Vehicle collision avoidance method, system and device based on differential evolution neural network

Abstract

The present invention relates to the field of vehicle automatic driving technology, specifically a method, system and device for avoiding collision of automobiles based on differential evolution neural network, which respectively obtains single-angle images of obstacles of various vehicle sensors in the environment around the vehicle; performs image segmentation on the obtained single-angle images to obtain label information of obstacles in the single-angle images; calculates the Euclidean distance of the vehicle from the obstacle based on the label information of the obstacle; establishes a differential evolution neural network model, and uses a differential evolution algorithm to iteratively optimize the vehicle chromosome encoding of the model; inputs the newly obtained Euclidean distance of the vehicle from the obstacle at different angles into the differential evolution neural network model, and outputs the steering angle of the vehicle; controls the vehicle to avoid collision based on the output steering angle. The method solves the high collision avoidance demand problem of automobile collision avoidance algorithm in complex traffic scenes, adapts to dynamically changing environments, and improves the obstacle avoidance capability and driving safety of vehicles.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle automatic driving, and specifically to a vehicle collision avoidance method, system and equipment based on differential evolution neural network. Background Art

[0002] Vehicle collision avoidance technology is a key link in the autonomous driving technology system. This technology requires the vehicle to comprehensively consider all the information obtained and guide the vehicle to make reasonable path planning before a collision. It plays an important role in vehicle intelligence and is therefore one of the key research technologies in the field of autonomous driving.

[0003] Most traditional car collision avoidance algorithms rely on a safe distance model to warn of potential collision risks, but this model cannot meet the collision avoidance requirements in complex traffic scenarios. For example, the artificial potential field method uses the concepts of gravity and repulsion to guide vehicles to avoid obstacles, but in complex traffic scenarios, this method may encounter local minima, causing the vehicle to fall into a local optimal solution and unable to reach the target point, limiting its application effect in the field of autonomous driving safety.

[0004] Compared with traditional algorithms, the car collision avoidance algorithm based on differential evolution neural network has significant advantages. The evolutionary algorithm searches for the optimal solution of the vehicle driving strategy by simulating the genetic and mutation mechanisms in biological evolution. At the same time, it uses neural networks to deal with nonlinear problems in complex traffic scenarios. In addition, the differential evolution algorithm can adaptively adjust the search strategy according to the difficulty and characteristics of the problem. Through strategies such as adaptive mutation rate and co-evolution of dual population structure, it can speed up the search of neural network structure and improve the efficiency and effect of the car collision avoidance algorithm. In this way, it can solve the high collision avoidance demand problem of the car collision avoidance algorithm in complex traffic scenarios, adapt to the dynamically changing environment, and improve the vehicle's obstacle avoidance ability and driving safety. Summary of the invention

[0005] To solve the above problems, the first aspect of the present invention provides a vehicle collision avoidance method based on a differential evolution neural network, comprising:

[0006] S1, based on vehicle sensors at different angles of the vehicle, obtaining single-angle images of obstacles in the vehicle's surrounding environment from each vehicle sensor;

[0007] S2. performing image segmentation on the acquired single-angle images to obtain label information of obstacles in the single-angle images of the vehicle; and calculating the Euclidean distance between the vehicle and the obstacle based on the label information of the obstacles in the single-angle images of the vehicle;

[0008] S3, establishing a differential evolution neural network model, performing model training based on the Euclidean distance of the vehicle from the obstacle at multiple angles, and iteratively optimizing the vehicle chromosome coding of the model using a differential evolution algorithm, wherein the vehicle chromosome coding is the vehicle control parameter of the neural network in the differential evolution neural network model;

[0009] S4, inputting the newly acquired Euclidean distances of the vehicle from the obstacle at different angles into the differential evolution neural network model, assigning the iteratively optimized vehicle control parameters to the weight parameters and bias value parameters of the neurons in each layer of the model neural network, and calculating and outputting the steering angle of the vehicle according to the input Euclidean distance;

[0010] S4.1. The Euclidean distances of the vehicle from the obstacle at different angles are used as input data, and are calculated through each layer of neurons in the input layer of the neural network. The input data is multiplied by the weight parameters of the neurons, and the bias value parameters are considered. The output data of each layer of the input layer is obtained by the activation function.

