Path searching method of unmanned vehicle

Abstract

The invention provides a path searching method for an unmanned vehicle, which comprises the following steps of: extracting feature points of an image by using an image feature detection algorithm, obtaining a key turning point according to the feature points of the image, and storing the key turning point into a starting table; performing data expansion for three times between the two key turning points of the optimization path, and inserting virtual points to smooth the optimization path; the loss amount distributed according to needs is given and serves as a part of an objective function in path optimization; and the objective function is minimized according to a cost minimum principle, and an excitation type path search algorithm is given to obtain an optimal path. According to the method, the optimal path is globally searched, and a traffic route planning problem is converted into an optimization problem to be solved, so that the trouble that a map is fast in aging and needs to be continuously updated is solved, the anti-noise capability of navigation positioning is enhanced, and obstacles are effectively avoided. According to the method, a loss evaluation function is used, so that the real-time performance of path planning is enhanced, and traffic emergencies can be effectively handled.

Description

Technical Field

[0001] The present invention relates to the technical field of image detection in the field of electronic information, and in particular to a path search method for an unmanned vehicle. Background Art

[0002] With the rapid development of economy and society, urban traffic congestion is becoming more and more serious, especially in some megacities where roads are crowded, blocked and traffic accidents are frequent. Therefore, unmanned vehicles require more accurate, safe and reasonable navigation to avoid traffic jams and reduce the accident rate. It is necessary to install a real-time path planning system on the unmanned vehicle. The system must have the function of global optimal path planning with optimized calculation, accurate navigation and rapid positioning. Therefore, it is urgent to study the optimal path planning method for unmanned vehicles, which has important theoretical value and practical significance.

[0003] Image features refer to an indicator that reflects the discontinuity of local characteristics of an image, which indicates the end of one area and the beginning of another. Existing path planning methods can be summarized as map pre-import method and local map construction method during driving. The map pre-import method can provide high-precision and intuitive navigation effects when the environment and target area are basically unchanged. It is suitable for static environments because the path is fixed and the navigation process is relatively reliable and accurate. However, it has poor real-time performance and high maintenance costs. With the changes of time and environment, the map data ages rapidly and needs to be updated frequently to maintain positioning accuracy. Since this method lacks the ability to respond quickly to dynamic changes, the delay and lag phenomenon are obvious, so it cannot meet the high requirements of today's unmanned driving systems for real-time and flexibility. The local map construction method during driving can update the map in real time as the vehicle travels, closely fitting the real scene, thereby maintaining the accuracy of navigation and positioning. This method responds more naturally and objectively when the environment changes. However, its disadvantages are that it is easily interfered by noise, the map construction process is complex and time-consuming, and the real-time performance is still insufficient. In summary, the existing unmanned vehicles have requirements for path planning, including good real-time performance, strong noise resistance, accurate navigation and positioning, intelligent obstacle avoidance, and the ability to respond quickly to emergencies. However, the map pre-import method is difficult to meet high real-time requirements due to its slow response to environmental changes and reduced positioning accuracy as the environment changes; and the local map construction method has shortcomings in anti-interference ability and obstacle handling, and cannot provide sufficient accuracy for the unmanned driving system. Therefore, the existing path planning methods face great challenges in the field of unmanned driving, and need to be further improved and optimized to meet these complex requirements. Summary of the invention

[0004] In view of the technical problem that current unmanned vehicles have no suitable path planning method to meet the requirements of high real-time and high precision, resulting in certain safety hazards, the present invention proposes a path searching method for unmanned vehicles, which improves efficiency and safety.

[0005] In order to achieve the above object, the technical solution of the present invention is implemented as follows: a path search method for an unmanned vehicle, the steps of which are as follows:

[0006] S1, extracting feature points of the image using an image feature detection algorithm, obtaining key turning points based on the feature points of the image, and storing the key turning points in a starting table;

[0007] S2, three data expansions are performed between two key turning points of the optimization path, and virtual points are inserted to smooth the optimization path;

[0008] S3, gives the amount of loss allocated on demand and is used as part of the objective function in path optimization;

[0009] S4, minimize the objective function according to the minimum cost principle, give an excitation path search algorithm, and obtain the optimal path.

