Intelligent vision-guided vehicle obstacle avoidance control optimization method

Abstract

The invention discloses an intelligent vision-guided automobile obstacle avoidance control optimization method, which belongs to the field of autonomous driving and is characterized in that it comprises a target detection module and an obstacle avoidance control module; wherein the target detection module comprises GCT-YOLOv5s, extracts key features through the C3GC module, utilizes the CARAFE operator to enhance feature fusion, and determines the relationship between classification and positioning through the TSCODE detection head, so as to realize the detection of traffic participant types in driving scenes; the obstacle avoidance control module comprises adaptive obstacle avoidance path planning and control, dynamically adjusts the repulsive field in combination with the traffic participant types, performs path smoothing processing, and adopts a pure tracking algorithm to perform path tracking control, so as to realize adaptive obstacle avoidance path planning and control of intelligent vehicles.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous driving, and more specifically, is an intelligent vision-guided automobile obstacle avoidance control optimization method. Technical Background

[0002] The complexity of urban roads, the dynamics of traffic flow, and the randomness of traffic participants' behavior have brought great challenges to the research on intelligent vehicle target detection and obstacle avoidance control. When performing obstacle avoidance control, intelligent vehicles need to accurately identify obstacle information, make reasonable decision-making plans, and accurately realize motion control. Traditional navigation and obstacle avoidance methods mainly rely on optical flow algorithms and binocular parallax calculations. Although they can calculate the displacement and depth information of obstacles, they have many disadvantages: high computational complexity, resulting in system delays; poor environmental adaptability, especially poor performance under conditions of changing lighting and sparse textures; low detection accuracy, making it difficult to identify and classify small targets and occluded objects; lack of target detection and feature extraction technology, which increases the difficulty of parameter tuning and affects the robustness and adaptability of the system.

[0003] Early target detection relied on manually designed features and classifiers, which were difficult to adapt to the complex environment in reality. With the rapid development of computer hardware and related fields such as deep learning technology, target detection algorithms continue to evolve. The emergence of two-stage detection methods such as R-CNN and single-stage detection methods such as YOLO has provided a wealth of options for achieving more accurate and efficient target detection tasks. However, the driving environment of autonomous driving is more complex and changeable, and there are still problems with the detection of small targets and occluded targets.

[0004] In terms of autonomous driving path planning, mainstream methods include artificial potential field method, graph search method, fuzzy logic control, fast random search tree and neural network method. Among them, the artificial potential field method is simple and feasible, but when applied to the path planning of smart cars, there are problems such as local minimum and unreachable targets, as well as long and uneven planned paths. At the same time, for vulnerable groups such as pedestrians, smart cars do not consider the safety differences of different types of obstacles during the path planning process. Path tracking control, its goal is to minimize the deviation between the vehicle and the reference path while completing the obstacle avoidance task. The pure tracking algorithm is simple to calculate and can achieve accurate tracking of the lateral path by controlling the front wheel angle. It is very popular in the field of smart car path tracking. Summary of the invention

[0005] To solve the above problems, the present invention proposes an intelligent vision-guided automobile obstacle avoidance control optimization method. The target detection module uses the machine vision technology GCT-YOLOv5s, that is, the YOLOv5s model that combines the fusion structure of the centralized comprehensive convolution and the global context module, the content-aware reorganization operator and the task-specific context decoupling head, combined with the C3GC module, that is, the fusion structure module of the centralized comprehensive convolution and the global context module, the CARAFE operator, that is, the content-aware reorganization operator and the TSCODE detection head, that is, the task-specific context decoupling detection head, to extract richer and more critical feature information, which is suitable for the detection of small targets and occluded targets in complex environments. The obstacle avoidance control module dynamically adjusts the repulsive field in combination with the type of traffic participants, performs obstacle avoidance path planning, and smoothes the path at the same time. The pure tracking algorithm is used for path tracking control to realize the adaptive obstacle avoidance path planning and control of the intelligent vehicle, which has higher integration and adaptability.

