TW201534512A - Control method about obstacle avoidance and navigation by binocular images - Google Patents

Abstract

This invention focus on the control method about obstacle avoidance and navigation by binocular images. A mobile platform is navigated based on binocular images without manual operation. It utilizes binocular stereo images by matching method to generate a vision disparity map. The obstacle avoidance algorithm uses the vision disparity map to find the barrier free path. The automatic navigation and obstacle avoidance method are according to the mobile platform motion to design obstacle avoidance path. The navigation control algorithm includes binocular image corresponding point matching, detecting barrier free path and path planning.

Description

Dual image obstacle avoidance path planning navigation control method

The dual image obstacle avoidance path planning navigation control method is a method for autonomous navigation of a mobile platform, and particularly relates to a method for autonomous navigation, which uses a dual camera to acquire images to determine a movable path and perform Plan to avoid obstacles.

At present, in the field of car navigation robots, there have been more and more researches on navigation obstacle avoidance. The common way is to use ultrasonic sensors or laser scanners to directly feed back the situation. It's a very straightforward way, but there are still some challenges that need to be broken, such as sharp-shaped obstacles or pits on the ground. The current situation is that monocular cameras can obtain image sensing data to achieve obstacle avoidance. Only a 2D image can be acquired by a single-eye camera to calculate 3D information. Many known information aids must be used for image calibration in order to obtain a proportional relationship between the size of the image and the world coordinates, and the obstacle must be obtained before the obstacle avoidance. Only by establishing a specific calculation method can the monocular camera avoid obstacles.

Therefore, the present invention uses a dual camera to obtain sensing data, and the obtained sensing data is subjected to stereoscopic space calculation, thereby enabling more accurate three-dimensional space regardless of obstacle shape, feature, and color. The calculation method enables the mobile platform to autonomously avoid obstacles.

The dual image obstacle avoidance path planning navigation control method refers to a method for enabling autonomous navigation of a mobile platform, and the two cameras are used as sensors for sensing an external environment and obtaining sensing data. After the image sensing data is processed by the moving average algorithm, the image is firstly graded, and the image is assumed to be continuous. Therefore, the previously generated disparity map and the two gradient images are calculated to generate a new disparity map; After the parallax map, according to the ground Like the linear feature of parallax, the feasible space is detected, and the geometric relationship between the ground and the camera is used to find out the features of the ground in the disparity map to detect the ground that can be walked.

The dual image obstacle avoidance path planning navigation control method adopts a binocular stereo vision matching method to generate a parallax map, and then designs an obstacle avoidance algorithm to process the parallax map to obtain an unobstructed space, and the method replaces the brightness difference by a brightness gradient difference. And add the smoothness of the disparity map, and process the left and right images to obtain a disparity map.

1‧‧‧Image obtained by the camera

11‧‧‧Processing of generating disparity maps

12‧‧‧ Path Planning Algorithm

13‧‧‧Determining the direction and speed of the car body

21‧‧‧ camera

22‧‧‧Optical Center

23‧‧‧ Optical center distance of two lenses

31‧‧‧Polar geometry limits

32‧‧‧ Any point in the space

33‧‧‧focal length

34‧‧‧ P _R

35‧‧‧ P _T

41‧‧‧polar line

42‧‧‧ Target image

43‧‧‧Reference image

51‧‧‧ original image

52‧‧‧ horizontal gradient image

53‧‧‧ vertical gradient image

54‧‧‧Edge detection image

61‧‧‧ reference window

62‧‧‧Search area

63‧‧‧Disparity

72‧‧‧Disparity map DM

81‧‧‧Programs from the bottom of the image

82‧‧‧Search for programs with accessible paths

83‧‧‧Reprogramming the width of the car and the starting point of the next search starting point

91‧‧‧ Pavement feasible space detection

92‧‧‧Search route points

93‧‧‧Delete or stack path points and perform loops to the last column

94‧‧‧ Estimated car width in the image

111‧‧‧Self-driving car width

Z‧‧‧In-depth information

Figure 1 is a flow chart of the method of the present invention.

2 is a schematic view showing the erection of a standard stereoscopic camera of the present invention.

Figure 3 is an image model of the standard stereoscopic vision of the present invention.

Figure 4 is a view of the standard stereoscopic image data of the present invention.

Figure 5 is a Sobel gradient and edge image processing of the present invention.

