CN106910191A - A kind of bearing channel recognition methods based on depth image - Google Patents

Abstract

The invention discloses a kind of bearing channel recognition methods based on depth image, its feature includes：1 obtains bearing depth data；2 noise processeds；3 bearings are recognized；4 raceway grooves are recognized；5 raceway grooves are fitted；The present invention during polytype bearing is measured, can eliminate the subjectivity error of manual detection, eliminate abrasion of the measurement apparatus to bearing, and can quickly recognize bearing channel and its size, ensure detection speed and precision higher, strong robustness can be applied to all kinds of bearing channel detections.

Description

A kind of bearing channel recognition methods based on depth image

Technical field

Field of machine vision of the present invention, specifically a kind of bearing channel recognition methods based on depth image.

Background technology

Bearing is the critical component in various plant equipment, and its quality is directly connected to the runnability of whole equipment, So detection to bearing quality it is critical that.The quality pack of bearing contains quality and apparent size, is passed for apparent size The method of system is to use artificial mensuration, and efficiency is low and subjectivity is strong.In order to overcome disadvantage mentioned above, there are many scholars and work The advanced automatic detection systems of journey Shi Faming.

Zhang Ganqiang et al. (A of CN 105571552) proposes a kind of Bearing testing method, is measured using contact pin type sensor Ditch track data, is measured using contact digital sensor and plate washer to bearing diameter.Chen Zhenfeng (A of CN 105277161) A kind of detection method for bearing channel is proposed, it is passed through using a kind of device for possessing top board, base plate and helical duct Whether the pressure sensor on the outside of helical duct on punch-pin is qualified to detect bearing.Cheng Jiang (A of CN 103968791) et al. is proposed A kind of full-automatic comprehensive detection device of deep groove ball bearing, this device uses air cylinder driven, and height, profit are measured using elevation carrection table Internal diameter is measured with measuring stick, and plate and screw rod are followed by regulation and adapt to different types of bearing.

The automated detection method of above scholar and engineer all overcomes the shortcoming of traditional-handwork measurement, automaticity Height, but there is also clearly disadvantageous：First, their equipment is all contact type measurement, there is abrasion in itself to bearing.Second, Adaptability is not strong, can only be directed to the measurement of the bearing of one or several types, once change it is accomplished by readjusting even change Structure.Third, speed and precision be not high, the sensor that it is used can only rough measure size, tool can not be accurately acquired Body accurate parameters, differentiate that error rate is high.It is domestic at present also no a kind of for the full-automatic, contactless of bearing, strong robustness Measuring method.

The content of the invention

The present invention is in order to solve the weak point that above-mentioned prior art is present, there is provided a kind of bearing groove based on depth image Road recognition methods, to during polytype bearing is measured, eliminate the subjectivity error of manual detection, eliminates measurement Abrasion of the device to bearing, and can quickly recognize bearing channel and its size, it is ensured that detection speed and precision higher, robustness By force.

In order to solve the above technical problems, the present invention is adopted the following technical scheme that being：

A kind of the characteristics of bearing channel recognition methods based on depth image of the present invention, comprises the following steps：

Step 1, acquisition bearing depth data：

The depth image M of bearing, a width of W, the height of the depth image M are gathered using line laser sensor and motion platform It is H to spend；

Step 2, noise processed：

Noise of the depth image M intermediate values beyond bearing height scope is removed using threshold method, sliding window is recycled Variance method removes the noise spot that peels off in the depth image M, obtains by the depth image M' after noise processed；

Step 3, bearing identification：

3.1st, the bianry image M of depth image M' is obtained using local thresholding method^bin；

3.2nd, to the bianry image M^binClosing operation of mathematical morphology is carried out, so as to obtain the bianry image M after closed operation_c ^bin；

3.3rd, the bianry image M is obtained using minimum enclosed rectangle_c ^binIn the codomain of connection two boundary rectangle, it is maximum The boundary rectangle of area is the outline comprising bearing, the central point O (x of the boundary rectangle of maximum area_c,y_c) it is bearing Centre point；

Step 4, raceway groove identification：

4.1st, the centre point O (x according to bearing_c,y_c) obtain the bearing circle being located on bearing diameter direction in depth image M' Loop data, is designated as Q={ w₁,w₂,…,w_i,…,w_n}；w_iRepresent ith pixel point；Remember the ith pixel point w_iCoordinate is (x_i,y_i)；1≤i≤n；