[0010] Preferably, the method of extracting the feature points of the image using the image feature detection algorithm is: using a matrix operator to extract the features of the traffic road condition map in front of the unmanned vehicle, and respectively obtaining the lateral gradient values ​​G x and the longitudinal gradient value G y , if the total gradient value in both directions is greater than the threshold δ, it is a feature point of the image.

[0011] Preferably, the lateral gradient value G is obtained by convolving the matrix operator with the traffic map. x and the longitudinal gradient value G y ;

[0012] The total gradient value G in the two directions is G = |G x |+|G y |;

[0013] The matrix operator is a 3*3 matrix operator;

[0014]

[0015] Among them, A and B are matrix operators, and I is the traffic map; is the convolution operator.

[0016] Preferably, the method for obtaining the key turning points based on the feature points of the image is: taking the total gradient value and gradient direction corresponding to the feature points of the image as feature information to construct a feature edge image; filtering the feature edge image using a low-pass filter, binarizing the filtered image, and the points with a pixel value of 1 in the binarized image are key turning points.

[0017] Preferably, the method of filtering the characteristic edge image using a low-pass filter is:

[0018] A low-pass filter H is used to filter the feature edge image f in the spatial domain. ′ (x, y) is smoothed to remove noise and redundant information, and the output image g(u, v) is:

[0019] g(u,v)=∑ x ∑ y f ′ (x,y)H(u-x+1,v-y+1);

[0020] Among them, u, v represent the pixel coordinates in the output image g(u, v), and x, y represent the feature edge image f ′ Pixel coordinates in (x,y);

[0021] The method for binarizing the filtered image is:

[0022] Select the threshold δ to binarize the output image g(u,v), and the processed image f(x,y) is:

[0023]

[0024] The key turning points appear on the binarized image f(x,y);

[0025] The starting table is used to record the initial key point information of the path, and the starting value is the gray value of the output image g(0,0); the path area of ​​interest is quickly located and processed through the starting table.

[0026] Preferably, the method for performing cubic data expansion between two key turning points of the optimization path is: adding virtual points between the two key turning points by selecting a cubic function with continuous function values ​​and first-order derivative values ​​at the endpoints of the two key turning points, and the cubic function is obtained based on the interpolation method.

[0027] Preferably, the cubic function is solved as follows: S(x) represents a cubic function interpolated in the horizontal direction X of the position node in the vertical direction Y, where X refers to the abscissa set of the position node, Y refers to the ordinate set of the position node, y=S(x), x∈X, y∈Y;

[0028] The second-order derivative of the cubic function S(x) at each node is S ″ (x k )=M k , k=0,1,…,λ; where M k Represents the location node x k The second-order derivative value at , λ represents the total number of position nodes in the path;

[0029] The second-order derivative of the cubic function S(x) is a linear function, that is,

[0030]

[0031] Among them, h k-1 =x k -x k-1 , M k-1 Represents the location node x k-1 The second derivative value at ;

[0032] For the second-order derivative S ″ (x) twice, and according to the node conditions: S(x k-1 )=y k-1 and S(x k )=y k , we can get the cubic interpolation function S(x).

[0033] Preferably, different loss amounts are defined in different road sections, and the loss amounts are determined according to the probability distribution of traffic flow density and congestion intensity;

[0034] The method for determining the loss amount is:

[0035] The occurrence of congestion is random, and the traffic flow on the road follows a negative exponential distribution. When the traffic flow density on the road is ξ(t), the probability of k congestion arrivals in the time period (0, t) is:

[0036]

[0037] Among them, {N(t),t≥0} is the number of congestion arrivals within (0,t];

[0038] The i-th road congestion intensity Y i The probability distribution function is: G(x) = P{Y i ≤x}, i=1,2,…; Assume that the congestion intensity at different intervals is independent and has the same probability distribution, G *(i) (x) is represented as the i-fold convolution of the probability distribution function G(x), where x represents the upper limit of the road congestion level;

[0039] Assume that the function of the loss of fuel consumption, vehicle wear and time caused by traffic congestion is F(t), and the interval time is T; when the congestion level on the traffic road reaches the threshold h, the processing capacity of traffic management drops to an unavailable state, then before time T, the congestion loss rate function is:

[0040]

[0041] Among them, S h represents the traffic congestion variable when approaching the threshold h;