Figure 6 is a schematic diagram of the estimated window disparity of the matching window of the present invention.

Fig. 7 is a flow chart showing the generation of a parallax map of the present invention.

Figure 8 is a schematic illustration of selected matching points of the present invention.

Figure 9 is a flow chart of the path planning of the present invention.

Figure 10 is a schematic illustration of the geometric relationship of the ground and camera of the present invention.

Figure 11 is a range view of the walkable space of the present invention.

Figure 12 is a perspective projection view of parallel lines of the present invention.

Figure 13 is a graph showing the results of the binarization of the walkable space of the present invention.

Fig. 14 is a view showing a vehicle width map in the parallax map estimation image of the present invention.

Figure 15 is a diagram showing the results of the path planning of the present invention.

Figure 16 is an explanatory diagram of the search path point algorithm of the present invention.

Figure 17 is a diagram showing an example of a calculated turning angle of the present invention.

Figure 18 is a schematic view showing the steering movement of the self-propelled vehicle of the present invention.

Figure 19 is a schematic view of the intersection of the minimum radius of gyration of the present invention.

Figure 20 is a block diagram of the hardware system of the present invention.

Figure 21 is a diagram showing the results of the parallax map of the outdoor scene and the experimental site of the present invention.

Figure 22 is a result of the framework of the feasible space and path planning of the present invention.

Figure 23 is a graph showing the results of the experiment one of the present invention.

Figure 24 is a diagram showing the results of the experiment 2 of the present invention.

Figure 25 is a diagram showing the results of the experiment three results of the present invention.

Figure 26 is a diagram showing the results of the experiment 4 of the present invention.

In order to fully understand the present invention, the preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. The statements in this specification are hereby implemented and are not intended to limit the technical scope of the invention.

The dual image obstacle avoidance path planning navigation control method of the present invention, as shown in FIG. 1, illustrates the method of the patent by a control flow chart, which includes: an image obtained by a camera (1), a process of generating a disparity map (11), The path planning algorithm (12) and the direction and speed of the car body (13), the operation of the system is to capture the scene image by two parallel cameras (21), and the two images are transmitted through the patented parallax map algorithm. To obtain 3D environmental image information, the environmental image information is processed by the obstacle avoidance algorithm of this patent, and the walkable space and path planning are detected, and an algorithm is designed to determine the moving direction and speed of the traveling vehicle by using the vehicle width. The first part of the text describes the algorithm of the disparity map, the second part describes the algorithm of obstacle avoidance path planning, the third part is the control algorithm that determines the speed and steering, and finally the experimental results are used to illustrate the function of this patent.

1. Parallax map: The following shows that the image information of the overlapping field of view is obtained by two cameras (21) to calculate the depth of field of the image, and the target point and the two images are illustrated by a target point in the overlapping field of view of the two cameras (21). Point relationship, as shown in Figure 2, the camera (21) has an optical center (22) (Optical Center), assuming that the two cameras (21) are erected, except for the baseline (Base Line, the amount of translation along the X axis), the other image coordinate systems are identical, and their optical axes (Optical Axes) are mutually Parallel, the optical center distance (23) of the two lenses is B.

The image model in the above case is shown in Figure 3. X , Y , and Z are world coordinates, assuming P _{( X , Y . Z} ) is any point in the space between the overlapping views of the two cameras (21) in the scene (32 ) and no shading phenomenon, and respectively along the two cameras (21) of each of the optical center (22) O, a projection for the reference image at the focal length f _{(33) (43) P R} (34) and the target image ( 42) Upper P _T (35), where P _R (34) is the imaging point of the spatial point P in the right image, and P _T (35) is the imaging point of the spatial point P in the left image. When the limitation of the standard stereoscopic geometry is met, the two image planes are on the same plane, and the polar lines (41) corresponding to the two images appear on the same line, so that P _R (34) and P _T (35) are in the left and right images. The positions on the plane Y axis are the same.

Continue, let the image coordinates of P _R (34) be ( x _R , y _R ), C _R be the image center point of the reference image (43), and the image coordinates of P _T (35) be ( x _T , y _T ), The depth information (Z) of a point in the scene can be obtained by the geometric relationship of similar triangles:

In Equation 1, B and f are constants, so the depth information (Z) of a point in the scene is required, the most important one is ( x _T - x _R ), which is generally called ( x _T - x _R ) is disparity. , expressed in d . How to find the corresponding point P _T (35) in the target image (42) from the P _R (34) of the reference image (43) becomes the core technology for designing stereo vision.