4.2nd, raceway groove identification is carried out to the bearing annulus data Q using based on lower convex domain decomposition merging algorithm, obtains ditch Track data is designated as Q_c；

Step 5, raceway groove fitting：

5.1st, using least square method LS to the ditch track data Q_cIt is fitted, obtains the first solution (x of fitting circle₀,y₀, r)；Wherein, x₀Represent abscissa of the fitting circle of bearing channel in the depth image M', y₀Represent the fitting of bearing channel Abscissa of the circle in the depth image M', r represents the radius of the fitting circle of bearing channel；

5.2nd, according to the just solution (x₀,y₀, r) using particle swarm optimization algorithm to the ditch track data Q_cOptimize and ask Solution, obtains the circular fitting parameter (x', y', r') of bearing channel.

The characteristics of present invention bearing channel recognition methods based on depth image according to claim 1, lies also in： Sliding window variance method SWVM in the step 2 is carried out as follows：

Step 2.1, definition sliding window：Define the length of side for g square window, on depth image with from left to right, from When the order of top to bottm is slided, the pixel positioned at window center position is denoted as p_x,y, coordinate is (x, y), then p_x,yBoundary constraint It is：X ∈ [(g-1)/2, W- (g-1)/2], y ∈ [(g-1)/2, H- (g-1)/2]；Note Ω_xyFor in depth image M', center pixel Coordinate is the pixel set in the square window of (x, y), thenWherein, p^(j) _xy Represent j-th pixel in square window, and pixel set Ω_xyCentral pixel point be

Step 2.2, initialization x=(g-1)/2-1, y=(g-1)/2；

Step 2.3, abscissa border detection：X+1 is assigned to x, judges whether x=W- (g-1)/2 sets up, if so, X=(g-1)/2-1 is then made, after y+1 is assigned into y, step 2.4 is performed, step 2.5 is otherwise performed；

Step 2.4, ordinate border detection：Judge whether y≤H- (g-1)/2 sets up, if so, step 2.5 is then performed, Otherwise, algorithm is terminated, so as to obtain the depth image M' for removing the noise that peels off；

Step 2.5, judge central pixel pointWhether it is non-zero points；If non-zero points, then enter step 2.6, if not Non-zero points, return and perform step 2.3；

Step 2.6, acquisition non-zero points set omega_xy ^v：

Step 2.6.1, initialization j=0；

Step 2.6.2, judge j=g²Whether set up, if set up, obtain non-zero point set Represent k-th non-zero points in corresponding square window, wherein N≤g², and And step 2.7 is performed, otherwise, j+1 is assigned to j, and judge j=g²Whether/2 set up, if set up, return to step 2.6.2, Otherwise perform step 2.6.3；

Step 2.6.3, judge j-th pixel p^(j) _xyWhether it is non-zero points, if non-zero points, then by p^(j) _xyIt is put into non- Zero point set omega_xy ^vIn, otherwise abandon j-th pixel p^(j) _xy；And return to step 2.6.2 is performed；

Step 2.7, using formula (1) obtain non-zero points set omega_xy ^vIn all effective pixel points spatial spreading degree

In formula (1), d_mnRepresent the Euclidean distance between m-th non-zero points and n-th non-zero points；And 1≤m ≠ n≤N；

Step 2.8, setting threshold value Τ₂If,Step 2.9 is then transferred to, step 2.3 is otherwise transferred to；

Step 2.9, available point set omega is represented with S_xy ^vThe variance of interior element, non-zero points set omega is represented with E_xy ^vInterior element Average；Threshold value T is set₃If, S>T₃, then by the central pixel point of square windowValue be set to " 0 ", otherwise protect Stay central pixel pointValue, and return to step 2.3.