[0042] Since traffic congestion requires a unit loss of c to restore, the total restoration cost C(T) is:

[0043]

[0044] Preferably, the implementation method of step S4 is: if the cost calculated by the total recovery cost C(T) is the smallest, an instruction is triggered for the path search pointer to perform path search learning and provide a search path. During the learning process, the path is modified while searching according to the principle of minimum cost, so as to obtain a search for an optimal path;

[0045] Introducing a function Indicates the number of roads selected by the 1st to Kth vehicles The maximum traffic flow rate that can be obtained when all n vehicles choose roads x j The maximum flow performance f obtained when n (x j )for

[0046] f n (x j )=J max

[0047] Among them, J max It represents the maximum smooth performance index after all vehicles select the path, m represents the total number of roads that can be selected, and n represents the total number of vehicles involved in the planning;

[0048] It is known that the number of roads selected by the K vehicles before entering is The number of roads selected by the Kth vehicle is x Kj , then the number of roads selected by the first K-1 vehicles is Consider the number of roads chosen by K vehicles as a K-step decision, then represents the optimal value of the indicator for K-step decision, g K (x Kj ) means the number of selected roads is x Kj The smooth flow probability value of the Kth step is, represents the optimal value of the indicator for the remaining K-1 decision steps. According to the optimality principle, we have

[0049]

[0050] Indicates that if the 1st to Kth vehicles choose the number of roads The optimal choice is x 1j ,…,x Kj , then x 1j ,…,x K-1,j The number of roads selected by the first K-1 vehicles must be The optimal allocation of .

[0051] Preferably, the problem is solved by a dynamic programming method Recursively calculate the optimal value of the indicator of the K-step decision; the loss amount allocated on demand is calculated by measuring the impact of congestion to obtain the smooth flow probability value g K (x Kj ) provided quantitative support.

[0052] Compared with the prior art, the invention has the following beneficial effects: 1) By using the first-order and second-order derivative methods, the discontinuous effect of the local characteristics of the image can be easily detected; in order to realize the detection of features, the spatial differential operator is used to perform convolution operation with the image, and the discrete z transform can also be used to realize the convolution operation. Specifically, the given digital matrix is ​​used as a template to perform discrete convolution operation on the image to realize the extraction of features. 2) After extracting the key turning points of the path, according to the mechanical analysis theory, the car is made to travel along a continuous curve according to a certain arc radius, so that it can travel balanced, effectively avoid obstacles, and ensure safety. Then, it is necessary to add virtual points between the two key turning points of the optimal path, that is, to make three data expansion points. By randomly transforming the original data, the diversity of the training data samples is increased, the robustness and generalization ability of the model are improved, the overfitting phenomenon is alleviated, and the problem of data imbalance is solved; 3) A traffic route planning method based on regional target selection and congestion conditions is proposed, and the congestion arrival probability and congestion loss rate function are defined; by using the congestion loss rate function, intelligent decision analysis can be realized, and traffic emergencies can be flexibly responded to, thereby improving the accuracy of navigation positioning. In addition, this method effectively solves the problems of regional road obstruction, traffic congestion, and work delays caused by poor local path planning, thereby significantly improving traffic efficiency and user experience. By globally searching for the optimal path, the traffic route planning problem is transformed into an optimization problem to be solved, thereby solving the problem of fast map aging and the need for constant updating, achieving enhanced anti-noise capabilities for navigation positioning, and effectively avoiding obstacles. In addition, a loss valuation function based on demand conditions is proposed. The use of the loss valuation function enhances the real-time performance of path planning, can effectively respond to traffic emergencies, solves the serious lag caused by poor global path planning, and solves accidents, damage, time costs or economic losses that may be caused by poor real-time performance, and is used to guide vehicles to choose the best driving path on traffic roads. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 It is the overall flow chart of the present invention.

[0055] Figure 2 This is a feature extraction diagram of the environment map planned in the simulation experiment of the present invention.

[0056] Figure 3 Schematic diagram of marking image feature points.

[0057] Figure 4 This is a schematic diagram of the planned path for complex traffic conditions according to the present invention.

[0058] Figure 5 This is an actual scene diagram of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] like Figure 1 As shown, a path search method for an unmanned vehicle, the steps are as follows:

[0061] S1, using an image feature detection algorithm to extract feature points of the image, obtaining key turning points based on the feature points of the image, and storing the key turning points in a starting table.