Since the polar line (41) is on the same horizontal line under the binocular vision architecture, it can be searched from the image data along the same column, as shown in Fig. 4, and because x _T ≦ x _R , it only needs to go in a single direction. Search can greatly reduce the search time, but how to find the best match point from the candidate is the focus of the matching algorithm in the disparity map. In order to find the corresponding matching points, it is necessary to do some image processing in advance to obtain more obvious matching features. The following will explain the image processing by image filtering to find the target image (42) and the reference image. Corresponding point of (43).

Regarding spatial filtering: the brightness of a general image pixel is represented by a grayscale value, and the gradient of a image refers to a change in brightness of a pixel, which is a partial differential of luminance to a pixel in an image space, such as an image I( u,v ) The gradient vector at ( u,v ) can be expressed as:

To realize the equation (3), it can be implemented by the associated operation. The operation is as follows. After the original image I( u,v ) is subjected to the convolution operation to obtain the new image G ( u,v ), it can be expressed by the following formula:

In the formula, the mask size of h ( u,v ) is an odd square matrix of W × W , a =( W -1)/2. Since equation (4) is directly processed on the image, h ( u , v ) can be called a spatial filter. Using equation (4), different masks h ( u , v ) can be designed to detect the edges of the image.

The use of spatial filters for edge detection can be handled by Sobel masks. The 3×3 Sobel mask is shown in equation (5), and the g _u of equation (5) is the horizontal mask, only for the image. The change in the horizontal direction is responsive, and the g _v of the formula (5) is a mask in the vertical direction. The operation of using the Sobel mask for equation (5) is:

Taking the original image (51) in Fig. 5 as the original image, and using the g _u and g _v masks to perform a gyroscopic product operation with the original image (51), the horizontal gradient image G _u can be obtained respectively, as shown in Fig. 5 . The horizontal gradient image (52) and the vertical gradient image G _v , as in the vertical gradient image (53) in FIG. The gradient image has directionality, the gradient from dark to bright is positive, the gradient from bright to dark is negative, and the two images of horizontal gradient image (52) and vertical gradient image (53) are squared. The root number gives the size of the original image (51) gradient. The gradient at the edge of the image will be relatively large. This feature is used to set a threshold T for discriminating the edge. As shown in equation (6), the pixel whose gradient is greater than the threshold T is the edge, and the lower the threshold T is detected. The more edges there are, the more likely the noise will affect the result. Conversely, a too high threshold T will lose the details. The image E binarized by this method is edge detection, such as the edge detection image in Fig. 5 (54).

About Local Matching: First determine the size of the matching window, extract the matching image from the reference image (43) (Reference Image), and then compare it in the target image (42) (Target Image) to find out The closest matching position finds the corresponding point of the reference image (43) and the target image (42). As shown in FIG. 6, centering on ( u , v ) of the left reference image (43), a matching matching window W is taken as a reference window (61), and a search area range (62) is formed, on the right. The target image (42) searches along the same polar line (41), and finds the matching window with the smallest degree of difference in the search area range (62). The difference degree calculation method has the difference absolute sum (SAD) or difference. The sum of squared values (SSD), as shown by equations (7) and (8), respectively, can be obtained by equation (7) or equation (8) along the polar line (41), and ( d _max - d _min ) can be obtained. The cost score map, and then search for the most suitable disparity value (63) d .

The matching method proposed in this patent replaces the luminance difference by the luminance gradient difference, adds the smoothness of the disparity map, and processes the left and right images to obtain a disparity map, as shown in FIG. 7. The brightness gradient difference can solve the matching error caused by the difference of the amount of light entering between the cameras (21). Since the camera (21) is continuously taking images during shooting, in the image sequence, the two adjacent images are related to each other. , the disparity map provided disparity map DM (72) is the last calculated output, taken from the disparity map matching window W DM (72), the absolute value of the sum of d and subtracting, when the parallax closer to the mean d and W When the value is used, the smaller the calculation result is, similar to the effect of averaging the disparity map. Since the noise of the camera (21) is usually isolated in the image and is discontinuous in the time series, this algorithm also has the effect of suppressing the noise of the camera (21).