[0011] S4.2, input the output data of each layer of the input layer to the hidden layer of the neural network, calculate through each layer of neurons in the hidden layer of the neural network, use the output data of each layer of the hidden layer as the input data of the corresponding layer, and calculate the output data of each layer of neurons in the hidden layer. The calculation formula is:

[0012] ,

[0013] in, For the hidden layer j The input data of the layer, For the j Layer node to i The weights corresponding to the layer nodes, N is the total number of neurons in the input layer, For the i The bias of the neurons in the layer, Hide i Indicates i Output data of a single neuron in layer 2; f is the activation function, ;

[0014] S4.3, the output data 2 is calculated by the output layer neurons of the neural network to obtain the output data 3, the output data 3 is the steering angle of the vehicle; the output layer neuron calculation formula is:

[0015] ,

[0016] in, F is the output data of the output layer, For the iLayer node to c The corresponding weights of the layer nodes, P is the total number of neurons in the hidden layer, b c is the bias value of the output layer, f c is the activation function of the output layer;

[0017] S5. Control the vehicle to avoid collision based on the output steering angle.

[0018] The differential evolution neural network model includes an information acquisition module, a neural network and an optimization module;

[0019] The information acquisition module is used to collect the Euclidean distance data of obstacles in a number of single-angle images of the vehicle sensor and input them into the neural network;

[0020] The neural network includes an input layer, an output layer and a hidden layer, and the connection mode of each layer is fully connected; the neural network is used to input multi-angle Euclidean distance data, and obtain output data of vehicle steering angle after processing in each layer;

[0021] The optimization module is used to establish a sample data set based on the Euclidean distances of vehicles from obstacles at multiple angles, and to iteratively optimize the vehicle chromosome encoding of the neural network in the model using a differential evolution algorithm.

[0022] The model training is performed based on the Euclidean distance of the vehicle from the obstacle at multiple angles. Specifically, the Euclidean distance of the obstacle at multiple angles in the sample data set is used as the input value, and the car collision avoidance result guided by a neural network optimized by a set of vehicle chromosome encoding is used as the target value to train the differential evolution neural network model.

[0023] As a specific implementation of the first aspect, the differential evolution algorithm is used to iteratively optimize the vehicle chromosome encoding of the neural network in the model, and the specific operation is:

[0024] S3.1. Initialize the population individuals, including the vehicle chromosome encoding length and encoding method;

[0025] S3.2, preset a fitness threshold, and perform fitness evaluation on the vehicle chromosome encoding according to the fitness threshold;

[0026] If a set of vehicle chromosome codes, based on the Euclidean distance of obstacles at multiple angles in the input sample data set, obtains a target value of the vehicle collision avoidance result that is less than the fitness threshold, then perform S3.3 operation;

[0027] S3.3, updating the vehicle chromosome encoding based on the evolution operator;

[0028] S3.3.1. Traversing the vehicle chromosome coded individuals in the populationX t , randomly select three non-repeating individuals from the population, respectively X a , X b and X c ;

[0029] S3.3.2. Randomly select two individuals from the three individuals and perform differential operation to obtain the differential vector Diff ;

[0030] S3.3.3. Add the remaining individuals to the differential vector to obtain the mutant individuals Muta ;

[0031] S3.3.4. Mutation of individuals based on the upper and lower limits of the determined search space Muta Perform bounds checking;

[0032] S3.4. Traversing the population individuals X t and variant individuals Muta , preset fitness evaluation indicators to select population individuals and obtain new population individual encoding information X t-new ;

[0033] S3.5. Preset the maximum number of iterations for iterative optimization; when all individuals in the population have been traversed once, enter the next iteration; if the number of iterations is equal to the maximum number of iterations, stop the iterative optimization to determine the vehicle chromosome coding, and determine the differential evolution neural network model of the vehicle chromosome coding; if the number of iterations is less than the maximum number of iterations, return to S3.2.

[0034] In a specific implementation, the vehicle chromosome encoding length is initialized in 3.1, and the vehicle chromosome encoding length is determined based on the hierarchical structure of the neural network and the number of neurons in each layer. LE The calculation formula is:

[0035] ,

[0036] in, It is i The number of neurons in the layer.

[0037] S3.4 X t-new The selection formula is as follows:

[0038] ,

[0039] in, Fitness() function represents the fitness evaluation index.

[0040] The label information includes the coordinates of the obstacle in the image, the characteristic scale of the obstacle, and the direction of the obstacle.

[0041] As a specific implementation of the first aspect, the output steering angle includes a left steering angle and a right steering angle of the vehicle.