[0006] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0007] The intelligent vision-guided vehicle obstacle avoidance control optimization method includes a target detection module and an obstacle avoidance control module. The method includes the following steps:

[0008] Step 1, target detection module: extract key features through the C3GC module, use the CARAFE operator to enhance feature fusion, determine the relationship between classification and positioning through the TSCODE detection head, improve YOLOv5 to build GCT-YOLOv5s for small targets and occluded targets in complex environments, and realize the detection of traffic participant types in driving scenes. The process includes the following sub-steps:

[0009] Step 1.1, C3GC module:

[0010] The Global Context module is integrated into the C3, i.e., centralized comprehensive convolutional structure, of the YOLOv5 backbone network. The Global Context module is added before the CBS, i.e., channel block compression layer in the C3 structure to form the C3GC module, which is used to highlight the useful information of the feature map and suppress irrelevant information, so that the C3GC module can focus on different areas to effectively detect objects.

[0011] The attention mechanism is a computational model that simulates the selectivity of human vision and cognitive processes. In the GlobalContext module, a set of key features in sequence transformation establish a global dependency connection between the input and output sequences for mapping through the self-attention mechanism. When the attention mechanism is used for image recognition, the information processed by the module is contained in the explicit sequence of items, and the formula is as follows:

[0012] α=(α1 ,α 2 ,....α i ,...,α n ),α i ∈R D (1)

[0013] In the formula, α represents the key feature of the feature map, i represents the number of feature vectors, n represents the number of feature vectors, and R D Represents the dimensional space where the feature vector is located.

[0014] The global attention pool uses a 1×1 convolution W k And Softmax function to generate attention weights and obtain global context features, assign different weights to input features from the spatial and channel dimensions to highlight useful information. The expression of attention weight is as follows:

[0015]

[0016] δ(·)=α+W v2 (Re LU(LN(W v1 (F(·))))),W v1 ∈R rC×H×W ,W v2 ∈R C×H×W (3)

[0017] Where F(·) is the attention weight generated by the Softmax function, δ(·) represents the bottleneck transformation, and N P represents the dimension of the input feature map, W k×j is a 1×1 convolution W of dimension j k , W k×m is a 1×1 convolution W of dimension m k , W v1 and W v2 There are two 1×1 convolutional layers, Re LU is the rectified linear unit, LN is the natural logarithm, and the size of its input feature map is H×W in the feature space, H represents the height of the feature map, W represents the width of the feature map, and the number of channels is rC and C respectively, C represents a channel of the input feature map, and r represents the dimension reduction factor.

[0018] Step 1.2, CARAFE operator:

[0019] In the feature fusion part, the upsampling CARAFE operator is introduced to replace the original upsampling structure of YOLOv5s. The CARAFE operator has a broad field of view and can effectively integrate the surrounding context information. It can dynamically generate adaptive kernels and can perceive the processing method of instance content, further enriching the semantic information of feature maps to enhance feature fusion.

[0020] Step 1.3, TSCODE detection head:

[0021] The TSCODE detection head is used to replace the traditional coupling head to balance the relationship between classification and localization and handle complex context conflicts. The classification task of the TSCODE detection head generates a feature map that is spatially coarse but semantically richer, while the localization task provides a feature map that contains more detailed texture and boundary information.

[0022] Classification Task The semantic context encoding part uses the l and P l+1 The feature maps of the two layers generate feature maps with richer semantics. Specifically, P l After 2 times of downsampling, l+1 Connect to generate the final This operation can not only utilize P l The sparsity of the salient features of the layer also benefits from P l+1 The layer-rich contextual semantics is calculated as follows:

[0023]

[0024] In the formula, Concat and DConv represent the connection layer and the shared downsampling convolution layer respectively, P l Represents the lth feature pyramid layer.

[0025] Positioning Task The encoding part for local detail preservation accepts feature maps from three pyramid levels, namely P l-1 , P l and P l+1 , P l-1 Provides more details and edge features, while P l+1 Provides a more comprehensive view of objects. First, P l Upsample twice and then with P l-1 Aggregate to get the feature map H l-1 , H l-1 After two downsampling, with P l and P after two upsampling l+1 Aggregation, finally generated This operation can obtain texture details and boundary information to accurately predict the location of the target in the image. The formula is expressed as:

[0026]

[0027] Where μ represents feature upsampling and DConv represents another shared downsampling convolutional layer.