The Cost calculation method of Figure 7 is as follows:

C ( x , y , d )= C _GRAD ( x , y , d )+ω* C _Smooth ( x , y , d ) (9)

Where ω is the weight ratio between the two, the larger the C _Smooth specific gravity, the smoother the disparity map is obtained, but when the specific gravity is too large, the calculation of C _Smooth is to use the disparity map of the last output, and in the non-texture area C _GRAD has a weak influence, so it is easy to leave a residual image of parallax in the non-textured area. After gradually performed from d to d _max d _min, each pixel will be (d _max d _min) of the image rates result, as shown, an image score 8 a multi-dimensional data, i.e. d-axis, then each The pixel searches for the lowest d of the score from the score result as the disparity value (63). This is because the smaller the Cost is, the higher the similarity is, so the lowest d is searched for each pixel, and d is used as the parallax. In the figure, the pixel has a disparity value (63).

If done from d _max d _min and then find the minimum, it requires a large amount of memory, therefore, the image of this patent do the whole operation, to reduce the amount of memory used. Before describing how to manipulate the disparity map for the whole operation, first define some code numbers as follows:

●Cost score map (C): used to calculate the temporary storage space of Cost. The pixel values of the initial image are infinite, and the size of the Cost score map (C) is consistent with the image.

● Disparity map (D): used to calculate the temporary storage space of the disparity map. The pixel values of the initial image are all zero, and the size of the disparity map (D) is consistent with the image.

● Output disparity map ( DM) : For the output of the disparity map, the initial image has zero pixel values, and the disparity map ( DM) is the same size as the image.

Parallax ( d ): The position difference (pixel) of the corresponding point in the horizontal direction, and the search range is d _min ~ d _max .

● W mask: consists of all 1s, and the convolution product for the image can replace the sum of the grayscale values of the adjacent pixels.

The method of obtaining the disparity map using the whole operation is as follows:

Step 1. Use the output disparity map DM (72) of the previous disparity map to copy the entire image of the disparity map DM (72) to D, and C to initialize the whole frame, the initial value is infinity.

Step 2. Obtain the brightness gradient image of the left and right images.

Step 3. The gradient image of the target image (42) is moved horizontally by d lines.

Step 4. Obtain the absolute values of the phase difference between the two sides of the vertical gradient and the horizontal gradient, respectively, and the absolute value of the difference between the disparity map DM (72) and the constant d of the previous disparity map.

Step 5. Use the W mask to make a gyro product. The mask size is the same as the matching window size, and the energy map of equation (11) is obtained.

Step 6. The obtained energy map is compared with C. The block smaller than C replaces the value of C with the current energy value, and the same block replaces the value of D with the current d , and is constantly modified.

Step 7. Return to step 3, d from d _min until the end of d _max and do the next step.

Step 8. D becomes a new disparity map, and D is copied to the disparity map DM (72) as the output result of the disparity map.

2. Image navigation obstacle avoidance control process: The patented binocular visual obstacle avoidance navigation control has the function of allowing the driverless car to have no environmental model before walking, and completely relies on the patented method to detect the parallax map in the sports environment. A walkable space, whereby the driverless car can draw a path of movement on its own, moving in an environment full of obstacles, avoiding collisions with obstacles and completing navigation control.

The image navigation control process of the driverless car studied in this patent is shown in Figure 9. The two cameras (21) are the only sensors in the system that sense the external environment. The two cameras (21) have been subjected to the extreme line geometry limitation. (31) Correction, after obtaining the image, first gradient the image, assuming that the image is continuous, so the previous generated disparity map and the two gradient images are calculated to generate a new disparity map. After obtaining the disparity map, according to the linear feature of the parallax of the ground image, the feasible space is detected. From the first column of the image, according to the initial vehicle width and the initial starting point, the path point in the disparity map is searched to determine whether the path point is reasonable and updated. The vehicle width and the starting point, wherein the updated vehicle width is the portion of the vehicle width (94) in the estimated image in Fig. 9, and the update starting point is the portion of the pavement feasible space detection (91) in Fig. 9. Copy, cycle to the last column, plan the obstacle avoidance path, that is, the part of the search path point (92) in Figure 9, determine the direction of the vehicle body from the path, delete or stack the path point as shown in Figure 9 and perform the loop to In the last column (93), the speed of the vehicle body is determined by detecting the obstacle in the minimum radius of gyration, that is, determining the direction and speed of the vehicle body (13). The algorithm for the detailed walkable path is described later.