Decompose merging algorithm DCBLCD based on lower convex domain in the step 4.2 is carried out as follows：

Step a, to the bearing annulus data Q={ w₁,w₂,…,w_i,…,w_nDecomposed：

Step a.1, definition stepped intervals be Δ, define B_fIt is the segmentation that bearing annulus data Q is decomposited, f justifies for bearing The subscript of the segmentation that loop data Q is decomposited, initializes i=1, f=1；

Step a.2, initialization Δ=2；

Step a.3, order segmentation B_f={ w_i,w_i+1,...,w_i+Δ, then it is segmented B_fThe curvilinear equation of composition is g (x), remembers w_zWith w_vIt is set B_fThe point of middle any two difference coordinate, coordinate is respectively (x_z,y_z) and (x_v,y_v), then there are g (x_z)=y_z,g(x_v) =y_v, and obtain w using formula (2)_zAnd w_vBetween under it is convex away from D_zv, so as to obtain segmentation B_fIn between all two differences Under it is convex away from, and constitute under it is convex away from set

D_zv=α g (x_z)+βg(x_v)-g(αx_z+βx_v)+τ (2)

α, β are scale factor in formula (2), and alpha+beta=1, and τ is constant；

Step a.4, use D_uIt is convex away from set under expressionIn u-th under it is convex away from then Wherein ε=Δ (1+ Δs)/2；JudgeWhether set up, if so, then perform step a.5, otherwise, perform step a.6；

A.5, by f-th step is segmented B_fWhether it is put into segmentation set Γ, and after i+ Δs are assigned into i, judges i >=n-1 Set up, if so, then represent and obtain segmentation set Γ={ B₁,B₂,...,B_γ,...,B_f, wherein B_γRepresent and be segmented in set Γ The γ segmentation, and perform step b, otherwise, f+1 is assigned to f, and return to step is a.2；

Step a.6, Δ+1 is assigned to Δ after, perform step a.3；

Step b, each segmentation to being segmented in set Γ are merged：

Step b.1, initialization γ=1, counting variable θ=0；

Step b.2, stepped intervals σ=1,

Step b.3, define B_γ,σExpression starting point is B_γ, and the constituted merging section of continuous σ segmentation, i.e.,By B_γ,σComprising annulus data point total number be designated as c；

B.4, according to formula (2) step calculates merging section B_γ,σIt is convex away from so as to obtain down under between middle any two difference It is convex away from setWherein λ=c (c-1)/2；

Step b.5, statistics under it is convex away from setIn less than 0 point number, be denoted as c^-, and calculate scale factor

Step b.6, defined parameters κ for control it is described under it is convex away from setThe factor of middle negative ratio, if η>κ, then hold B.8 b.7 row step, otherwise perform step；

Step b.7, section B will be merged_γ,σIt is put into merging section set H, and after γ+σ+1 are assigned into γ, judges that γ=f is No establishment, if so, then represent to obtain to merge to be segmented and gather It is in HIndividual merging section, And perform step c, otherwise return to step b.3；

Step b.8, σ+1 is assigned to σ after, judge whether γ+σ+1=f set up, if so, then represent obtain merge section SetOtherwise, after θ+1 being assigned into θ, return to step is b.2；

Step c, from section set H is merged filter out ditch track data Q_c：

NoteRepresentComprising annulus data point number, selected comprising a point number from the merging section set H At mostAs the ditch track data Q_c, i.e.,

Compared to tradition machinery formula detection method, the beneficial effects of the invention are as follows：

1st, the present invention uses software automatic survey, it is not necessary to the participation of people, is missed so as to avoid the subjective of manual detection Difference, improves the uniformity of Bearing testing.

2nd, the present invention can effectively remove bearing data noise using sliding window variance method SWVM is slided, and reduce and make an uproar Sound recognizes the influence with measurement to the later stage, improves certainty of measurement.

3rd, the present invention can be identified effectively using based on lower convex domain decomposition merging DCBLCD from bearing annulus data Raceway groove circular arc data, and circular arc border is accurately determined, recognition accuracy is high.

4th, the robustness of this method is high, and different types of bearing can be detected, without complicated mechanical machine The design of structure and conversion.

Brief description of the drawings

Fig. 1 is recognition methods block diagram of the present invention；

Fig. 2 is the bearing channel identification region measurement data figure of recognition methods of the present invention；

Fig. 3 is the bearing channel datagram for having identified shown in recognition methods of the present invention；

Fig. 4 is the bearing channel data circular fitting design sketch shown in recognition methods of the present invention.

