) Wv can be measured based on the observation width, the measurement result of the inter-vehicle distance Ldist, and the photographing conditions of the monocular camera 2. ing.

[Processing of image processing unit 3 (extraction of preceding vehicle information)]
Thus, based on the monocular image of the monocular camera 2, the image processing unit 3 detects the observation position Oy and the observation width iwv as described below.

  The detection of the observation position Oy and the observation width iwv of the preceding vehicle 6 existing on the photographed image plane 2b can be performed by various image processing of the monocular image. In this embodiment, the preceding vehicle recognition by the so-called edge histogram is performed. Performed by algorithm image processing.

  In this case, the monocular image of each frame of the monocular camera 2 A / D converted by the image processing unit 3, for example, the monocular image Pa in FIG. An edge detection region V-ROI and a horizontal edge detection region H-ROI are set.

  6 is set to a position and size for capturing the preceding vehicle 6 based on various experiments and the like, and both regions H-ROI and V-ROI are This is a belt-like region that intersects in a cross shape at the center of the image.

  First, the differential image in the vertical direction is density-projected with respect to the region V-ROI set almost at the center in the frame line, and the edge histogram Gv of the vertical edge in FIG. 7 is obtained. Detects parts with strong vertical line components such as both ends.

  Next, a region H-ROI is set in the vertical direction substantially at the center of the interval between the two peaks detected from the edge histogram Gv, and a differential image in the horizontal direction is projected on the region H-ROl. An edge histogram Gh is obtained, and peaks at the upper and lower ends of the preceding vehicle having a strong horizontal line component of the histogram Gh are detected.

  Based on the peak detection information of the vertical and horizontal edge histograms, the lower end of the preceding vehicle in which the horizontal edge peak is detected is recognized, and the underline midpoint coordinates on the image plane 2b in FIG. (Ox, Oy) is detected, and the observation position in the height direction of the reference portion is detected from the coordinates Oy. At the same time, the preceding vehicle on the image plane 2b in FIG. The vehicle width (horizontal width) iwv is detected as the observation width.

  At this time, the underline midpoint coordinates (Ox, Oy) move up and down depending on the distance of the preceding vehicle 6 regardless of the size of the preceding vehicle 6, and the observation position immediately follows the distance of the preceding vehicle 6 from the own vehicle 1. Then go up and down.

[Processing of Filter Operation Unit 4]
If the vehicle width Wv of the preceding vehicle 6 is known, the inter-vehicle distance Ldist can be obtained by detecting the vehicle width iwv of the preceding vehicle 6 on the captured image plane 2b from the relationship of the equation (34). When the preceding vehicle width Wv is unknown, the inter-vehicle distance Ldist cannot be obtained only by detecting the vehicle width iwv.

Therefore, in the filter calculation unit 4 to which the observation position and the observation width obtained by the image processing unit 3 are input, the inter-vehicle distance Ldist is filtered by filtering the nonlinear state space model Γ _{2 in} the case where the vehicle width Wv is unknown. Is estimated.

In the case of this embodiment, the vehicle width Wv is also estimated simultaneously by filtering the linear state space model Γ ₁ .

Then, using the H∞ filter, adaptive filters of the following state space models Γ ₁ and Γ ₂ are formed.

(A) a linear state-space model gamma ₁ of the adaptive filter that estimates a linear state-space model gamma ₁ of the adaptive filter wheel width Wv of the formula (36) in the number of the next 38, the formula of the number 39 (37).

  In these equations, θk is the vehicle width state quantity at the time point k of the preceding vehicle 6, ζlk is an inter-vehicle distance parameter of the observation matrix Hk at the time point k, and θk, ζlk, and Hk are the following equation 40 (38) to Expression 42 (42).

(B) Adaptive filter of the nonlinear state space model Γ ₂ The adaptive filter of the nonlinear state space model Γ ₁ for estimating the inter-vehicle distance Ldist is expressed by the following equations (41) and (42).

  In these equations, ξk is the inter-vehicle distance at the time point k, ζ2k is the observation position Oy at the time point k, hk (ξk) is the observation matrix at the time point k, and the following equations (43) to (47) It is represented by Formula (45).

The state space model Γ ₁ having the vehicle width of the preceding vehicle 5 as the state quantity θk includes the inter-vehicle distance ζlk as an unknown parameter in the observation matrix Hk. By using the estimation result of the state space model Γ ₂ as this parameter, The state quantity θk and the unknown parameter ζlk can be estimated simultaneously in the framework of the adaptive filter.

Specifically, the adaptive filter of the state space model Γ ₂ is formed by the nonlinear H∞ filter, and the estimated value ξk of the inter-vehicle distance Ldist is obtained first, and the estimated value ξk is used to obtain the state space model Γ _1. The estimated value θk of the vehicle width Wv can be obtained by forming the above adaptive filter by the linear H∞ filter.

In this case, the adaptive filter filters of the state space models Γ ₁ and Γ ₂ conforming to the expressions (36) to (45) are connected in series as shown in the block diagram of the Z conversion in FIG. .