[0028] Step 2, obstacle avoidance control module: Based on step 1, the repulsion field is dynamically adjusted in combination with the type of traffic participants, obstacle avoidance path planning is performed, the path is smoothed, and a pure tracking algorithm is used for path tracking control to achieve adaptive obstacle avoidance path planning and control for smart cars. The process includes the following sub-steps:

[0029] Step 2.1, Dynamic adjustment of repulsive field:

[0030] Step 2.1.1, introduce repulsion attenuation factor:

[0031] A repulsive force attenuation factor is introduced into the repulsive force function. When the distance between the target point and the obstacle is close, the repulsive force generated by the obstacle gradually decreases. The improved repulsive force potential field function expression is as follows:

[0032]

[0033] Where m is the position of the smart car, k r is the repulsion gain coefficient, m obs is the distance between the smart car and the obstacle, n is a fixed value, is the repulsive force attenuation factor, m g is the distance between the smart car and the target point, (x goal ,y goal ) is the location information of the target point, (x obs ,y obs ) is the location information of the obstacle.

[0034] General After substituting into formula (6), the partial derivatives of x and y are calculated to obtain:

[0035]

[0036] It can be seen from formula (7) that when the distance between the target point and the obstacle is close to the vehicle, the repulsive force generated by the obstacle will decrease rapidly. The present invention sets the trigger range of the function to s. If the distance between the vehicle and the target point is greater than the trigger range, the original repulsive force function will be used to work; otherwise, the improved repulsive force potential field function will replace the original repulsive force function. In the process of formula derivation, m g is a constant, s is m g The range of change is as follows:

[0037]

[0038] Step 2.1.2: Introduce target type adaptation factor:

[0039] Introducing the target type adaptation factor ξ in formula (8), we get formula (9) to dynamically adjust the repulsion field. For different types of traffic participants detected by the target detection module, ξ has different values ​​to adaptively change the repulsion field and perform adaptive obstacle avoidance path planning. For vulnerable groups such as pedestrians, ξ takes a larger value, and for ordinary vehicles, ξ takes a normal value, but it should not be too large or too small. If it is too large, the path planning will be to the road boundary, and if it is too small, it will collide with obstacles. After introducing the target type adaptation factor, the repulsion function formula is as follows:

[0040]

[0041] Step 2.2: Path smoothing:

[0042] Step 2.2.1, temporary target point selection:

[0043] Construct a prediction artificial potential field method, which includes two steps: predicting the future path and selecting temporary target points;

[0044] Step 2.2.1.1, predict future path:

[0045] When predicting forward a fixed distance, the improved algorithm of step 2.1 is used, which uses only two intuitive parameters: the prediction horizon λ and the step size Δ. The parameter λ determines the length of the forward prediction path, and the second parameter Δ is used to manipulate the distance between consecutive points. The next position of the waypoint is generated using the improved algorithm of step 2.1 with a constant step size Δ until the predefined prediction horizon λ is reached.

[0046] Step 2.2.1.2, select a new temporary target point:

[0047] After determining the starting point and the target point, use the improved algorithm in step 2.1 to calculate the resultant force of the starting point or temporary target point, and determine the equation of the line where the resultant force is located. The point where the resultant force equation is calculated for the first time is the starting point of the vehicle. When calculating the resultant force equation in the future, the selected points are all obtained from the previous prediction. Through the vertical projection of each point on the path and the straight line equation, the mathematical formula of the vertical distance is used to obtain the point farthest from the straight line equation as the new temporary target point, until the distance between the vehicle and the global target point at the latest temporary target point is less than λ, and the global target point is used as the temporary target point.

[0048] Step 2.2.2, polynomial curve fitting:

[0049] On the basis of the artificial potential field prediction method PAPF, the polynomial curve fitting method is added to design a smooth artificial potential field prediction method. The polynomial curve fitting smoothes some discrete path points in the form of polynomials to solve the problem of sudden curvature changes in the discrete point set path obtained by the artificial potential field method. Let t be the fitting order of the polynomial, then the calculation formula of the polynomial curve fitting is:

[0050]

[0051] Where (x, y) represents the coordinates of discrete path points, and a represents the coefficient constant.