Specific embodiment

In the present embodiment, a kind of bearing channel recognition methods based on depth image has that speed is fast, high precision, robustness Strong the characteristics of, all kinds of bearing channel detections are can apply to, its step includes 1 data acquisition, 2 noise filterings, the identification of 3 bearings, 4 Raceway groove is recognized and the fitting of 5 raceway grooves.Wherein：

The noise filtering of the 2nd step, is using sliding window variance method SWVM (SlidingWindowVariance Method), differentiated by calculating the space length of data and the variance of value in image block positioned at window center data whether It is noise；

3rd step bearing is recognized, is that bearing is identified from background data, using Minimum Enclosing Rectangle method MER (Minimum ExternalRectangle), first obtains the profile of image, then according to the minimum external of profile with morphological approach Rectangular characteristic carries out handsome choosing, obtains inner circle excircle configuration；

4th step raceway groove is recognized, is that raceway groove circular arc data are identified, and algorithm is merged using being decomposed based on lower convex domain DCBLCD (Decomposition andConsolidationAlgorithmBasedonLower ConvexDomain) is recognized Raceway groove；

5th step raceway groove is fitted, and is to obtain just solution using least square method LS (Least Square) fitting circule methods, then with grain Swarm optimization PSO (Particle Swarm Optimization) is optimized, and then obtains precision circular fitting higher； Specifically, as shown in figure 1, being to carry out as follows：

Step 1, acquisition bearing depth data：

Line laser sensor is placed in the surface of motion platform, bearing is placed on the moving platform, using line laser Sensor and motion platform gather the depth image M of bearing, a width of W of depth image M, are highly H；

Step 2, noise processed：

In order to reduce operand, first with threshold method removal depth image M intermediate values substantially beyond bearing height scope Discrete noise, sets threshold value T₁, threshold value T herein₁1.5 times of bearing maximum height are set to, sliding window variance method is recycled The noise spot that peels off with good data mixing in SWVM removal depth images M, obtains by the depth image after noise processed M', specifically, SWVM algorithms are carried out in accordance with the following steps：

Step 2.1, definition sliding window：It is the square window of g to define the length of side, and g values are bigger, and operand is bigger, and value is smaller, Locality is stronger, and g generally takes 5, when being slided with order from left to right, from top to bottom on depth image, positioned at window center The pixel of position is denoted as p_x,y, coordinate is (x, y), then p_x,yBoundary constraint be：X ∈ [(g-1)/2, W- (g-1)/2], y ∈ [(g-1)/2,H-(g-1)/2]；Note Ω_xyFor in depth image M', center pixel coordinate is the pixel in the square window of (x, y) Set, thenWherein, p^(j) _xyJ-th pixel in square window is represented, and Pixel set Ω_xyCentral pixel point be

Step 2.2, initialization x=(g-1)/2-1, y=(g-1)/2；

Step 2.3, abscissa border detection：X+1 is assigned to x, judges whether x=W- (g-1)/2 sets up, if so, X=(g-1)/2-1 is then made, after y+1 is assigned into y, step 2.4 is performed, step 2.5 is otherwise performed；

Step 2.4, ordinate border detection：Judge whether y≤H- (g-1)/2 sets up, if so, step 2.5 is then performed, Otherwise, algorithm is terminated, so as to obtain the depth image M' for removing the noise that peels off；

Step 2.5, judge central pixel pointWhether it is non-zero points；If non-zero points, then enter step 2.6, if not Non-zero points, return and perform step 2.3；

Step 2.6, acquisition non-zero points set omega_xy ^v：

Step 2.6.1, initialization j=0；

Step 2.6.2, judge j=g²Whether set up, if set up, obtain non-zero point set Represent k-th non-zero points in corresponding square window, wherein N≤g², and And step 2.7 is performed, otherwise, j+1 is assigned to j, and judge j=g²Whether/2 set up, if set up, return to step 2.6.2, Otherwise perform step 2.6.3；

Step 2.6.3, judge j-th pixel p^(j) _xyWhether it is non-zero points, if non-zero points, then by p^(j) _xyIt is put into non- Zero point set omega_xy ^vIn, otherwise abandon j-th pixel p^(j) _xy；And return to step 2.6.2 is performed；

Step 2.7, using formula (1) obtain non-zero points set omega_xy ^vIn all effective pixel points spatial spreading degree

In formula (1), d_mnRepresent the Euclidean distance between m-th non-zero points and n-th non-zero points；And 1≤m ≠ n≤N.From Definition can be seen thatIt is Ω_xy ^vIn between all non-zero points and the non-zero points closest with it distance average value,Table Show the spatial distribution characteristic of data point in window,Bigger, data distribution is more discrete,Smaller, data get over uniform close, when When represent in window W there is no zero point.