[0062] The specific implementation method of step S1 is as follows:

[0063] The feature extraction is performed on the overall path map (i.e., traffic map) of the environment ahead of the unmanned vehicle. The extracted operators are assumed to be two n*n matrix operators, and the lateral gradient values ​​G are obtained respectively. x and the longitudinal gradient value G y , if the total gradient value in the two directions is greater than a certain threshold δ, then the point is considered to be a feature point of the image.

[0064] The n*n matrix operator is a 3*3 matrix operator.

[0065] i. Convolution operation between matrix algorithm and target image

[0066]

[0067] Here, A and B are matrix operators, and I is the target image; is the convolution operator.

[0068] ii. Gradient value and its direction

[0069] Calculate the gradient value of the image, that is, the horizontal gradient value G of each pixel of the target image x And the longitudinal gradient value G y The total gradient value is calculated using the following formula:

[0070]

[0071] Calculate the approximate gradient value of the image, using the approximate value without square root as follows: |G| = |G x |+|G y |,This processing method actually utilizes the city block distance (Manhattan distance), making the operation more in line with natural and actual traffic scenes and more realistically reflecting the gradient changes in the image.

[0072] The gradient direction is calculated using the following formula:

[0073]

[0074] In path search, the gradient direction helps identify the direction information of each feature point in the image, which is important for identifying road edges and key turning points. The gradient direction provides directional information of pixel value changes, which helps unmanned vehicles more accurately identify road boundaries and shapes in complex road environments and enhance their path recognition and positioning capabilities.

[0075] In the subsequent path planning step, the gradient direction is used to determine the connectivity between feature points. This directional information can help the unmanned driving system identify continuous edges and form complete road boundaries, thereby optimizing path planning.

[0076] Given a threshold δ, when the total gradient value G>δ, the gradient value and direction of the point are considered as the eigenvalue of the feature point, and the eigenvalue information of the feature point is stored. The eigenvalue information includes the total gradient value G and the gradient direction θ. The feature point is an edge point, and also has multiple features such as bifurcation point, corner point, inflection point, etc.

[0077] The pixel feature values ​​in the flat area are shielded, and only the feature value information of the stored pixel points (the gradient value and direction of the feature points mentioned above) is displayed to extract the feature edge image of the target image.

[0078] In order to determine the key turning point, the present invention is based on the gray value. According to formula (1), the algorithm pointer scans the image from left to right and from top to bottom, and performs convolution operation on the image gray value. According to the convolution result, the maximum value of the convolution is selected as the feature, and the image point corresponding to this value is the key turning point of the image to be searched, and the gray value of this point is the feature value of the key turning point, and the key turning point is marked.

[0079] In order to enhance the feature points, the extracted feature points are further smoothed: let the feature edge image be f ′ (x, y), use a low-pass filter H to filter the feature edge image f in the spatial domain ′ (x, y) is smoothed to filter out noise and redundant information, making the image boundary clearer. The output image is g(u, v):

[0080] g(u,v)=∑ x ∑ y f ′ (x,y)H(u-x+1,v-y+1) (5)

[0081] After filtering by formula (5), an image with relatively complete target features can be obtained. Then, an appropriate threshold δ is selected to perform binarization processing on the output image g(u,v), highlighting the target image of interest and weakening the background area of ​​no interest, thereby reducing the amount of calculation and focusing on the target area of ​​interest. The processed image f(x,y) is:

[0082]

[0083] By analyzing the image after smoothing filtering and binarization, the key turning points can be shown on the binarized image f(x,y). These points provide important references for path optimization through the changes in grayscale values ​​and the characteristics of the angles. The starting table is used to record the initial key point information of the path, and the starting value is the grayscale value of g(0,0). The starting table can be used to quickly locate and process the path area of ​​interest, which helps to quickly build the navigation path.

[0084] Based on the above image feature extraction algorithm, feature extraction is performed on the path map planned in the experiment (this feature is the obstacle contour), such as Figure 2 The simulation results show that the proposed image feature extraction algorithm can completely outline the contour features of the target object and extract the contours of all obstacles set in path planning, which illustrates the feasibility and effectiveness of this feature extraction algorithm.