[0052] The number of polynomial curve fittings is usually determined by factors such as data distribution, fitting requirements, and data volume. High approximation accuracy can be achieved with less noise. When using high-order polynomials to fit curves, overfitting is prone to occur, especially in the presence of noise and outliers, where the Runge phenomenon may be observed. Therefore, when choosing the number of fittings, it is necessary to weigh the fitting effect and the risk of overfitting to ensure that the fitting results are both accurate and generalized.

[0053] Step 2.3, pure tracking algorithm:

[0054] The pure tracking algorithm is a geometric calculation model based on human driving habits. It tracks the desired path by controlling the front wheel angle of the driverless vehicle. Based on the vehicle kinematic model, the pure tracking algorithm is used to control the intelligent car to track the obstacle avoidance path adaptively planned by the smooth prediction artificial potential field method. According to the pure tracking model, the sine theorem can be obtained:

[0055]

[0056]

[0057]

[0058] Where, L d is the preview distance of the pure tracking model, R is the turning radius of the vehicle, and α is the angle between the vehicle axis and the line connecting the center of the vehicle's rear axle to the preview path point.

[0059] From formula (13), we can get the curvature From the pure pursuit model

[0060] The desired path is tracked by controlling the front wheel steering angle of the smart car. The specific formula is as follows:

[0061]

[0062] Where L is the wheelbase, δ is the front wheel turning angle, and e is the lateral error.

[0063] The beneficial effects of the present invention are as follows: the present invention proposes an intelligent vision-guided automobile obstacle avoidance control optimization method, which can solve the detection of multi-scale targets, small targets and occluded targets in complex road environments. The obstacle avoidance path planning and control strategy is formulated after the obstacle type is identified, so that the intelligent car exhibits adaptive smooth obstacle avoidance path planning and control characteristics. This design aims to enable the intelligent car to accurately identify the obstacle type, and further perform adaptive obstacle avoidance path planning and control based on the obstacle type, so as to ensure the safety and intelligence of the intelligent car when facing a variety of driving scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is the overall technical solution diagram of the present invention.

[0065] Figure 2 This is the GCT-YOLOv5s network structure diagram.

[0066] Figure 3 It is the flow chart of obstacle avoidance control module. DETAILED DESCRIPTION

[0067] The present invention is described in detail below with reference to the accompanying drawings and examples.

[0068] Figure 1 It is an overall technical scheme diagram of the intelligent vision-guided automobile obstacle avoidance control optimization method. The present invention includes a target detection module 1 and an obstacle avoidance control module 2. The target detection module 1 is composed of a C3GC module, i.e., a fusion structure module of a centralized comprehensive convolution and a global context module, a CARAFE operator, i.e., a content-aware reorganization operator, and a TSCODE detection head, i.e., a task-specific context decoupling detection head. The three sub-modules finally construct a GCT-YOLOv5s for small targets and occluded targets in a complex environment, i.e., a YOLOv5s that combines a fusion structure of a centralized comprehensive convolution and a global context module, a content-aware reorganization operator, and a task-specific context decoupling head, to realize the detection of the type of traffic participants in the driving scene. The type of traffic participants detected by the target detection module is sent to the obstacle avoidance control module 2, which is dynamically adjusted by the repulsive field, path smoothing processing, and a pure tracking algorithm. Three sub-modules realize the adaptive obstacle avoidance path planning and control of the intelligent car.

[0069] The method comprises the following steps:

[0070] Step 1, target detection module 1: The image information is sequentially extracted through the C3GC module to extract key features, the CARAFE operator enhances feature fusion, and the TSCODE detection head determines the relationship between classification and positioning. The GCT-YOLOv5s of small targets and occluded targets in complex environments is constructed by improving YOLOv5 to realize the detection of traffic participant types in driving scenes. Figure 2It can be seen that the process includes the following sub-steps:

[0071] Step 1.1, the image information is input to the backbone network at the input end;

[0072] Step 1.2: Extract features from the image information input in step 1.1 through the backbone network, where the C3GC module:

[0073] The Global Context module is integrated into the C3, i.e., centralized comprehensive convolutional structure, of the YOLOv5 backbone network. The Global Context module is added before the CBS, i.e., channel block compression layer in the C3 structure to form the C3GC module, which is used to highlight the useful information of the feature map and suppress irrelevant information, so that the C3GC module can focus on different areas to effectively detect objects.