Then, based on the relationship between the state space models Γ ₁ and Γ ₂ , the filter operation unit 4 is provided with the Z-transform state estimation filter of FIG. 9 using an H∞ filter, and the H∞ filter of the nonlinear state space model Γ ₂ is provided. The estimated value ξk of the inter-vehicle distance Ldist is obtained from the block F2, and the estimated vehicle width Wv is adjusted by adjusting the observation matrix Hk of the block F1 of the H∞ filter of the linear state space model Γ ₁ according to the inter-vehicle distance Ldist. The value θk is obtained.

  In blocks F1 and F2 in FIG. 9, the estimated θk and ξk are output as estimated values zlk and z2k through the unit matrix Lk.

  Then, when the processing of the image processing unit 3 and the filter processing unit 4 is shown in a flowchart, it becomes like, for example, steps S1 to S5 in FIG. 10, and even if the vehicle width Wv is unknown based on these processes, H∞ It is possible to accurately estimate and measure the inter-vehicle distance Ldist with a configuration that is robust against noise characteristic changes and modeling errors using a filter, and at the same time, estimates the vehicle width Wv of the preceding vehicle 6. Can be measured.

[Experimental result]
Next, experimental results of simulation using the state estimation filter of FIG. 9 will be described.

  In this experiment, the vehicle width Wv of the preceding vehicle 6 on the image plane 2b is calculated from the vehicle speed of the own vehicle 1 and the preceding vehicle 6, the inter-vehicle separation, the camera parameters, and the like, and a normal noise of α = 1 is added thereto. The added signal was used as an observation signal.

  And it experimented with the following three driving | running | working patterns. That is, the first travel pattern (travel pattern 1) is a case where the inter-vehicle distance Ldist is constant, and the second travel pattern (travel pattern 2) is a case where acceleration is performed after 30 seconds ([s]). Yes, the third travel pattern (travel pattern 3) is a case where the vehicle decelerates after 30 [s].

In any driving pattern, the vehicle width Wv of the preceding vehicle is 1600 [mm], the host vehicle speed is a constant vehicle speed of 30 [km / h], and the initial value in the filtering algorithm of the H∞ filter is θ _{0 | − 1} = 1800 [mm] and ξ _{0 | −1} = 30 [m].

(A) Result of traveling pattern 1 When traveling with the inter-vehicle distance dist kept constant, the estimation results of FIGS. 11 and 12 were obtained. FIG. 11 shows the result of the inter-vehicle distance Ldist, and FIG. 12 shows the result of the vehicle width Wv.

  In both figures, the vertical axis represents distance, the horizontal axis represents time in units of 100 milliseconds, a solid line a is a true value characteristic, and a solid line b is an estimated value characteristic. The inter-vehicle distance Ldist was set to 6 [m].

(B) Travel pattern 2
When the preceding vehicle 6 accelerated after 30 [s], the estimation results of FIGS. 13 and 14 were obtained. FIG. 13 shows the result of the inter-vehicle distance Ldist, and FIG. 14 shows the result of the vehicle width Wv.

  In both figures, the vertical axis represents distance, the horizontal axis represents time in units of 100 milliseconds, and the inter-vehicle distance Ldist is set to 6 [m].

  The solid line a in both figures is the true value characteristic, the solid line b is the estimated value characteristic, and the broken line c is the characteristic when estimated by a so-called extended Kalman filter for comparison.

(C) Running pattern 3
When the preceding vehicle 6 decelerates after 30 [s], the estimation results of FIGS. 15 and 16 were obtained. FIG. 15 shows the result of the inter-vehicle distance Ldist, and FIG. 16 shows the result of the vehicle width Wv.

  In both figures, the vertical axis represents distance, the horizontal axis represents time in units of 100 milliseconds, and the inter-vehicle distance Ldist was set to 30 [m].

  The solid line a in both figures is the true value characteristic, the solid line b is the estimated value characteristic, and the broken line c is the characteristic when estimated by a so-called extended Kalman filter for comparison.

  Based on the above experimental results, it has been found that there is a difference in followability between the estimation using the state estimation filter of the embodiment and the conventional estimation using the extended Kalman filter. Specifically, for example, from FIGS. 13 and 14, it has been found that the convergence to the true value immediately after the start of the system operation is shorter than the conventional estimation, and from FIGS. 13 and 15, acceleration and deceleration are performed. When the inter-vehicle distance starts to change, it has been found that there is no response delay found in conventional estimation.

  Furthermore, it has been found that the estimation value is more sensitive to noise than the conventional estimation, and only a slight fluctuation occurs in the estimated value.

  As is well known, the Kalman filter can change the follow-up characteristics of the filter by setting system noise and observation noise. However, if these parameters are not adjusted, the characteristics of the Kalman filter are changed. It is not easy to approach the good estimation characteristics using the state estimation filter of FIG. 9, and in particular, when it is difficult to know the state of system noise, observation noise, etc. in advance, the inter-vehicle distance Ldist and the vehicle width Wv As for the estimation, a highly accurate and stable estimation like the state estimation filter of FIG. 9 using the H∞ filter cannot be performed.