[0085] S2, perform three data expansion points between key turning points: perform three data expansions between two key turning points of the optimization path and insert virtual points to smooth the path.

[0086] The specific implementation process of step S2 is as follows:

[0087] A virtual point is added between the two key turning points of the optimization path. In order to ensure the smoothness and continuity of the path, the selection of the cubic function must satisfy the requirement of having continuous function values ​​and first-order derivative values ​​at the endpoints of the two key turning points. The specific form of the cubic function is obtained based on the interpolation method.

[0088] Suppose S(x) represents a cubic function interpolated in the horizontal direction X of the position node in the vertical direction Y, X refers to the abscissa set of the position node, Y refers to the ordinate set of the position node, y=S(x), x∈X, y∈Y.

[0089] Assume that the second-order derivative of the cubic function S(x) at each node is

[0090] S ″ (x k )=M k , k=0,1,…,λ;

[0091] Among them, M k Represents the location node x k The second-order derivative value at , λ represents the total number of position nodes in the path.

[0092] The second-order derivative of the cubic function S(x) is a linear function, that is,

[0093]

[0094] Among them, h k-1 =x k -x k-1 , M k-1 Represents the location node x k-1 The second-order derivative value at .

[0095] If the second-order derivative S ″ (x) twice, and according to the node conditions: S(x k-1 )=y k-1 and S(x k )=y k , we can get the cubic interpolation function S(x), and then by constructing cubic interpolation polynomials for the nodes in each interval and determining the boundary conditions of the second-order derivative, we can gradually calculate the virtual points between the two key turning points. The setting of virtual points can smooth the path and reduce the drastic changes in the path curvature. After adding virtual points, a smooth path that is more suitable for practical applications is generated, which can be used for real-time adjustment and optimization in the path tracking of unmanned driving systems.

[0096] S3 gives the amount of loss allocated on demand, which is used to measure the impact of traffic congestion and is used as part of the objective function in path optimization. Different amounts of loss are defined for different road sections, and the amount of loss is determined based on the probability distribution of traffic flow density and congestion intensity.

[0097] The specific process of step S3 is as follows:

[0098] According to the selection of regional targets, the occurrence of congestion is random, and the traffic flow on the road obeys the negative exponential distribution, so when the traffic flow density on the road is ξ(t), the probability of k congestion arrivals in the time period (0, t) can be defined as:

[0099]

[0100] Among them, {N(t),t≥0} is the number of congestion arrivals within (0,t].

[0101] Let the i-th road congestion intensity Y i The probability distribution function is: G(z) = P{Y i ≤x}, i=1,2,…, and assume that the congestion intensity at different intervals is independent of each other and has the same probability distribution, G *(i) (x) is represented as the i-fold convolution of the probability distribution function G(x), where x represents the upper limit of the road congestion level.

[0102] Assume that the loss function of fuel consumption, vehicle wear and tear, time, etc. caused by traffic congestion is F(t), and the interval time is T; when the congestion level on the traffic road reaches the threshold h, the processing capacity of traffic management drops to an unavailable state, then before time T, the congestion loss rate function is:

[0103]

[0104] Among them, S h It represents the traffic congestion variable when approaching the threshold h, obtained through the traffic signal stream collector.

[0105] Since traffic congestion requires a unit loss of c to restore, the total restoration cost C(T) is:

[0106]

[0107] The total cost of restoration C(T) can be used in path planning to: (1) evaluate the potential costs of different paths under traffic congestion; (2) as one of the objective functions of the optimal path search, combined with the path loss F(T) to achieve comprehensive optimization; (3) provide a cost evaluation basis for path planning to help the intelligent scheduling system formulate a reasonable traffic diversion plan. The calculation of the total cost of restoration needs to be combined with the virtual point data after the path smoothing process in step S2 to determine the congestion distribution characteristics of each road section, thereby improving the estimation accuracy.

[0108] Figure 1 The evaluation function is consistent with the description of formula (10). The evaluation function can be defined as a comprehensive indicator used to measure the quality of a path during path planning. This function needs to consider: (1) the degree of congestion on the road section; (2) the restoration cost C(T); and (3) the smooth performance of the path.

[0109] S4, according to the minimum cost principle (implemented by calculating the minimum value of the total restoration cost using formula (10)), an excitation path search algorithm is given to obtain the optimal path.