[0074] The attention mechanism is a computational model that simulates the selectivity of human visual and cognitive processes. In the GlobalContext module, a set of key features in sequence transformation establishes global dependency connections between input and output sequences for mapping through the self-attention mechanism. When the attention mechanism is used for image recognition, the information processed by the module is contained in the explicit sequence of items, and the formula is as follows:

[0075] α=(α 1 ,α 2 ,....α i ,...,α n ),α i ∈R D (1)

[0076] In the formula, α represents the key feature of the feature map, i represents the number of feature vectors, n represents the number of feature vectors, and R D Represents the dimensional space where the feature vector is located.

[0077] The global attention pool uses a 1×1 convolution W k And Softmax function to generate attention weights and obtain global context features, assigning different weights to input features from the spatial and channel dimensions to highlight useful information; the expression of attention weights is as follows:

[0078]

[0079] δ(·)=α+W v2 (Re LU(LN(W v1 (F(·))))),W v1 ∈R rC×H×W ,W v2 ∈R C×H×W (3)

[0080] Where F(·) is the attention weight generated by the Softmax function, δ(·) represents the bottleneck transformation, and N P represents the dimension of the input feature map, W k×j is a 1×1 convolution W of dimension j k , W k×m is a 1×1 convolution W of dimension m k , W v1 and W v2 There are two 1×1 convolutional layers, Re LU is the rectified linear unit, LN is the natural logarithm, and the size of its input feature map is H×W in the feature space, H represents the height of the feature map, W represents the width of the feature map, and the number of channels is rC and C respectively, C represents a channel of the input feature map, and r represents the dimension reduction factor.

[0081] Step 1.3: Use the neck to aggregate and refine the features extracted in step 1.2, usually focusing on strengthening the spatial and semantic information at different scales and further refining the features. The CARAFE operator:

[0082] In the feature fusion part, the upsampling CARAFE operator is introduced to replace the original upsampling structure of YOLOv5s. The CARAFE operator has a broad field of view and can effectively integrate the surrounding context information. It can dynamically generate adaptive kernels and can perceive the processing method of instance content, further enriching the semantic information of feature maps to enhance feature fusion.

[0083] Step 1.4: Finally, the detection end processes the features provided in step 1.3, generates predictions for each candidate, completes target detection, and realizes the detection of traffic participant types in the driving scene. The TSCODE detection head:

[0084] The feature map enhanced by steps 1.2 and 1.3 is sent to the detection end for detection and output of target prediction box, target category and confidence score to indicate different types of traffic participants. The traditional coupling head is replaced by the TSCODE detection head to balance the relationship between classification and positioning and handle complex context conflicts. The classification task of the TSCODE detection head generates a feature map that is spatially rough but semantically richer, while the positioning task provides a feature map containing more detailed texture and boundary information.

[0085] Classification Task The semantic context encoding part uses the l and P l+1 The feature maps of the two layers generate feature maps with richer semantics. Specifically, P l After 2 times of downsampling, l+1 Connect to generate the final This operation can not only utilize Pl The sparsity of the salient features of the layer also benefits from P l+1 The layer-rich contextual semantics is calculated as follows:

[0086]

[0087] In the formula, Concat and DConv represent the connection layer and the shared downsampling convolution layer respectively, P l Represents the lth feature pyramid layer.

[0088] Positioning Task The encoding part for local detail preservation accepts feature maps from three pyramid levels, namely P l-1 , P l and P l+1 , P l-1 Provides more details and edge features, while P l+1 Provides a more comprehensive view of objects. First, P l Upsample twice and then with P l-1 Aggregate to get the feature map H l-1 , H l-1 After two downsampling, with P l and P after two upsampling l+1 Aggregation, finally generated This operation can obtain more texture details and boundary information to accurately predict the location of the target in the image. The formula is expressed as:

[0089]

[0090] Where μ represents feature upsampling and DConv represents another shared downsampling convolutional layer.

[0091] Step 1.5, hyperparameter settings: number of iterations: 300; warm-up learning rounds: 3; warm-up learning rate momentum: 0.8; batch size: 16; optimizer: SGD; weight decay coefficient: 0.0005.