Step 2.8, setting threshold value Τ₂, in the present embodiment, take Τ₂It is 1, represents that the non-zero points in window are non-with its arest neighbors Average distance between zero point is 1, ifRepresent that the spatial spreading degree of data in window is larger, exceeded setting Threshold value, then be transferred to step 2.9, is otherwise transferred to step 2.3；

Step 2.9, available point set omega is represented with S_xy ^vThe variance of interior element, non-zero points set omega is represented with E_xy ^vInterior element Average；Threshold value T is set₃If, S>T₃, then by the central pixel point of square windowValue be set to " 0 ", i.e., quite In removing the point.Otherwise retain central pixel pointValue, and return to step 2.3.

Step 3, bearing identification：

3.1st, the bianry image M of depth image M' is obtained using local thresholding method^bin, increased income using image algorithm in this example AdaptiveThreshold () function in storehouse opencv 2 is processed；

3.2nd, to bianry image M^binClosing operation of mathematical morphology is carried out, is increased income in storehouse opencv 2 using image algorithm in this example MorphologyEx () function pair M^binProcessed, so as to obtain the bianry image M after closed operation_c ^bin；

3.3rd, bianry image M is obtained using minimum enclosed rectangle (MinimumExternal Rectangle, MER)_c ^binIn The codomain of connection two boundary rectangle, minAreaRect () is increased income using image algorithm in storehouse opencv 2 in this example to calculate The minimum enclosed rectangle of two-value connected domain, because only this target of bearing in detection image, does not have other objects, so most The boundary rectangle of large area is the outline comprising bearing, the central point O (x of the boundary rectangle of maximum area_c,y_c) it is bearing Centre point；

Step 4, raceway groove identification：

4.1st, the centre point O (x according to bearing_c,y_c) obtain the bearing circle being located on bearing diameter direction in depth image M' Raceway groove identification numeric field data in loop data, such as Fig. 2 is exactly to be located at data on bearing annulus, is designated as Q={ w₁,w₂,…,w_i,…, w_n}；w_iRepresent ith pixel point；Note ith pixel point w_iCoordinate is (x_i,y_i)；1≤i≤n；

4.2nd, merging algorithm (Decomposition and Consolidation are decomposed using based on lower convex domain AlgorithmBased on Lower ConvexDomain, DCBLCD) raceway groove identification is carried out to bearing annulus data Q, obtain Ditch track data as shown in Figure 3 is designated as Q_c, Q_cIdeal zone corresponding diagram 2 in channel region, it is a central angle at 90 degree The circular arc of left and right.Q_cAcquisition be broadly divided into decomposition and union operation to data in Q, wherein operation splitting is the number in Q According to several segmentations are decomposed into, each segmentation is a set for data, and union operation be by above-mentioned segmentation adjacent to each other by According to the merging of certain rule selectivity.Specifically follow the steps below：

Step a, to bearing annulus data Q={ w₁,w₂,…,w_i,…,w_nDecomposed：

Step a.1, definition stepped intervals be Δ, define B_fIt is the segmentation that bearing annulus data Q is decomposited, f justifies for bearing The subscript of the segmentation that loop data Q is decomposited, initializes i=1, f=1；

Step a.2, initialization Δ=2；

Step a.3, order segmentation B_f={ w_i,w_i+1,...,w_i+Δ, then it is segmented B_fThe curvilinear equation of composition is g (x), remembers w_zWith w_vIt is set B_fThe point of middle any two difference coordinate, coordinate is respectively (x_z,y_z) and (x_v,y_v), then there are g (x_z)=y_z,g(x_v) =y_v, and obtain w using formula (2)_zAnd w_vBetween under it is convex away from D_zv, so as to obtain segmentation B_fIn between all two differences Under it is convex away from, and constitute under it is convex away from set

D_zv=α g (x_z)+βg(x_v)-g(αx_z+βx_v)+τ (2)

