CN110956179A

CN110956179A - Robot path skeleton extraction method based on image refinement

Info

Publication number: CN110956179A
Application number: CN201911201016.8A
Authority: CN
Inventors: 严锡君; 候添琪; 刘旭东; 刁宏志; 蒋悦
Original assignee: Hohai University HHU
Current assignee: Hohai University HHU
Priority date: 2019-11-29
Filing date: 2019-11-29
Publication date: 2020-04-03

Abstract

The invention discloses a robot path skeleton extraction method based on image refinement, which comprises the following steps: selecting a proper step length to perform pooling treatment on the robot road original image by utilizing a reforming group idea; performing Zhang-suen thinning on the obtained pooling result; carrying out mask processing on the thinned pooled image; and reducing the pooling template, and further performing pooling and Zhang-suen thinning on the image to obtain the robot path skeleton. The invention can reduce the calculation complexity, accelerate the thinning speed and greatly reduce the time for completing the whole task on the premise of not obviously losing the image information.

Description

Robot path skeleton extraction method based on image refinement

Technical Field

The invention belongs to the field of image processing or robot control, and particularly relates to a robot path skeleton extraction method.

Background

At present, robots, as an important aspect of intelligent technology, are hot spots in research and development in this field. The teaching and practice activities of the education robot related to the teaching and practice activities are developed vigorously in all colleges and universities, and play a role in guidance and leading.

Among educational robots, there is a type of practical robot that simulates an actual activity by recognizing a simulation path and controlling a motion trajectory, such as simulating a medical service. Generally, such wheeled robots acquire information about a ground path and surrounding environment by using sensors such as space, distance, and photoelectric sensors, and control the rotation speed of each motor to change the motion and direction of the robot, start and stop the robot, and complete the entire motion process as quickly as possible. Since the photoelectric sensor can only detect the path of the actual arrival point and cannot predict the path in advance, when the path changes direction, the robot generally adopts a method of stopping and then changing the direction, so that the movement speed is greatly reduced.

In order to increase the running speed of the robot and overcome the defects caused by the detection of the photoelectric sensor, many research and development personnel apply the image processing technology to the robot. The images collected by the camera are converted into quantitative parameters which can provide direction information for the mobile robot, and then the path is further processed and extracted. The path extraction generally includes four parts, namely preprocessing, binarization, post-processing, skeleton extraction and the like, the skeleton extraction is the most critical step in the path extraction, and the skeleton can convey structural information of an original image, including the position, direction, length and the like of line segments, which can represent the content of the composed image.

The embedded system has limited computing resources, and the execution rate of the image processing algorithm is very important. For the same image, it is worth exploring how to reduce the processing time under the condition that the error of the extracted parameter is not large. When the traditional Zhang-suen thinning algorithm is used for processing images, the images are traversed each time, and after pixel points meeting requirements are deleted, the outermost circle of pixel points of object contents are stripped. The processing method has a good processing effect on the complex image, but the execution efficiency is low.

Disclosure of Invention

In order to solve the technical problems mentioned in the background art, the invention provides a robot path skeleton extraction method based on image refinement.

In order to achieve the technical purpose, the technical scheme of the invention is as follows:

a robot path skeleton extraction method based on image refinement comprises the following steps:

(1) selecting a proper step length to perform pooling treatment on the robot road original image by utilizing a reforming group idea;

(2) zhang-suen thinning is carried out on the pooling result obtained in the step (1);

(3) performing mask processing on the thinned pooled image in the step (2);

(4) and reducing the pooling template, and further performing pooling and Zhang-suen thinning on the image to obtain the robot path skeleton.

Further, in the step (1), the pooling process adopts a maximum pooling method, and the pixel point with the maximum feature value is used as a pooling result of the pooling template.

Further, in the step (3), the thinned pooled image is sampled and enlarged to the size of the original image to obtain an image a, and the image a and the original image are subjected to an and operation to complete the masking processing.

Further, before the step (1), the robot road original image needs to be preprocessed, and the preprocessing processes include gray level extraction, binarization, corrosion and expansion in sequence.