  The present invention is not limited to the above-described embodiment, and various modifications other than those described above can be made without departing from the spirit of the present invention.

For example, the image processing unit 3, the observed position of the at different algorithms between the preceding vehicle recognition algorithm by the edge histogram may detect the swath width.

  The obstacle in front of the subject vehicle to be measured is not limited to the preceding vehicle 6, and may be either a traveling object or a stationary object.

Next, the adaptive filter filters of the state space models Γ ₁ and Γ ₂ of the filter calculation unit 4 may be configured by a digital filter or the like equivalent to the H∞ filter.

  And the measurement result of this invention can be used for various driving | running | working control of the own vehicle 1, such as ACC and a damage reduction automatic brake system.

  By the way, in order to reduce the number of equipment parts of the own vehicle 1, for example, the present invention can be applied to the case where the monocular camera 2 of FIG. 1 is also used as a sensor for other control such as follow-up running control.

1 is a block diagram of an embodiment of the present invention. FIG. 2 is an explanatory diagram of the relationship of the coordinate system of FIG. 1. FIG. 3 is a side view of FIG. 2. It is explanatory drawing of the observation position of FIG. 1, and observation width. It is explanatory drawing of the coordinate transformation of FIG. FIG. 2 is a shooting diagram of the monocular camera of FIG. 1. It is explanatory drawing of the process image of FIG. FIG. _{2 is} a block diagram of state space models Γ ₁ and Γ ₂ of the filter operation unit in FIG. 1. FIG. 9 is a block diagram of a state estimation filter of the filter operation unit of FIG. 1 in accordance with the state space models Γ ₁ and Γ ₂ of FIG. 8. It is a flowchart for process description of FIG. It is a characteristic view of the example of estimation of the inter-vehicle distance of the running pattern 1. It is a characteristic view of the example of estimation of the vehicle width of the running pattern 1. It is a characteristic view of the example of estimation of the inter-vehicle distance of the driving | running | working pattern 2. FIG. FIG. 6 is a characteristic diagram of an example of estimating a vehicle width of a running pattern 2; It is a characteristic view of the example of estimation of the inter-vehicle distance of the running pattern 3. It is a characteristic view of the example of estimation of the vehicle width of the driving pattern 3.

Explanation of symbols

1 Car 2 Monocular Camera 3 Image Processing Unit 4 Filter Calculation Unit

Claims (6)

A monocular camera is installed in the vehicle,
By the image processing of the captured image in front of the vehicle of the monocular camera, the position in the height direction of the reference portion of the obstacle ahead of the vehicle such as the preceding vehicle on the captured image and the lateral width of the obstacle are observed positions. And detected as an observation width ,
By filtering the adaptive filter of the nonlinear state space model with the observation position as input, the distance from the own vehicle to the obstacle is estimated based on the observation position and the photographing condition of the monocular camera ,
By filtering the adaptive filter of the linear state space model with the observation width as an input, the coefficient of the observation matrix of the adaptive filter of the linear state space model is adjusted by the estimated value of the distance of the adaptive filter of the nonlinear state space model. Estimating the width,
Wherein the distance from the captured image the obstacle, obstacle measuring method characterized by measuring the width at the same time. Reference portion of the obstacle, horizontal shot image of a monocular camera, lower central portion der of the obstacle detected from the vertical edge histogram is, and the width of the obstacle, detecting the peak of the both edges histogram obstacle measuring method according to claim 1, wherein to Rukoto. The obstacle measuring method according to claim 1 , wherein the adaptive filter is an H∞ filter . A monocular camera mounted on the vehicle,
The captured image of the front of the vehicle of the monocular camera is subjected to image processing, and the position in the height direction of the reference portion of the obstacle ahead of the vehicle such as the preceding vehicle on the captured image and the width of the obstacle are observed positions. And image processing means for measuring as an observation width,
By filtering the adaptive filter of the nonlinear state space model with the observation position as an input, the distance from the own vehicle to the obstacle is estimated based on the observation position and the photographing conditions of the monocular camera, and the observation width is By filtering the adaptive filter of the linear state space model as an input, the coefficient of the observation matrix of the adaptive filter of the linear state space model is adjusted by the estimated value of the distance of the adaptive filter of the nonlinear state space model to estimate the lateral width And an obstacle measuring device comprising a filter calculating means for simultaneously measuring the distance and the width of the obstacle. The image processing means detects the lower end center of the obstacle as a reference part of the obstacle from the horizontal and vertical edge histograms of the captured image of the monocular camera, and the obstacle from the peak of the both edge histograms. The obstacle measuring device according to claim 4, wherein the width of the obstacle is detected. The obstacle measuring apparatus according to claim 4 or 5 , wherein the adaptive filter is an H∞ filter .