[0092] Step 1.6, weight file training: After setting the hyperparameters in step 1.5, start training to generate the weight file and obtain the final obstacle information perception module.

[0093] Step 2, obstacle avoidance control module 2: Based on step 1, the repulsive field is dynamically adjusted in combination with the type of traffic participants, obstacle avoidance path planning is performed, and the path is smoothed. A pure tracking algorithm is used for path tracking control to achieve adaptive obstacle avoidance path planning and control for smart cars. Figure 3 It can be seen that the process includes the following sub-steps:

[0094] Step 2.1, Dynamic adjustment of repulsive field:

[0095] When the smart car starts obstacle avoidance path planning, it first initializes the parameters to calculate the potential field attraction and repulsion, and determines whether the smart car is trapped in the local minimum. If the smart car is not trapped in the local minimum, it continues to predict forward. If the smart car is trapped in the local minimum, the repulsive field is dynamically adjusted to break the force balance. The local minimum refers to the phenomenon that under the action of the resultant force, the smart car repeatedly oscillates in a small area, causing the smart car to be unable to move forward or delay its movement.

[0096] Step 2.1.1, introduce repulsion attenuation factor:

[0097] A repulsive force attenuation factor is introduced into the repulsive force function. When the distance between the target point and the obstacle is close, the repulsive force generated by the obstacle gradually decreases. The improved repulsive force potential field function expression is as follows:

[0098]

[0099] Where m is the position of the smart car, k r is the repulsion gain coefficient, m obs is the distance between the smart car and the obstacle, n is a fixed value, is the repulsive force attenuation factor, m g is the distance between the smart car and the target point; (x goal ,y goal ) is the location information of the target point, (x obs ,y obs ) is the location information of the obstacle.

[0100] General After substituting into formula (6), the partial derivatives of x and y are calculated to obtain:

[0101]

[0102] It can be seen from formula (7) that when the distance between the target point and the obstacle is close to the vehicle, the repulsive force generated by the obstacle will decrease rapidly. The present invention sets the trigger range of the function to s. If the distance between the vehicle and the target point is greater than the trigger range, the original repulsive force function will be used to work; otherwise, the improved repulsive force potential field function will replace the original repulsive force function. In the process of formula derivation, m g is a constant, s is m g The range of change is as follows:

[0103]

[0104] Step 2.1.2: Introduce target type adaptation factor:

[0105] Introducing the target type adaptation factor ξ into formula (8), we get formula (9) to dynamically adjust the repulsive field. The obstacle information obtained by the target detection module 1 is used to detect different traffic participants by the perception module to obtain the types of different traffic participants. ξ has different values ​​to adaptively change the repulsive field and perform adaptive obstacle avoidance path planning (at this moment, it is assumed that the distance information has been obtained based on other sensor devices of the intelligent vehicle). For vulnerable groups such as pedestrians, ξ takes a larger value, and for ordinary vehicles, ξ takes a normal value, but it should not be too large or too small. If it is too large, the path planning will be to the road boundary, and if it is too small, it will collide with obstacles. The repulsive function formula after introducing the target type adaptation factor is as follows:

[0106]

[0107] Step 2.2: Path smoothing:

[0108] Step 2.2.1, temporary target point selection:

[0109] After the repulsive field is dynamically adjusted in step 2.1, a prediction artificial potential field method is constructed, which includes two steps: predicting the future path and selecting a temporary target point;

[0110] Step 2.2.1.1, predict future path:

[0111] When predicting forward a fixed distance, the improved algorithm of step 2.1 is used, which uses only two intuitive parameters: the prediction range λ and the step size Δ; the parameter λ determines the length of the forward prediction path, and the second parameter Δ is used to manipulate the distance between consecutive points. The next position of the path point is generated using the improved algorithm of step 2.1 with a constant step size Δ until the predefined prediction range λ is reached. The specific process is as follows:

[0112] (1) First, calculate the resultant force of the starting point or temporary target point, determine the equation of the straight line where the resultant force is located, and define it as the force direction of the current original artificial potential field method;

[0113] (2) Use the improved artificial potential field method in step 2.1 to predict a fixed distance forward with a constant step size Δ to generate a collision-free continuous path;

[0114] (3) To calculate the vertical distance between each path point and the straight line equation established in step (1), use the mathematical formula for vertical distance to obtain the point farthest from the straight line equation by perpendicularly projecting each point on the path onto the straight line equation;

[0115] (4) Select the point with the farthest distance as the new temporary target point;

[0116] (5) Use the improved artificial potential field method in step 2.1 to re-plan a path to the temporary target point to generate a shorter distance.