Adopt the beneficial effect that above-mentioned technical scheme brought:

aiming at the defects of high computational complexity and low execution efficiency of a classical path extraction algorithm, the invention provides a path skeleton extraction method based on image refinement, which eliminates redundant information of an image by using the concept of a reforming group, realizes Zhang-suen refinement according to a template with self-adaptive size change, reduces the computational complexity, accelerates the refinement speed and greatly reduces the time for completing the whole task on the premise of not obviously losing image information.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is a diagram of a robotic racetrack before refinement;

FIG. 3 is a diagram of a reforming cluster for an Ising model;

FIG. 4 is a schematic diagram of maximum pooling at template step size of 2;

FIG. 5 is a schematic diagram of masking.

Detailed Description

The technical scheme of the invention is explained in detail in the following with the accompanying drawings.

The invention designs a robot path skeleton extraction method based on image refinement, which comprises the following steps of:

step 1: selecting a proper step length to perform pooling treatment on the robot road original image by utilizing a reforming group idea;

step 2: performing Zhang-suen refinement on the pooling result obtained in the step 1;

and step 3: carrying out shade processing on the thinned pooled image in the step 2;

and 4, step 4: and reducing the pooling template, and further performing pooling and Zhang-suen thinning on the image to obtain the robot path skeleton.

One, realize

The method comprises the steps of firstly carrying out thinning processing on the original image with coarser granularity, firstly obtaining an approximate skeleton image, and meanwhile, using a masking operation to keep detailed contents.

1) Image features

Fig. 2 shows that the images before the thinning processing of the invention have the characteristics that the width of the guide line is fixed and far larger than one pixel, the images are simple, and only the path, the background and the like are included for the acquired track images. After pre-processing (gray scale, binarization, erosion, expansion), the black part in the image is a path after binarization and after erosion expansion, and the pixel width may be different for different application environments. By utilizing the coarse granularity of the target image, the running of the algorithm can be accelerated.

2) Reforming group

In theory, the Renormalization group (Renormalization group) is a mathematical method that refers to observing the system's changes at different scales, called scale transformation, and is closely related to scale invariance and conformal invariance.

The renormalization group coarsely granulizes the image to be processed using the scale invariance of the system (i.e., different scale scales are observed, the property being explored does not change), irrelevant features are attenuated in the scaling, and the observed information is preserved. For example, fig. 3 is a schematic diagram of a reforming group of an Ising model, the cell information of 3x3 in the Ising model is integrated into one cell after 3-fold scale change, and the information of a small cell is replaced by the information of a large-scale cell, so that the same critical property of the Ising model at different scales is researched.

3) Image pooling and refinement

For the image processed in the path recognition algorithm, the pixel information of N × N is mapped to a single large-scale pixel by using the concept of reforming group, which is called Pooling (Pooling), the unit cell C (N) with the size of N, and the image information represented by the set { C (N) } of C (N) is called image extraction T (N) with the granularity of N, thereby realizing the image processing with different granularities. In the process, extracting { t (N) | N ═ 1,2,4,8 … } of the images decomposed into different granularities is used to speed up the image processing.

Pooling, unlike convolution operations, acts on non-overlapping regions in the image. There are many ways of pooling, including average pooling, maximum pooling, random pooling, and the like. Considering that the images before and after the pooling in the task are both binary images, the invention selects the maximum pooling method. The maximum pooling refers to the pooling result of taking the pixel point with the maximum characteristic value as the template in the pooling template.

FIG. 4 is a schematic diagram of maximum pooling at template step size of 2. Let m be the pooled image and N be the size of the pooled template. Refining the pooled image to obtain m 'because the scale of the m' is 1/N of the original image²So the thinning speed of the image after the pooling is about N²And (4) doubling. And on the premise of no image distortion, a proper pooling template is selected, so that the effect of improving the image processing efficiency is obvious.