[0042] The second aspect provides a car collision avoidance system based on a differential evolution neural network, comprising:

[0043] A sensor module, for acquiring single-angle images of obstacles in the surrounding environment of the vehicle respectively based on vehicle sensors at different angles of the vehicle;

[0044] An image module is used to perform image segmentation on the acquired single-angle images to obtain label information of obstacles in the single-angle images of the vehicle; and calculate the Euclidean distance between the vehicle and the obstacle based on the label information of the obstacles in the single-angle images of the vehicle;

[0045] A model training module is used to establish a differential evolution neural network model, perform model training based on the Euclidean distance of the vehicle from the obstacle at multiple angles, and iteratively optimize the vehicle chromosome coding of the model using a differential evolution algorithm. The vehicle chromosome coding is the vehicle control parameter of the neural network in the differential evolution neural network model.

[0046] A processing module is used to input the newly acquired Euclidean distances of the vehicle from the obstacle at different angles into the differential evolution neural network model, assign the iteratively optimized vehicle control parameters to the weight parameters and bias value parameters of the neurons in each layer of the model neural network, and calculate and output the steering angle of the vehicle according to the input Euclidean distance; use the Euclidean distances of the vehicle from the obstacle at different angles as input data, calculate through each layer of neurons in the input layer of the neural network, multiply the input data with the weight parameters of the neurons, and consider the bias value parameters, and obtain the output data one of each layer of the input layer by the activation function; input the output data one of each layer of the input layer to the hidden layer of the neural network, calculate through each layer of neurons in the hidden layer of the neural network, use the output data one input to each layer of the hidden layer as the input data of the corresponding layer, and calculate to obtain the output data two of the neurons in each layer of the hidden layer; calculate the output data two through the neurons in the output layer of the neural network to obtain output data three, and the output data three is the steering angle of the vehicle;

[0047] The control module is used to control the vehicle to avoid collision based on the output steering angle.

[0048] The third aspect provides a car collision avoidance device based on a differential evolution neural network, comprising a processor and a memory, wherein the processor implements the car collision avoidance method based on a differential evolution neural network as described above when executing program data stored in the memory.

[0049] The beneficial effect is that the present invention is a method for avoiding automobile collisions based on a differential evolution neural network. Single-angle images of obstacles in the vehicle's surrounding environment are acquired from different angles, and the acquired multiple single-angle images are segmented to obtain label information of the vehicle at a single angle. At the same time, the label information of the vehicle at a single angle is input into a differential evolution neural network model to output the steering angle of the vehicle, wherein the differential evolution neural network model is obtained by training the search space corresponding to single-angle images at different angles and the differential evolution algorithm. The vehicle guides the vehicle to avoid collisions based on the steering angle output by the differential evolution neural network model, thereby solving the high collision avoidance requirement problem of the vehicle collision avoidance algorithm in complex traffic scenarios and improving the vehicle's obstacle avoidance capability and driving safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of a process flow of a vehicle collision avoidance method provided by the present invention;

[0051] Figure 2 This is a diagram of the differential evolution neural network model;

[0052] Figure 3 It is a motion diagram of a single time step of the vehicle simulation model;

[0053] Figure 4 This is a schematic diagram of vehicle sensor perception;

[0054] Figure 5 This is a schematic diagram of the update when the vehicle sensor encounters an obstacle;

[0055] Figure 6 This is a schematic diagram of vehicle collision detection;

[0056] Figure 7 This is a dynamic simulation training diagram. DETAILED DESCRIPTION

[0057] As mentioned above, the traditional car collision avoidance algorithm cannot meet the collision avoidance requirements in complex traffic scenarios. In this regard, an embodiment of the present invention provides a car collision avoidance method based on a differential evolution neural network. The contents provided by the present invention are described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] Figure 1 A schematic diagram of a process flow of a vehicle collision avoidance method provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the specific implementation steps of the method are as follows:

[0059] S1, based on vehicle sensors at different angles of the vehicle, obtaining single-angle images of obstacles in the vehicle's surrounding environment from each vehicle sensor;

[0060] By setting up multiple vehicle sensors at different angles at the front of the vehicle, when an obstacle is sensed in front, the sensor images collected at different angles in front of the vehicle at the same time can be obtained.

[0061] The angle detection range of the vehicle sensor is from 0° to 360°, and the angles between adjacent sensors are the same.