α, β are scale factor in formula (2), and alpha+beta=1, in this example the equal value of α, β for 0.5, τ be constant, be an error Tolerance parameter, in a lower convex domain, in theory, and when τ=0, D_zv＞ 0 is permanent to be set up, but in actual acquired data because The presence of noise data, D_zv＞ 0 is frequently not permanent establishment, so needing to set a constant factor τ for tolerance noise, this example Middle value is 0.01；

Step a.4, use D_uIt is convex away from set under expressionIn u-th under it is convex away from then Wherein ε=Δ (1+ Δs)/2；JudgeWhether set up, if so, illustrate to be segmented B_fIn have and be unsatisfactory for lower convexity matter Two points, do not continue to extension, perform step a.5, otherwise, continue to extend, and perform step a.6；

A.5, by f-th step is segmented B_fWhether it is put into segmentation set Γ, and after i+ Δs are assigned into i, judges i >=n-1 Set up, if so, then represent and obtain segmentation set Γ={ B₁,B₂,...,B_γ,...,B_f, wherein B_γRepresent and be segmented in set Γ The γ segmentation, and perform step b, otherwise, f+1 is assigned to f, and return to step is a.2；

Step a.6, Δ+1 is assigned to Δ after, perform step a.3；

Step b, each segmentation to being segmented in set Γ are merged：

Step b.1, initialization γ=1, counting variable θ=0；

Step b.2, stepped intervals σ=1,

Step b.3, define B_γ,σExpression starting point is B_γ, and the constituted merging section of continuous σ segmentation, i.e.,By B_γ,σComprising annulus data point total number be designated as c；

B.4, according to formula (2) step calculates merging section B_γ,σIt is convex away from so as to obtain down under between middle any two difference It is convex away from setWherein λ=c (c-1)/2；

Step b.5, statistics under it is convex away from setIn less than 0 point number, be denoted as c^-, and calculate scale factor The value is bigger, illustrates that two segmentation differences are bigger, is more not belonging to same figure below domain.

Step b.6, defined parameters κ for control under it is convex away from setThe factor of middle negative ratio, κ is taken in this example is 0.2, if η>B.7 κ, then perform step, otherwise performs step b.8；

Step b.7, section B will be merged_γ,σIt is put into merging section set H, and after γ+σ+1 are assigned into γ, judges that γ=f is No establishment, if so, then represent to obtain to merge to be segmented and gather It is in HIndividual merging section, And perform step c, otherwise return to step b.3；

Step b.8, σ+1 is assigned to σ after, judge whether γ+σ+1=f set up, if so, then represent obtain merge section SetOtherwise, after θ+1 being assigned into θ, return to step is b.2；

Step c, from section set H is merged filter out ditch track data Q_c：

NoteRepresentComprising annulus data point number, selected from section set H is merged most comprising point numberAs ditch track data Q_c, i.e.,

Step 5, raceway groove fitting：

5.1st, using least square method LS to ditch track data Q_cIt is fitted, obtains the first solution (x of fitting circle₀,y₀,r)；Its In, x₀Represent abscissa of the fitting circle of bearing channel in depth image M', y₀Represent the fitting circle of bearing channel in depth map As the ordinate in M', r represents the radius of the fitting circle of bearing channel；

5.2nd, it is not optimal solution by the solution that least square method LS is obtained is a kind of estimation, so according to just solution (x₀,y₀, particle swarm optimization algorithm (Particle Swarm Optimization Algorithm, PSO) r) is reused to raceway groove Data Q_cSolution is optimized, the circular fitting parameter (x', y', r') of bearing channel is obtained, there is the preferable just solution, institute of LS It is very fast with PSO convergence rates.Final fitting effect such as Fig. 4, the point of solid black is the ditch track data for identifying, and black surround Hollow point is final round matched curve.

Embodiments of the invention are these are only, the scope of the claims of the invention is not thereby limited, it is every to be said using the present invention Equivalent structure or equivalent flow conversion that bright book and accompanying drawing content are made, or directly or indirectly it is used in other related technology necks Domain, is included within the scope of the present invention.