4) Shade cover

The thinned pooled image m' is the backbone extraction of the coarse granularity of the original image. And S (m ') in m ' is a point set of which the median value in m ' is 1. For the point x ∈ S (m'), the region of the original image N × N is corresponded. In order to extract detail information in the original image, it is necessary to up-sample and enlarge m' to the original image size, and perform an and operation with the original image m. FIG. 5 is a schematic diagram of masking, where each element m' is associated with a corresponding four-pixel composition structure phase of the original image during the masking operation.

Second, experiment

And carrying out gray extraction, binaryzation, expansion and refinement on the image in the test data. Testing the total time consumed for image processing using the classical Zhang-Suen refinement algorithm and different pooling granularities Zhang Suen. The time consumption of the refined algorithm and the classical algorithm of the invention is shown in the following table 1:

TABLE 1

Time (ms)	Classical Zhang-Suen	Pooling formwork 8 x 8	Pooling template 4 x 4	Pooling template 2 x 2
					Straight line	28.6	21.6	18.6	15.6
Right angle bend	31.7	17.7	17.7	13.7
					Crossroad	47.2	25.5	20.2	22.2
T-shaped road junction	49.3	24.3	21.3	18.3

Columns 3-5 in Table 1 indicate the time consumed by the method of the present invention under different pooling templates. Comparing the time consumption of the method of the invention with the time consumption of the classical method, the time optimization effect of the invention is obvious, and the time for processing a picture is reduced from the original 50ms to the average 20 ms. Therefore, the image processing speed can be accelerated, and the computing resources are saved. For the unsmooth phenomenon of the skeleton caused by coarse granularity refinement, all pixel points are comprehensively considered in the parameter extraction process, and the influence caused by the unsmooth phenomenon is reduced.

The whole path extraction algorithm comprises five steps of path graying, binaryzation, corrosion, expansion, skeleton extraction and the like. The experiment is mainly carried out on straight lines, right-angle bends, crossroads and T-shaped road junctions in the path, and the experimental results are shown in table 2.

TABLE 2

Time (ms)	Graying	Binarization method	Etching of	Expansion of	Skeletal extraction	Average time consumption
							Straight line	26	2.4	7.8	1.6	15.6	53.4
Right angle bend	26	2.4	7.8	2.5	13.7	52.4
							Crossroad	26	2.5	7.8	2	22.2	60.5
T-shaped road junction	26	2.5	7.8	1.9	18.3	54.5

The time required for the different stages is different due to the different principles and implementations of the different algorithms. Graying involves floating point operations, which take a long time. The binarization only traverses the image once, and the required time is very short. The image post-processing selects erosion and expansion, and the graph is traversed once in each operation, so that only a few operations are actually enough to meet the requirements. The most time consuming of skeleton extraction is the refinement operation. Experiments show that the Zhang-suen thinning algorithm has strong capability of processing complex path images, does not need to identify cross paths in advance, belongs to an algorithm with stronger universality, but has relatively higher time consumption and higher processing efficiency and can support the image refresh rate of 8 frames.

The embodiments are only for illustrating the technical idea of the present invention, and the technical idea of the present invention is not limited thereto, and any modifications made on the basis of the technical scheme according to the technical idea of the present invention fall within the scope of the present invention.

Claims

1. A robot path skeleton extraction method based on image refinement is characterized by comprising the following steps:

(3) performing mask processing on the thinned pooled image in the step (2);

2. The method for extracting the robot path skeleton based on the image refinement of claim 1, wherein in the step (1), the pooling process adopts a maximum pooling method, and the pixel point with the maximum feature value is used as the pooling result of the pooling template.

3. The method for extracting robot path skeleton based on image refinement of claim 1, wherein in step (3), the refined pooled image is sampled and enlarged to the size of an original image to obtain a map a, and the map a and the original image are subjected to an and operation to complete a masking process.

4. The method for extracting the robot path skeleton based on the image refinement as claimed in claim 1, wherein before the step (1), the robot road original image needs to be preprocessed, and the preprocessing processes are gray scale extraction, binarization, corrosion and expansion.