[0062] S2. performing image segmentation on the acquired single-angle images to obtain label information of obstacles in the single-angle images of the vehicle; and calculating the Euclidean distance between the vehicle and the obstacle based on the label information of the obstacles in the single-angle images of the vehicle;

[0063] Based on vehicle sensors at different angles, several single-angle images of obstacles in front of the vehicle are obtained respectively, and image segmentation is performed on the several single-angle images to obtain label information of obstacles in the single-angle images of the vehicle.

[0064] The image segmentation algorithm is used to detect the feature points in the single-angle image of the vehicle in sequence, and the label information in the feature points is extracted; the label information includes the coordinates of the obstacle in the image, the characteristic size of the obstacle, and the direction of the obstacle.

[0065] Based on the obstacle label information of each single-angle image, the Euclidean distance between the vehicle and the obstacle in each image is calculated in turn.

[0066] S3, establishing a differential evolution neural network model, performing model training based on the Euclidean distance of the vehicle from the obstacle at multiple angles, and iteratively optimizing the vehicle chromosome coding of the model using a differential evolution algorithm, wherein the vehicle chromosome coding is the vehicle control parameter of the neural network in the differential evolution neural network model;

[0067] like Figure 2 As shown, a differential evolution neural network model is established, and the differential evolution neural network model includes an information acquisition module, a neural network and an optimization module. The information acquisition module is used to collect the Euclidean distance data of obstacles in a number of single-angle images of the vehicle sensor and input them into the neural network;

[0068] The neural network includes an input layer, an output layer and a hidden layer, and the connection mode of each layer is fully connected; the neural network is used to input multi-angle Euclidean distance data, and obtain output data of vehicle steering angle after processing in each layer;

[0069] The optimization module is used to establish a sample data set based on the Euclidean distances of vehicles from obstacles at multiple angles, and to iteratively optimize the vehicle chromosome encoding of the neural network in the model using a differential evolution algorithm.

[0070] In this embodiment, it should be noted that the model training is performed based on the Euclidean distance of the vehicle from the obstacle at multiple angles, specifically, the Euclidean distance of the obstacle at multiple angles in the sample data set is used as the input value, and the car collision avoidance result guided by a neural network optimized by a set of vehicle chromosome encoding is used as the target value to train the differential evolution neural network model.

[0071] The differential evolution algorithm is used to iteratively optimize the vehicle chromosome encoding of the model, and the specific operation is:

[0072] S3.1. Initialize the population individuals, including the vehicle chromosome encoding length and encoding method;

[0073] Initializing the vehicle chromosome coding length, wherein the vehicle chromosome coding length is determined based on the hierarchical structure of the neural network and the number of neurons in each layer, and the vehicle chromosome coding length is equal to the sum of the vehicle control parameters of the neural network;

[0074] For those with L layer of neural network, vehicle chromosome encoding length LE The calculation formula is:

[0075] ,

[0076] in, It is i The number of neurons in the layer;

[0077] Initialize the vehicle chromosome encoding method. The vehicle chromosome encoding is the vehicle control parameters of the neural network in the differential evolution neural network model. The vehicle control parameters of the neural network are also the weight parameters of the edge connections of each layer of neurons and the bias value parameters of the neurons.

[0078] The upper and lower limits of the corresponding search space are determined according to the vehicle chromosome encoding method, and the value range of the parameter is dynamically adjusted.

[0079] S3.2, preset a fitness threshold, and perform fitness evaluation on the vehicle chromosome encoding according to the fitness threshold;

[0080] If a set of vehicle chromosome codes, based on the Euclidean distance of obstacles at multiple angles in the input sample data set, obtains a target value of the vehicle collision avoidance result that is less than the fitness threshold, then perform S3.3 operation;

[0081] If the target value of the vehicle collision avoidance result is greater than or equal to the fitness threshold, the chromosome encoding of this group of vehicles is retained.