The patented driverless vehicle navigation method utilizes the geometric relationship between the ground and the camera (21) to find out the features of the ground in the parallax map to detect the ground that can be walked. The geometric relationship between the ground image parallax and the camera (21) is shown in Fig. 10. In Fig. 10, the optical center point (22) of the camera (21) is O , the height h from the ground, and the Y axis and the Z axis are the camera (21). The coordinate system, the focal length (33) of the image plane is f , so that P is any point on the ground space (32) in the image, the Y- axis direction from the center point C of the image is y, and the angle between h and the Z- axis is θ angle. α =tan ^-1 ( y / f ), known from equation (1) and Figure 10 , so the d in the disparity map is:

In the middle ,versus Is a constant coefficient. From the formula (12), the ground is d = a in the disparity map. y + b, which is a straight line, so this feature is used to find the ground that can be walked without obstacles, and the result is shown in FIG.

Regarding the part of the estimated vehicle width, after judging the walkable space by the parallax map, it is also necessary to know the vehicle width of the driverless vehicle in the parallax map to satisfy the demand for the travelable path planning. Imagine a pair of parallel lines on both sides of the car, extending infinitely along the front of the car body. As shown in Figure 12, this parallel line will intersect at the vanishing point in the image, indicated by V point, the path starting point S The position is in line with the V point, and it is known by the formula (12) and FIG. And learned from equation (1) , can be pushed to the following formula:

Let the distance between the actual right end point X ₁ and the left end point X ₂ of the vehicle body be the vehicle width W=( X ₁ - X ₂ ), and substitute X ₁ and X ₂ into the equation (13) to obtain x _R2 and x _R1 in the image respectively. , the width w of the bicycle in the image is:

It is known from the equation (14) that the traveling vehicle width w is proportional to the parallax d of the walkable space.

In the disparity map, the number of pixels occupied by the vehicle width in the column direction is estimated, and the search path point is provided as a basis for allowing the driverless vehicle to pass. After searching for the position of the path point, the w parallax d of the position in the disparity map is taken. The median of the values, recalculating the car width w in the disparity map.

This patent search path is from far to near, far disparity is located close to zero, when w is too short, too little information, prone to false positives, it defines the minimum number of Pixel w. The search method of the walkable path will be described below by placing the camera (21) in the middle of the driverless vehicle.

Regarding the path search, as shown in FIG. 13, let E be the target point, and the imaging position of the destination in the driving direction, and set S as the starting point of the path, and the imaging position in the center of the vehicle width, because the driverless car is toward the most Go straight in the direction, so set the destination to go with the path start point. The search path starts from the first column and travels with the E point. Search for the left and right directions. Read the data according to the length w of the vehicle to find the position without obstacles. In other words, the space can be doubled. The information of the image is zero. When the vehicle width length w <1 or too short, errors or too little information are unreliable, so the minimum vehicle width must be set. The maximum range for searching to the right or searching for the left is the image width plus w/ 2 to the left minus w/ 2, and the possible outcomes of the search are (1) no right or left to walk, (2) right Or search for a walkable space to one of the left, and (3) search for a walkable space to the right or left. In case (1), search directly to the next column. In case (2) or (3), there is a walkable space. In order to make the path points connect to a continuous curve, and finally connect with the S point, according to the search position and the bottom of the image. the difference between the current time ħ unmanned vehicle search at a ratio relationship between the position (horizontal direction) and a difference video border l set point, 13, as ħ (Bottom-i), l is the (waypoints found - S ). After finding the position that can be walked, the marker point of the path is made in the center of the position. In order to speed up the search speed, the total value of the self-read data is used to make the movement of the reading track, and if the search is to the left and right, the reading is performed. If the trajectory is completely beyond the image range, the path point cannot be found on this column, and the next column is searched. When the path point is found, the vehicle width w of the driver's car in the next row of the disparity map is calculated according to the middle number of the disparity value (63) d at the same position in the disparity map, as shown in FIG. The search starting point of the next column is linearly approached to the S point. In order to solve the problem that the width of the car body in the image tends to exceed the image size when searching near the bottom of the image, the condition for designing the search for the column is read. The trajectory is completely outside the image range, so it is reasonable to have the path point (the center of the read trajectory) appear outside the image. When the path point is not found, the starting point position of the next column search is again with the E point, and the previous path point is deleted. The delete mark in FIG. 13 indicates that all the path points before the column are invalid. Through the above calculations, the walkable space can be calculated. The result is shown in Fig. 15, where a and c are the ranges of the walkable space, and b is the path for calculating the planned vehicle body walking. The algorithm for path search summarized by the above description is shown in Fig. 16. The ground in the image extends from bottom to top to infinity, so the search path searches for the path that can be walked from the farthest point, so the search path When clicking, the last column is searched from the first column, as shown in the third row of Figure 16, which is the program sequence (81) detected from the bottom of the image. In the ninth to sixteenth steps of FIG. 16, the program for searching for the barrier-free path (82), if the 14th line is not established, the path point is found, and it is necessary to judge whether the image is completely exceeded, and the 17th to 28th lines in Fig. 16 Yes, if it is not exceeded, reset the program width (83) of the car width and the next search start position and record it. The sample result of this algorithm is shown in Fig. 15. The blue lines indicating a, b and c are the paths connecting the travelable path points, and the central line segment b is the walking area according to the width of the vehicle. The form car body depends on the space occupied by the path in the image.