Claims (3)

1. a kind of bearing channel recognition methods based on depth image, it is characterised in that comprise the following steps：

Step 1, acquisition bearing depth data：

Using the depth image M, a width of W of the depth image M of line laser sensor and motion platform collection bearing, highly it is H；

Step 2, noise processed：

Noise of the depth image M intermediate values beyond bearing height scope is removed using threshold method, sliding window variance is recycled Method removes the noise spot that peels off in the depth image M, obtains by the depth image M' after noise processed；

Step 3, bearing identification：

3.1st, the bianry image M of depth image M' is obtained using local thresholding method^bin；

3.2nd, to the bianry image M^binClosing operation of mathematical morphology is carried out, so as to obtain the bianry image M after closed operation_c ^bin；

3.3rd, the bianry image M is obtained using minimum enclosed rectangle_c ^binIn the codomain of connection two boundary rectangle, maximum area Boundary rectangle be the outline comprising bearing, the central point O (x of the boundary rectangle of maximum area_c,y_c) it is the center of circle of bearing Point；

Step 4, raceway groove identification：

4.1st, the centre point O (x according to bearing_c,y_c) obtain the bearing annulus number being located on bearing diameter direction in depth image M' According to being designated as Q={ w₁,w₂,…,w_i,…,w_n}；w_iRepresent ith pixel point；Remember the ith pixel point w_iCoordinate is (x_i, y_i)；1≤i≤n；

4.2nd, raceway groove identification is carried out to the bearing annulus data Q using based on lower convex domain decomposition merging algorithm, obtains raceway groove number According to being designated as Q_c；

Step 5, raceway groove fitting：

5.1st, using least square method LS to the ditch track data Q_cIt is fitted, obtains the first solution (x of fitting circle₀,y₀,r)；Its In, x₀Represent abscissa of the fitting circle of bearing channel in the depth image M', y₀Represent the fitting circle of bearing channel in institute The abscissa in depth image M' is stated, r represents the radius of the fitting circle of bearing channel；

5.2nd, according to the just solution (x₀,y₀, r) using particle swarm optimization algorithm to the ditch track data Q_cSolution is optimized, is obtained To the circular fitting parameter (x', y', r') of bearing channel.

2. the bearing channel recognition methods based on depth image according to claim 1, it is characterized in that：In the step 2 Sliding window variance method SWVM carry out as follows：

Step 2.1, definition sliding window：Define the length of side for g square window, on depth image with from left to right, on to Under order slide when, the pixel positioned at window center position is denoted as p_x,y, coordinate is (x, y), then p_x,yBoundary constraint be：x ∈ [(g-1)/2, W- (g-1)/2], y ∈ [(g-1)/2, H- (g-1)/2]；Note Ω_xyFor in depth image M', center pixel coordinate It is the pixel set in the square window of (x, y), thenWherein, p^(j) _xyExpression side J-th pixel in shape window, and pixel set Ω_xyCentral pixel point be

Step 2.2, initialization x=(g-1)/2-1, y=(g-1)/2；

Step 2.3, abscissa border detection：X+1 is assigned to x, judges whether x=W- (g-1)/2 sets up, if so, then make x =(g-1)/2-1, after y+1 is assigned into y, performs step 2.4, otherwise performs step 2.5；

Step 2.4, ordinate border detection：Judge whether y≤H- (g-1)/2 sets up, if so, step 2.5 is then performed, it is no Then, algorithm is terminated, so as to obtain the depth image M' for removing the noise that peels off；

Step 2.5, judge central pixel pointWhether it is non-zero points；If non-zero points, then enter step 2.6, if not non-zero Point, returns and performs step 2.3；

Step 2.6, acquisition non-zero points set omega_xy ^v：

Step 2.6.1, initialization j=0；

Step 2.6.2, judge j=g²Whether set up, if set up, obtain non-zero points set omega_xy ^v={ p '_xy ⁽¹⁾,p ′_xy ⁽²⁾,...p′_xy ^(k)...p′_xy ^(N),Represent k-th non-zero points in corresponding square window, wherein N≤g², and simultaneously perform Step 2.7, otherwise, is assigned to j, and judge j=g by j+1²Whether/2 set up, if set up, return to step 2.6.2 otherwise holds Row step 2.6.3；

Step 2.6.3, judge j-th pixel p^(j) _xyWhether it is non-zero points, if non-zero points, then by p^(j) _xyIt is put into non-zero point set Close Ω_xy ^vIn, otherwise abandon j-th pixel p^(j) _xy；And return to step 2.6.2 is performed；