[0082] S3.3. Update the vehicle chromosome code based on the evolution operator. The specific operation is as follows:

[0083] S3.3.1. Traversing the vehicle chromosome coded individuals in the population X t , randomly select three non-repeating individuals from the population, respectively X a , X b and X c ;

[0084] S3.3.2. Randomly select two individuals from the three individuals and perform differential operation to obtain the differential vector Diff , the difference vector Diff The calculation formula is:

[0085] ,

[0086] in, X a , X b are two individuals selected at random;

[0087] S3.3.3. Add the remaining individuals to the differential vector to obtain the mutant individuals Muta , the calculation formula for variant individuals is:

[0088] ,

[0089] in, X c are the remaining individuals; Diff is the difference vector, F Represents the differential weight, set to 1;

[0090] S3.3.4. Mutation of individuals based on the upper and lower limits of the determined search space Muta Perform bounds checking;

[0091] S3.4. Traversing the population individuals X t and variant individuals Muta , preset fitness evaluation indicators to select population individuals and obtain new population individual encoding information X t-new , X t-new The selection formula is as follows:

[0092] ,

[0093] in, Fitness () function represents the fitness evaluation index;

[0094] S3.5. Preset the maximum number of iterations for iterative optimization; when all individuals in the population have been traversed once, enter the next iteration; if the number of iterations is equal to the maximum number of iterations, stop the iterative optimization to determine the vehicle chromosome coding, and determine the differential evolution neural network model of the vehicle chromosome coding; if the number of iterations is less than the maximum number of iterations, return to S3.2.

[0095] To facilitate the understanding of those skilled in the art, and to further explain the simulation process of the differential evolution neural network model, this application sets up a simulation environment to conduct relevant simulation experiments, and examples are introduced as follows:

[0096] like Figure 3 As shown in the figure, a vehicle simulation model is established in the simulation environment. Given the speed of the vehicle simulation model and a time step (1 second), the movement of the vehicle simulation model in a time step is calculated by the speed and time step. Given the direction of the vehicle, Figure 3 It is marked in the form of an arrow; given the center of mass position of the vehicle, the center of mass position is used for the horizontal driving angle of the vehicle simulation model in the simulation environment.

[0097] When the vehicle is driving in the simulation environment, its initial angle is θ ,exist Figure 3 At the solid point in the figure, the vehicle's rotation angle is output based on the differential evolution neural network model. α , thus instructing the vehicle to change direction and move. At this time, the adjusted vehicle horizontal driving angle is the new value θ new , which is different from the initial angle θ And the rotation angle α There are the following relationships:

[0098] .

[0099] A vehicle sensor model is established based on a vehicle simulation model. In the present invention, sensors are used to perceive the vehicle's surrounding environment in real time, and single-angle images of obstacles in the vehicle's surrounding environment are obtained from each vehicle sensor. The vehicle simulation model is equipped with several sensors for simulation experiments of the simulated environment, such as Figure 4 As shown in the figure, the vehicle sensor model can achieve a semicircular sensing area from 0° to 180°, completely covering the sensor range in front and on the sides of the vehicle. Each line in the figure represents the sensor direction of the vehicle at a specific angle, and takes the vehicle as the starting point to preset the sensor's sensing distance.

[0100] When simulating the vehicle simulation model in this simulation environment, the vehicle sensor model includes 19 angle values, and each angle is evenly distributed to ensure that the vehicle has a comprehensive perception of the surrounding environment and monitors the surrounding conditions in multiple directions in real time, thereby effectively detecting obstacles.

[0101] After obtaining the single-angle images of each vehicle sensor of obstacles in the vehicle's surrounding environment based on the vehicle sensor model, it is necessary to perform image segmentation operations on the obtained single-angle images to obtain label information of the obstacles in the single-angle images.

[0102] In this simulation, the vehicle uses the beam emitted by the sensor to update the Euclidean distance between the vehicle and obstacles (including other vehicles and environmental boundaries) in real time. For each sensor, the system searches for and records the distance to the nearest obstacle. If no obstacle is detected, the sensor reading defaults to the preset maximum sensing distance; if an obstacle is detected, the actual distance from the sensor to the obstacle is recorded and used as the sensor reading. Figure 5 This is a schematic diagram of the update when the vehicle sensor encounters an obstacle.

[0103] A differential evolution neural network model is established, and the vehicle chromosome encoding of the model is iteratively optimized using a differential evolution algorithm; the newly acquired Euclidean distances of the vehicle from obstacles at different angles are input into the differential evolution neural network model, and the steering angle of the vehicle is output; the vehicle is controlled to avoid collision based on the output steering angle.