3. Direction and speed of the vehicle: After planning the path, take the path point near the S end to calculate the angle θ with the vertical line of S , and use θ as the basis for the steering control of the driverless vehicle, as shown in Figure 17.

The general vehicle is controlled by the front wheel, and mainly depends on the speed difference steering of the left and right wheels. The steering motion is as shown in Fig. 18, so that v _L and v _R are the speeds of the left and right wheels respectively (m/sec), v _C is the speed at which the car advances (m/sec), ω _C is the angular velocity of the turn (rad/sec) of the vehicle, and w is the distance between the left and right wheels (m). The equation of motion is as follows:

If the minimum radius of gyration of the car is k _min , the equation is:

At the defined minimum radius of gyration, the maximum upper limit ω _{C _max of the} maximum angular velocity ω _C can be calculated from equation (16).

In terms of speed, since the minimum radius of gyration is limited, when an obstacle occurs in the intersection of the left and right minimum gyration radii, the driverless car cannot be avoided, and the driverless car should perform the braking action. The trajectory of the two minimum radius of gyrations is in the intersection of the images, similar to the shape of a triangle. As shown in Fig. 19, the area enclosed by triangles is easier to search. When an obstacle appears in this area of the feasible space binarization map, The command speed is reduced to zero. In order to prevent the car from getting off the bus, the area of the intersection area is enlarged, and the action of deceleration is performed according to the situation of obstacles in this area.

4. Driverless navigation control experiment: The experimental system of the dual image obstacle avoidance path planning navigation control method of this patent, as shown in Fig. 20, the experimental vehicle is an electric vehicle driven by a motor with left and right front wheels, and the vehicle body is loaded with two cameras. (21) An environmental image is obtained by the image capturing device. First, the image of the environment is obtained by the parallel left and right cameras (21), the parallax map is generated by the computer for image processing, the feasible space is identified, and the direction and speed are determined by the path planning algorithm (12), and the command is sent to the motor. The control circuit and the motor feedback signal processing interface circuit instantaneously calculate the rotational speed required for the left and right wheel motors, and output a control voltage signal to the motor driver of the left and right wheels to operate the motor to achieve the forward and steering control of the driverless vehicle. During the navigation process, the camera (21) updates the direction of the driverless car as the driverless car moves in the environment, and can correct the driving path at any time even if a new obstacle appears. The path planning test and navigation test are described later.

Path planning test: Experiments performed in different scenarios to illustrate the functions of this patent. As shown in FIG. 21, there are two outdoor scenes, one indoor scene, and the amount of light entering the left and right cameras (21) in FIG. 21 is not the same, but for the method of the present patent, the parallax map results have little effect and match. As a result, few noises appeared. Figure 22 takes the disparity map to detect the feasible space and then performs the path planning. The frame selection range indicated by the left figure is the feasible space detected by the method of the patent, and the line segment shown by the right icon is the path planned by the method of this patent. As a result, the range indicated by the line segment is the area occupied by the vehicle width in the image. The experimental results show that the different methods in the indoor and outdoor can be used to detect the feasible space and complete the path planning.