[0117] Step 2.2.1.2, select a new temporary target point:

[0118] After determining the starting point and the target point, use the improved algorithm in step 2.1 to calculate the resultant force of the starting point or temporary target point, and determine the equation of the line where the resultant force is located. The point where the resultant force equation is calculated for the first time is the starting point of the vehicle. When calculating the resultant force equation in the future, the selected points are all obtained by the previous prediction. Through the vertical projection of each point on the path and the straight line equation, the mathematical formula of the vertical distance is used to obtain the point farthest from the straight line equation as the new temporary target point. Until the distance between the vehicle and the global target point at the latest temporary target point is less than λ, the global target point is used as the temporary target point. The specific process is as follows:

[0119] (1) Calculate the resultant force F at the starting point or temporary target point;

[0120] (2) Create a linear function (i.e., Ax+By+C=0, where x and y are the coordinates of the path point; A, B, and C are the function parameters for obtaining the direction of the force), called the direction line;

[0121] (3) Calculate the distance d between each predicted path point and the direction line in the predicted path using the following method:

[0122]

[0123] In the formula, the denominator are A and B in the equation of the line, (x 0 ,y 0 ) is the location of the predicted waypoint;

[0124] (4) Select the point farthest from the straight line equation as the new temporary target point;

[0125] (5) Use the improved artificial potential field method in step 2.1 to predict a fixed distance forward with a constant step size Δ. When the distance between the vehicle's latest temporary target point and the global target point is less than λ, use the global target point as the temporary target point.

[0126] Step 2.2.2, polynomial curve fitting:

[0127] In step 2.2.1, when the global target point is used as a temporary target point, the path is smoothed, that is, a polynomial curve fitting method is added to the predictive artificial potential field (PAPF) method to design a smooth predictive artificial potential field method. Polynomial curve fitting fits and smoothes some discrete path points in the form of polynomials to solve problems such as sudden changes in curvature of the discrete point set path planned by the artificial potential field method. Let t be the fitting order of the polynomial, then the calculation formula of the polynomial curve fitting is:

[0128]

[0129] Where (x, y) represents the coordinates of discrete path points, and a represents the coefficient constant.

[0130] The number of polynomial curve fittings is usually determined by factors such as data distribution, fitting requirements, and data volume. High approximation accuracy can be achieved under low noise influence. When using high-order polynomials to fit curves, overfitting problems are prone to occur, especially in the presence of noise and outliers, where the Runge phenomenon may be observed. Therefore, when choosing the number of fittings, it is necessary to weigh the fitting effect and the risk of overfitting to ensure that the fitting results are both accurate and generalized.

[0131] Step 2.3, pure tracking algorithm:

[0132] The pure tracking algorithm is a geometric calculation model based on human driving habits. It tracks the desired path by controlling the front wheel angle of the driverless vehicle. Based on the vehicle kinematic model, the pure tracking algorithm is used to control the intelligent vehicle to track the obstacle avoidance path adaptively planned in steps 2.1 and 2.2. According to the pure tracking model, the sine theorem can be used to obtain:

[0133]

[0134]

[0135]

[0136] Where, L d is the preview distance of the pure tracking model, R is the turning radius of the vehicle, and α is the angle between the vehicle axis and the line connecting the center of the vehicle's rear axle to the preview path point.

[0137] From formula (14), we can get the curvature From the pure pursuit model

[0138] The desired path is tracked by controlling the front wheel steering angle of the smart car. The specific formula is as follows:

[0139]

[0140] Where L is the wheelbase, δ is the front wheel turning angle, and e is the lateral error.