[0104] In the simulation environment, the vector cross product method is used to detect and judge the collision of vehicles. Figure 6 As shown in the figure, the sensor line segment is regarded as a point with a specific direction, and the obstacle value in the simulation environment is digitized. The intersection of the sensor line segment and the obstacle line segment is solved by the vector cross product method, so as to determine the intersection of the sensor line segment and other line segments. Once the line segments intersect, it means that the vehicle has collided, and the next chromosome needs to be replaced. If all chromosomes have been used up, the differential evolution algorithm must enter the next iteration. With the help of algebraic calculations, the relationship between the vehicle and the obstacle can be carefully portrayed, and the vehicle collision detection can be performed.

[0105] At the same time, if Figure 7As shown in the figure, multiple vehicle simulation models are set up in the simulation environment for dynamic simulation training. The simulation environment is set to a free space area, in which eight vehicles are driving freely. Each vehicle simulation model is equipped with an artificial neural network. Subsequently, the neural network of each vehicle is trained by the differential evolution algorithm so that it can drive without colliding with the environment boundary or other vehicles, that is, the collision avoidance of the vehicle is required to be able to handle not only static obstacles, but also to identify and respond to dynamic vehicles.

[0106] Under the constraints of the dynamic movement of the vehicle and static obstacles, the differential evolution algorithm is used to iteratively optimize the vehicle chromosome encoding of the model. The neural network model outputs the steering angle of the vehicle, and controls the vehicle to avoid collision based on the output steering angle.

[0107] The above examples are an introduction to the relevant experimental design of simulation, in order to further explain the simulation process of the differential evolution neural network model. It should be noted that the simulation environment shown here is only one of many simulation environments. In addition, in the actual application of this model, a more specific simulation environment needs to be set according to different training purposes.

[0108] S4, inputting the newly acquired Euclidean distances of the vehicle from the obstacle at different angles into the differential evolution neural network model, assigning the iteratively optimized vehicle control parameters to the weight parameters and bias value parameters of the neurons in each layer of the model neural network, and calculating and outputting the steering angle of the vehicle according to the input Euclidean distance;

[0109] S4.1. The Euclidean distances of the vehicle from the obstacle at different angles are used as input data, and are calculated through each layer of neurons in the input layer of the neural network. The input data is multiplied by the weight parameters of the neurons, and the bias value parameters are considered. The output data of each layer of the input layer is obtained by the activation function.

[0110] As a specific implementation method, the Euclidean distances of the vehicle from the obstacle at different input angles are obtained as input data and processed by a differential evolution neural network model, that is, calculation processing is performed by neurons of the neural network in the model.

[0111] The number of neurons in the neural network input layer is the number of single-angle images of obstacles acquired by the vehicle. The input data is calculated and processed to obtain output data 1, and the calculation formula is:

[0112] ,

[0113] in, X is the input signal of each layer, W is the weight corresponding to the input layer neuron,b is the bias value, f is the activation function, O is the output signal of each layer.

[0114] S4.2, inputting the output data 1 of each layer of the input layer into the hidden layer of the neural network, performing calculations through each layer of neurons in the hidden layer of the neural network, using the output data 1 input to each layer of the hidden layer as the input data of the corresponding layer, and calculating to obtain the output data 2 of each layer of neurons in the hidden layer;

[0115] The neurons in the hidden layer select the nonlinear activation function Sigmoid function with a threshold, and its formula is:

[0116] ,

[0117] Hidden layer i The output calculation formula of a single neuron in the layer is:

[0118] ,

[0119] in, For the hidden layer j The input data of the layer, For the j Layer node to i The weights corresponding to the layer nodes, N is the total number of neurons in the input layer, For the i The bias of the neurons in the layer, Hide i Indicates i Output data of a single neuron in layer 2;

[0120] S4.3, calculating the output data 2 through the neurons of the output layer of the neural network to obtain the output data 3, wherein the output data 3 is the steering angle of the vehicle;

[0121] The output calculation formula of the output layer neurons is:

[0122] ,

[0123] in, F is the output data of the output layer, For the i Layer node to c The corresponding weights of the layer nodes, P is the total number of neurons in the hidden layer, b c is the bias value of the output layer, f c is the activation function of the output layer;

[0124] Based on the trained differential evolution neural network model; the Euclidean distances of the vehicle from the obstacle at different angles are input into the differential evolution neural network model as input data, and the steering angle of the vehicle is output, and the output steering angle includes the left steering angle and the right steering angle of the car.

[0125] S5. Control the vehicle to avoid collision based on the output steering angle.