[0110] The specific process of step S4 is as follows:

[0111] If the cost calculated by formula (10) is the smallest, then an instruction is triggered for the path search pointer to perform path search learning and give a search path. During the learning process, the path is modified while searching according to the principle of minimum cost, thereby obtaining an optimal path search. The entire optimal path search process is the completion of the establishment of the optimal planning model. Introduce a function Indicates the number of roads selected by the 1st to Kth vehicles The maximum flow rate that can be obtained when f n (x j ) represents the number of roads x chosen by all n vehicles j The maximum flow performance obtained when

[0112] f n (x j )=J max (11)

[0113] Among them, J max It represents the maximum smooth performance index after all vehicles select their paths, m represents the total number of roads that can be selected, and n represents the total number of vehicles involved in the planning.

[0114] It is known that the number of roads selected by K vehicles before entering is The number of roads selected by the Kth vehicle is x Kj , then the number of roads selected by the first K-1 vehicles is Consider the number of roads chosen by K vehicles as a K-step decision, then represents the optimal value of the indicator for K-step decision, g k (x Kj ) means the number of selected roads is x Kj The smooth flow probability value of the Kth step, that is, the index value, represents the optimal value of the indicator for the remaining K-1 decision steps. According to the optimality principle, we have

[0115]

[0116] This means that if the 1st to Kth vehicles choose the road number The optimal choice is x 1j ,…,x Kj , then x 1j ,…,x K-1,j The number of roads selected by the first K-1 vehicles must be The optimal allocation of , where j = 1, 2, …, m, 1 ≤ K ≤ n.

[0117] Formula (12) recursively calculates the optimal index value of the K-step decision through the dynamic programming method. The key turning points and virtual point interpolation processing in step S2 provide more potential selection points for path planning, and the loss amount allocated on demand in step S3 measures the impact of congestion and is the smooth probability value g in formula (12) K (x Kj ) provides quantitative support. By combining the results of step S2 and step S3, formula (12) can be used to gradually optimize the path selection of each vehicle and finally achieve the optimal solution for the overall path planning.

[0118] According to the above path optimization method and formula (12), the environment model set in the experiment is as follows: Figure 3 Mark the feature points as shown, that is, mark different obstacles with different letters, and mark each new area according to the following marking rules: ① Mark all pixels as the letter I; ② Find the starting point row by row from the upper left corner and mark the point as IP; The condition for being a starting point is that the point is marked as I and is not in the outline that has been drawn; ③ According to the principle of "looking to the left" and "turning clockwise", find the same grayscale I point and add the mark according to the marking rules; ④ After completing a closed contour, find a new contour starting point. Until all contours are marked.

[0119] Here, the determination of each newly marked IP point, that is, the condition that it can become the starting point is: search row by row from left to right and from top to bottom. ① Construct a record table area, and clear the record table area when searching for a new row; ② Enter the record table when encountering the letter A, exit the record table when encountering D, and ignore when encountering R; if the empty record table encounters I, it is a new IP point; if the non-empty record table encounters I and the grayscale value of the pixel is different from the grayscale value of the previous point A in the record table, then the point is a new IP point, otherwise it is not. Here, A is the direction of the starting or upward or left pixel out; R is the direction of the pixel out to the right; D is the direction of the pixel out to the bottom.

[0120] The traditional path search method selects the marked points on the existing map and searches according to the marked points. The present invention first gives an arbitrary starting point for search, builds the surrounding environment while searching, changes the path direction at any time according to the evaluation function, selects the next branch point, and calculates the probability of obstacles appearing on this branch. Obstacles are avoided based on this probability value. The entire search process is carried out while calculating and building the surrounding environment map, thus forming an optimal search path. Figure 4 In the example, the map area increases, the complexity increases, there are more concave obstacles in the upper half of the map, and the obstacles in the lower half are randomly distributed. The present invention can pass through concave obstacles and randomly distributed obstacles with fewer turning points, and the quality of the generated path is significantly improved compared with the traditional algorithm.