Navigation test: The purpose of the experiment is to test whether the test vehicle can pass multiple obstacles in different environments. The following description shows that the experimental vehicle uses the algorithm of this patent to complete the obstacle avoidance path planning. Experimental results.

The first experimental site is an indoor obstacle environment. As shown in Figure 23, the experimental vehicle avoids the obstacle and passes through a 0.75 m wide door. The width of the experimental vehicle is 0.6 m. The vehicle width is set by adjusting the parameters in the program. And according to the position of the camera (21) on the experimental vehicle, adjust the ratio of the camera (21) in the width of the vehicle. After these corrections, the experimental vehicle can smoothly pass through the narrow door. During the experiment, the left and right wheels of the experimental vehicle are recorded. The value of the motor encoder is used to calibrate the walking trajectory of the experimental vehicle, as shown by the dashed line in Fig. 23. In addition, a piece of cardboard is placed on the ground and aligned with the ground. Since the method of this patent mainly uses stereoscopic vision to identify a feasible space, It should be free from the interference of the ground cardboard, and can go directly from above. It is confirmed by the experimental results that the method of this patent has this function; the experimental site is the corridor inside the building, as shown in Figure 24, the corridor of the corridor is 3m long and 13.5m long. And placing a plurality of obstacles on the corridor, the experimental vehicle detects the feasible space according to the parallax map matched by the binocular camera (21), and plans the motion route through the path planning algorithm (12) of this patent. Through the corridor, the experimental vehicle constantly corrected the direction of the driving due to the obstruction of the wall and obstacles. According to the experiment, the experimental vehicle found a feasible path in the obstacle area of 13.5m and successfully passed the obstacle area; the experimental site 3 was outdoor 5m long and 3m. Wide asphalt road, there is no wall limitation under the wide field, and multiple obstacles are placed on the road. The experimental vehicle can drive on the asphalt road and can smoothly avoid obstacles. The experimental results are shown in Figure 25; The venue is the same as the site 3, and the obstacles are placed in the same position as the site 3. However, the initial position of the experimental vehicle is changed, and pedestrians are arranged as moving obstacles. The experimental results are shown in Figure 26, because there is no surrounding area under the wide field. With the limitation of the wall, the experimental vehicle can easily leave the obstacle area. When the experimental vehicle meets the pedestrian moving on the roadside, the experimental vehicle can turn correctly and avoid the pedestrian walking in the area without obstacles. The environmental test proved by various experimental scenes, this patent The proposed dual image obstacle avoidance path planning navigation control method can control the vehicle to successfully avoid different types of obstacles, and By hazard, so the method of the present patent can be applied to an unmanned navigation system.

5. Conclusion: The dual image obstacle avoidance path planning navigation control method of this patent takes a moving environment image through two cameras (21), obtains a disparity map through image processing, detects a feasible space from the disparity map, and plans a navigation path. , according to the width of the car to determine the speed and steering control, controllable The vehicle travels to achieve the goal of driverlessness.

In conclusion, it is only a preferred embodiment of the invention, and is not intended to limit the scope of the invention. That is, the equivalent changes and modifications made by the scope of the patent application of the present invention are covered by the scope of the invention.

1‧‧‧Image obtained by the camera

11‧‧‧Processing of generating disparity maps

12‧‧‧ Path Planning Algorithm

13‧‧‧Determining the direction and speed of the car body

Claims (2)

A dual image obstacle avoidance path planning navigation control method mainly adopts a binocular visual matching method, and then designs an obstacle avoidance algorithm to process the parallax map to obtain an unobstructed space, and the method replaces the luminance difference by a luminance gradient difference, and adds a disparity map. Smoothness, processing left and right images to obtain a disparity map. A dual image obstacle avoidance path planning navigation control method is obtained by taking a binocular camera image, and after obtaining a parallax map by image processing, the image is processed by a moving average algorithm, and the image is firstly graded, and the image is assumed to be continuous, so The parallax map generated at one time and the two gradient images are calculated to generate a new parallax map. After obtaining the parallax map, according to the linear characteristics of the parallax of the ground image, the feasible space is detected, and the geometric relationship between the ground and the camera is used to find the ground. The features in the disparity map are used to detect the ground that can be walked.