[0126] The vehicle optimizes its driving strategy based on the steering angle of the vehicle output by the differential evolution neural network model, effectively achieving collision avoidance.

[0127] In general, the differential evolution neural network model described in the present invention can adaptively adjust the search strategy according to the difficulty and characteristics of the problem, speed up the search of the neural network structure, improve the efficiency and effect of the vehicle collision avoidance algorithm, solve the high collision avoidance demand problem of the vehicle collision avoidance algorithm in complex traffic scenarios, adapt to the dynamically changing environment, and improve the vehicle's obstacle avoidance ability and driving safety.

[0128] In addition, the present application also provides a car collision avoidance system based on differential evolution neural network, including:

[0129] A sensor module, for acquiring single-angle images of obstacles in the surrounding environment of the vehicle respectively based on vehicle sensors at different angles of the vehicle;

[0130] An image module is used to perform image segmentation on the acquired single-angle images to obtain label information of obstacles in the single-angle images of the vehicle; and calculate the Euclidean distance between the vehicle and the obstacle based on the label information of the obstacles in the single-angle images of the vehicle;

[0131] A model training module is used to establish a differential evolution neural network model, perform model training based on the Euclidean distance of the vehicle from the obstacle at multiple angles, and iteratively optimize the vehicle chromosome coding of the model using a differential evolution algorithm. The vehicle chromosome coding is the vehicle control parameter of the neural network in the differential evolution neural network model.

[0132] A processing module is used to input the newly acquired Euclidean distances of the vehicle from the obstacle at different angles into the differential evolution neural network model, assign the iteratively optimized vehicle control parameters to the weight parameters and bias value parameters of the neurons in each layer of the model neural network, and calculate and output the steering angle of the vehicle according to the input Euclidean distance; use the Euclidean distances of the vehicle from the obstacle at different angles as input data, calculate through each layer of neurons in the input layer of the neural network, multiply the input data with the weight parameters of the neurons, and consider the bias value parameters, and obtain the output data one of each layer of the input layer by the activation function; input the output data one of each layer of the input layer to the hidden layer of the neural network, calculate through each layer of neurons in the hidden layer of the neural network, use the output data one input to each layer of the hidden layer as the input data of the corresponding layer, and calculate to obtain the output data two of the neurons in each layer of the hidden layer; calculate the output data two through the neurons in the output layer of the neural network to obtain output data three, and the output data three is the steering angle of the vehicle;

[0133] The control module is used to control the vehicle to avoid collision based on the output steering angle.

[0134] Finally, a car collision avoidance device based on a differential evolution neural network is also provided, comprising a processor and a memory, wherein the processor implements the car collision avoidance method based on a differential evolution neural network as described above when executing program data stored in the memory.

Claims

1. A car collision avoidance method based on differential evolution neural network, characterized in that: include: S1, based on vehicle sensors at different angles of the vehicle, obtaining single-angle images of obstacles in the vehicle's surrounding environment from each vehicle sensor; S2. performing image segmentation on the acquired single-angle images to obtain label information of obstacles in the single-angle images of the vehicle; and calculating the Euclidean distance between the vehicle and the obstacle based on the label information; S3, establishing a differential evolution neural network model, performing model training based on the Euclidean distance of the vehicle from the obstacle at multiple angles, and iteratively optimizing the vehicle chromosome encoding of the model using a differential evolution algorithm, wherein the vehicle chromosome encoding is a vehicle control parameter of the neural network in the differential evolution neural network model; S4, inputting the newly acquired Euclidean distances of different angles into the differential evolution neural network model, assigning the iteratively optimized vehicle control parameters to the weight parameters and bias value parameters of the neurons in each layer of the neural network, and calculating and outputting the steering angle of the vehicle according to the input Euclidean distance; S4.

1. Taking the Euclidean distance at different angles as input data, calculating through each layer of neurons in the input layer of the neural network, multiplying the input data with the weight parameter of the neuron, and considering the bias value parameter, the output data of each layer of the input layer is obtained by the activation function; S4.2, inputting the output data 1 of each layer of the input layer into the hidden layer of the neural network, performing calculations through each layer of neurons in the hidden layer of the neural network, and calculating output data 2 of each layer of neurons in the hidden layer; S4.3, calculating the output data 2 through the neurons of the output layer of the neural network to obtain the output data 3, wherein the output data 3 is the steering angle of the vehicle; S5. Control the vehicle to avoid collision based on the output steering angle.