[0121] In order to verify the effectiveness of the optimal path search algorithm proposed in this invention, this algorithm is deployed on the digital Meipai of the mobile robot platform developed by our laboratory. Figure 5 The actual object shown. A relatively complex test environment is designed for the path search method designed in the present invention. A path is generated using the unmanned driving path search method proposed in the present invention. The generated path first moves forward for a distance, then turns right to avoid the obstacle on the left. Although the path generated by the method proposed in the present invention is close to the obstacle, due to the existence of the expansion layer, the robot can still safely avoid the obstacle. At this time, the mobile robot is surrounded by obstacles and is prone to collision. Therefore, it needs to move at a lower speed. After completing the right turn, it moves to the left. At this time, there are obstacles on both sides of the robot. Combined with the environmental model and the expansion layer mark, a dynamic programming algorithm based on the principle of minimum cost is used to gradually calculate the optimal path from the starting point to the end point. It is far away from the obstacles on both sides, which improves the safety of the robot's passage. The subsequent path gradually moves away from the obstacles and finally reaches the end position.

[0122] In summary, in an actual environment where obstacles appear randomly in size, direction and distance from each other, the path generated by the path search method proposed in the present invention can plan a path as far away from obstacles as possible on the basis of shortening the path length as much as possible (the path formed while searching and evaluating, calculated by the evaluation function of each step of the search) (the evaluation formula (10) and the optimal decision formula (12) ensure that the decision to search this path with low cost and good performance is the path far away from obstacles), thereby improving the travel efficiency and safety of the mobile robot.

[0123] In order to verify that the established global optimal path planning model is stable and robust, a stability experiment is given here (obtained from the experiment of the constructed network structure model for searching the optimal machine learning, where a second-order five-layer neural network structure is established to simulate the path planning method of this patent). The results show that: the error gradually decreases with the increase in the number of experiments, tends to be stable and convergent, indicating that the established global optimal path planning model (including key turning point selection, interpolation between turning points, evaluation function, and optimal decision) has stable performance and strong noise resistance.

[0124] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A path search method for an unmanned vehicle, characterized in that: The steps are as follows: S1, extracting feature points of the image using an image feature detection algorithm, obtaining key turning points based on the feature points of the image, and storing the key turning points in a starting table; S2, three data expansions are performed between two key turning points of the optimization path, and virtual points are inserted to smooth the optimization path; S3, gives the amount of loss allocated on demand and is used as part of the objective function in path optimization; S4, minimize the objective function according to the minimum cost principle, give an excitation path search algorithm, and obtain the optimal path.

2. The path search method for an unmanned vehicle according to claim 1, characterized in that: The method of extracting feature points of an image using an image feature detection algorithm is as follows: using a matrix operator to extract features from a traffic map in front of an unmanned vehicle, and obtaining lateral gradient values ​​G x and the longitudinal gradient value G y , if the total gradient value in both directions is greater than the threshold δ, it is a feature point of the image.

3. The path search method for an unmanned vehicle according to claim 2, characterized in that: The lateral gradient value G is obtained by convolving the matrix operator with the traffic map. x and the longitudinal gradient value G y ; The total gradient value G in the two directions is G = |G x |+|G y |; The matrix operator is a 3*3 matrix operator; Among them, A and B are matrix operators, and I is the traffic map; is the convolution operator.

4. The path search method for an unmanned vehicle according to any one of claims 1 to 3, characterized in that: The method for obtaining the key turning point according to the feature points of the image is as follows: taking the total gradient value and gradient direction corresponding to the feature points of the image as feature information to construct a feature edge image; The feature edge image is filtered using a low-pass filter, and the filtered image is binarized. The points with pixel values ​​of 1 in the binarized image are key turning points.

5. The path search method for an unmanned vehicle according to claim 4, characterized in that: The method of filtering the feature edge image using a low-pass filter is: A low-pass filter H is used to filter the feature edge image f in the spatial domain. ′ (x, y) is smoothed to remove noise and redundant information, and the output image g(u, v) is: g(u,v)=∑ x ∑ y f ′ (x,y)H(u-x+1,v-y+1); Among them, u, v represent the pixel coordinates in the output image g(u, v), and x, y represent the feature edge image f ′ Pixel coordinates in (x,y); The method for binarizing the filtered image is: Select the threshold δ to binarize the output image g(u,v), and the processed image f(x,y) is: The key turning points appear on the binarized image f(x,y); The starting table is used to record the initial key point information of the path, and the starting value is the gray value of the output image g(0,0); the path area of ​​interest is quickly located and processed through the starting table.