Claims

1. Intelligent vision-guided vehicle obstacle avoidance control optimization method, characterized in that: It includes a target detection module and an obstacle avoidance control module; wherein the target detection module includes GCT-YOLOv5s, that is, YOLOv5s that combines a fusion structure of a centralized comprehensive convolution and a global context module, a content-aware reorganization operator, and a task-specific context decoupling head, extracts key features through a C3GC module, that is, a fusion structure module of a centralized comprehensive convolution and a global context module, uses a CARAFE operator, that is, a content-aware reorganization operator to enhance feature fusion, and determines the relationship between classification and positioning through a TSCODE detection head, that is, a task-specific context decoupling detection head, to achieve detection of traffic participant types in driving scenes; the obstacle avoidance control module includes adaptive obstacle avoidance path planning and control, dynamically adjusts the repulsive field in combination with the type of traffic participants, performs path smoothing, and uses a pure tracking algorithm for path tracking control to achieve adaptive obstacle avoidance path planning and control for smart cars; The GCT-YOLOv5s of the target detection module, i.e., YOLOv5s combining the fusion structure of the centralized comprehensive convolution and the global context module, the content-aware reorganization operator and the task-specific context decoupling head, includes a C3GC module, a CARAFE operator, and a TSCODE detection head; The C3GC module, i.e., the fusion structure module of the centralized comprehensive convolution and the global context module, refers to integrating the GlobalContext, i.e., the global context module, into the C3, i.e., the centralized comprehensive convolution structure, of the YOLOv5 backbone network, and adding the Global Context module before the CBS, i.e., the channel block compression layer in the C3 structure to form the C3GC module; The CARAFE operator, i.e., the content-aware recombination operator, refers to the introduction of an upsampling CARAFE operator in the feature fusion part to replace the original upsampling structure of YOLOv5s, which is used to enrich the semantic information of the feature map to enhance feature fusion; The TSCODE detection head, i.e., the task-specific context decoupling detection head, includes a classification task and a localization task, wherein the classification task The semantic context encoding part uses the l and P l+1 The feature maps of the two layers generate feature maps with richer semantics; positioning task The encoding part for local detail preservation accepts feature maps from three pyramid levels, namely P l-1 , P l and P l+1 , P l-1 Provides more details and edge features, while P l+1 Provide a more comprehensive view of objects; The adaptive obstacle avoidance path planning and control of the obstacle avoidance control module includes dynamic adjustment of the repulsive field by introducing a repulsive force attenuation factor and a target type adaptation factor, path smoothing, and a pure tracking algorithm; The dynamic adjustment of the repulsive field includes introducing a repulsive attenuation factor and introducing a target type adaptation factor; wherein, the introduction of the repulsive attenuation factor refers to introducing the repulsive attenuation factor into the repulsive function, and the improved repulsive potential field function expression is: Where m is the position of the smart car, k r is the repulsion gain coefficient, m obs is the distance between the smart car and the obstacle, n is a fixed value, is the repulsive force attenuation factor, m g is the distance between the smart car and the target point; In the process of formula derivation, the present invention sets s as the trigger range of the function, m g is a constant, that is, s is m g The range of change is as follows: The introduction of the target type adaptation factor refers to the introduction of the target type adaptation factor ξ in formula (2). After the introduction of the target type adaptation factor, the repulsion function formula is as follows: The repulsive field is dynamically adjusted according to formula (3).

2. The intelligent vision-guided vehicle obstacle avoidance control optimization method according to claim 1 is characterized in that: The path smoothing process includes temporary target point selection and polynomial curve fitting; wherein, the temporary target point selection includes predicting the future path and selecting a new temporary target point; wherein, the predicting the future path means that the smart car predicts a fixed distance forward, determines the prediction range λ, and generates path points with a constant step size Δ until the predefined range is reached; The selecting of a new temporary target point means that after determining the starting point and the target point, the equation of the resultant straight line at the starting point or the temporary target point is calculated, and the point farthest from the straight line is found as the new temporary target point through the vertical projection of the path points, until the vehicle selects the global target point as the latest temporary target point; The polynomial curve fitting refers to the design of a smooth prediction artificial potential field method by adding a polynomial curve fitting method on the basis of the prediction artificial potential field method when the global target point is selected as the temporary target point; The pure tracking algorithm refers to an obstacle avoidance path adaptively planned by using a pure tracking algorithm to control the intelligent car tracking based on a vehicle kinematic model and combining the smooth prediction artificial potential field method of the traffic participant type.

Images