Step 2.7, using formula (1) obtain non-zero points set omega_xy ^vIn all effective pixel points spatial spreading degree l：

In formula (1), d_mnRepresent the Euclidean distance between m-th non-zero points and n-th non-zero points；And 1≤m ≠ n≤N；

Step 2.8, setting threshold value T₂If, l >=T₂, then step 2.9 is transferred to, otherwise it is transferred to step 2.3；

Step 2.9, available point set omega is represented with S_xy ^vThe variance of interior element, non-zero points set omega is represented with E_xy ^vInterior element it is equal Value；Threshold value T is set₃If, S>T₃, then by the central pixel point of square windowValue be set to " 0 ", otherwise retain in Imago vegetarian refreshmentsValue, and return to step 2.3.

3. the bearing channel recognition methods based on depth image according to claim 1, it is characterized in that：The step 4.2 In based on lower convex domain decompose merge algorithm DCBLCD carry out as follows：

Step a, to the bearing annulus data Q={ w₁,w₂,…,w_i,…,w_nDecomposed：

Step a.1, definition stepped intervals be Δ, define B_fIt is the segmentation that bearing annulus data Q is decomposited, f is bearing annulus data The subscript of the segmentation that Q is decomposited, initializes i=1, f=1；

Step a.2, initialization Δ=2；

Step a.3, order segmentation B_f={ w_i,w_i+1,...,w_i+Δ, then it is segmented B_fThe curvilinear equation of composition is g (x), remembers w_zAnd w_vFor Set B_fThe point of middle any two difference coordinate, coordinate is respectively (x_z,y_z) and (x_v,y_v), then there are g (x_z)=y_z,g(x_v)=y_v, And obtain w using formula (2)_zAnd w_vBetween under it is convex away from D_zv, so as to obtain segmentation B_fIn between all two differences under it is convex Away from, and it is convex away from set under composition

D_zv=α g (x_z)+βg(x_v)-g(αx_z+βx_v)+τ (2)

α, β are scale factor in formula (2), and alpha+beta=1, and τ is constant；

Step a.4, use D_uIt is convex away from set under expressionIn u-th under it is convex away from then Wherein ε=Δ (1+ Δs)/2；JudgeWhether set up, if so, then perform step a.5, otherwise, perform step a.6；

A.5, by f-th step is segmented B_fIt is put into segmentation set Γ, and after i+ Δs are assigned into i, judges whether i >=n-1 sets up, If so, then represent and obtain segmentation set Γ={ B₁,B₂,...,B_γ,...,B_f, wherein B_γRepresent the in segmentation set Γ γ segmentation, and step b is performed, otherwise, f+1 is assigned to f, and return to step is a.2；

Step a.6, Δ+1 is assigned to Δ after, perform step a.3；

Step b, each segmentation to being segmented in set Γ are merged：

Step b.1, initialization γ=1, counting variable θ=0；

Step b.2, stepped intervals σ=1,

Step b.3, define B_γ,σExpression starting point is B_γ, and the constituted merging section of continuous σ segmentation, i.e.,Will B_γ,σComprising annulus data point total number be designated as c；

B.4, according to formula (2) step calculates merging section B_γ,σUnder between middle any two difference it is convex away from so that obtain down it is convex away from SetWherein λ=c (c-1)/2；

Step b.5, statistics under it is convex away from setIn less than 0 point number, be denoted as c^-, and calculate scale factor

Step b.6, defined parameters κ for control it is described under it is convex away from setThe factor of middle negative ratio, if η>κ, then perform step Suddenly b.7, otherwise step is performed b.8；

Step b.7, section B will be merged_γ,σBe put into merging section set H, and after γ+σ+1 are assigned into γ, judge γ=f whether into It is vertical, if so, then represent and obtain merging segmentation set It is in HIndividual merging section, and hold Row step c, otherwise return to step are b.3；

Step b.8, σ+1 is assigned to σ after, judge whether γ+σ+1=f set up, if so, then represent obtain merge section setOtherwise, after θ+1 being assigned into θ, return to step is b.2；

Step c, from section set H is merged filter out ditch track data Q_c：

NoteRepresentComprising annulus data point number, selected from the merging section set H most comprising point numberAs the ditch track data Q_c, i.e.,