6. The path search method for an unmanned vehicle according to claim 4, characterized in that: The method for performing cubic data expansion between two key turning points of the optimization path is: adding virtual points between the two key turning points by selecting a cubic function with continuous function values ​​and first-order derivative values ​​at the endpoints of the two key turning points, and the cubic function is obtained based on the interpolation method.

7. The path search method for an unmanned vehicle according to claim 6, characterized in that: The method for solving the cubic function is: S(x) represents the cubic function interpolated in the vertical direction Y in the horizontal direction X of the position node, X refers to the abscissa set of the position node, Y refers to the ordinate set of the position node, y=S(x), x∈X, y∈Y; The second-order derivative of the cubic function S(x) at each node is S ″ (x k )=M k , k=0,1,…,λ; where M k Represents the location node x k The second-order derivative value at , λ represents the total number of position nodes in the path; The second-order derivative of the cubic function S(x) is a linear function, that is, Among them, h k-1 =x k -x k-1 , M k-1 Represents the location node x k-1 The second derivative value at ; For the second-order derivative S ″ (x) twice, and according to the node conditions: S(x k-1 )=y k-1 and S(x k )=y k , we can get the cubic interpolation function S(x).

8. The path search method for an unmanned vehicle according to claim 6 or 7, characterized in that: Different loss amounts are defined for different road sections, and the loss amounts are determined based on the probability distribution of traffic flow density and congestion intensity; The method for determining the loss amount is: The occurrence of congestion is random, and the traffic flow on the road follows a negative exponential distribution. When the traffic flow density on the road is ξ(t), the probability of k congestion arrivals in the time period (0, t) is: Among them, {N(t),t≥0} is the number of congestion arrivals within (0,t]; The i-th road congestion intensity Y i The probability distribution function is: G(x) = P{Y i ≤x}, i=1,2,…; Assume that the congestion intensity at different intervals is independent and has the same probability distribution, G *(i) (x) is represented as the i-fold convolution of the probability distribution function G(x), where x represents the upper limit of the road congestion level; Assume that the function of the loss of fuel consumption, vehicle wear and time caused by traffic congestion is F(t), and the interval time is T; when the congestion level on the traffic road reaches the threshold h, the processing capacity of traffic management drops to an unavailable state, then before time T, the congestion loss rate function is: Among them, S h represents the traffic congestion variable when approaching the threshold h; Since traffic congestion requires a unit loss of c to restore, the total restoration cost C(T) is:

9. The path search method for an unmanned vehicle according to claim 8, characterized in that: The implementation method of step S4 is as follows: if the cost calculated by the total recovery cost C(T) is the smallest, an instruction is triggered for the path search pointer to perform path search learning and provide a search path. During the learning process, the path is modified while searching according to the principle of minimum cost, thereby obtaining a search for an optimal path; Introducing a function Indicates the number of roads selected by the 1st to Kth vehicles The maximum traffic flow rate that can be obtained when all n vehicles choose roads x j The maximum flow performance f obtained when n (x j ) is f n (x j )=J max Among them, J max It represents the maximum smooth performance index after all vehicles select the path, m represents the total number of roads that can be selected, and n represents the total number of vehicles involved in the planning; It is known that the number of roads selected by the K vehicles before entering is The number of roads selected by the Kth vehicle is x Kj , then the number of roads selected by the first K-1 vehicles is Consider the number of roads chosen by K vehicles as a K-step decision, then represents the optimal value of the indicator for K-step decision, g K (x Kj ) means the number of selected roads is x Kj The smooth flow probability value of the Kth step is, represents the optimal value of the indicator for the remaining K-1 decision steps. According to the optimality principle, we have Indicates that if the 1st to Kth vehicles choose the number of roads The optimal choice is x 1j ,…,x Kj , then x 1j ,…,x K-1,j The number of roads selected by the first K-1 vehicles must be The optimal allocation of .

10. The path search method for an unmanned vehicle according to claim 9, characterized in that: Solving the problem using dynamic programming Recursively calculate the optimal value of the indicator of the K-step decision; the loss amount allocated on demand is calculated by measuring the impact of congestion to obtain the smooth flow probability value g K (x Kj ) provided quantitative support.

Images