2. The vehicle collision avoidance method based on differential evolution neural network according to claim 1 is characterized in that: The differential evolution neural network model includes an information acquisition module, a neural network and an optimization module; The information acquisition module is used to collect the Euclidean distance data of obstacles in a number of single-angle images of the vehicle sensor and input them into the neural network; The neural network includes an input layer, an output layer and a hidden layer, and the connection mode of each layer is fully connected; the neural network is used to input multi-angle Euclidean distance data, and obtain output data of vehicle steering angle after processing in each layer; The optimization module is used to establish a sample data set based on the Euclidean distances of vehicles from obstacles at multiple angles, and to iteratively optimize the vehicle chromosome encoding of the neural network in the model using a differential evolution algorithm.

3. The vehicle collision avoidance method based on differential evolution neural network according to claim 2 is characterized in that: The model training is performed based on the Euclidean distance of the vehicle from the obstacle at multiple angles. Specifically, the Euclidean distance of the obstacle at multiple angles in the sample data set is used as the input value, and the car collision avoidance result guided by a neural network optimized by a set of vehicle chromosome encoding is used as the target value to train the differential evolution neural network model.

4. The vehicle collision avoidance method based on differential evolution neural network according to claim 2 is characterized in that: The differential evolution algorithm is used to iteratively optimize the vehicle chromosome encoding of the neural network in the model, and the specific operation is: S3.

1. Initialize the population individuals, including the vehicle chromosome encoding length and encoding method; S3.2, preset a fitness threshold, and perform fitness evaluation on the vehicle chromosome encoding according to the fitness threshold; If a set of vehicle chromosome codes, based on the Euclidean distance of obstacles at multiple angles in the input sample data set, obtains a target value of the vehicle collision avoidance result that is less than the fitness threshold, then perform S3.3 operation; S3.3, updating the vehicle chromosome encoding based on the evolution operator; S3.3.

1. Traversing the vehicle chromosome coded individuals in the population X t , randomly select three non-repeating individuals from the population, respectively X a , X b and X c ; S3.3.

2. Randomly select two individuals from the three individuals and perform differential operation to obtain the differential vector Diff ; S3.3.

3. Add the remaining individuals to the differential vector to obtain the mutant individuals Muta ; S3.3.

4. Mutation of individuals based on the upper and lower limits of the determined search space Muta Perform bounds checking; S3.

4. Traversing the population individuals X t and variant individuals Muta , preset fitness evaluation indicators to select population individuals and obtain new population individual encoding information X t-new ; S3.

5. Preset the maximum number of iterations for iterative optimization; when all individuals in the population have been traversed once, enter the next iteration; If the number of iterations is equal to the maximum number of iterations, the iterative optimization is stopped to determine the vehicle chromosome encoding, and the differential evolution neural network model of the vehicle chromosome encoding is determined; If the number of iterations is less than the maximum number of iterations, return to S3.

2.

5. The vehicle collision avoidance method based on differential evolution neural network according to claim 4 is characterized in that: The vehicle chromosome encoding length is determined based on the hierarchical structure of the neural network and the number of neurons in each layer. LE The calculation formula is: ， in, It is i The number of neurons in the layer.

6. The vehicle collision avoidance method based on differential evolution neural network according to claim 4 is characterized in that: S3.4 X t-new The selection formula is as follows: ， in, Fitness () function represents the fitness evaluation index.

7. The vehicle collision avoidance method based on differential evolution neural network according to claim 1 is characterized in that: The label information includes the coordinates of the obstacle in the image, the characteristic scale of the obstacle, and the direction of the obstacle.

8. The vehicle collision avoidance method based on differential evolution neural network according to claim 1 is characterized in that: The output steering angle includes a left steering angle and a right steering angle of the vehicle.

9. A car collision avoidance system based on differential evolution neural network, characterized in that: include: A sensor module, for acquiring single-angle images of obstacles in the surrounding environment of the vehicle respectively based on vehicle sensors at different angles of the vehicle; An image module is used to perform image segmentation on the acquired single-angle images to obtain label information of obstacles in the single-angle images of the vehicle; and calculate the Euclidean distance between the vehicle and the obstacle based on the label information of the obstacles in the single-angle images of the vehicle; A model training module is used to establish a differential evolution neural network model, perform model training based on the Euclidean distance of the vehicle from the obstacle at multiple angles, and iteratively optimize the vehicle chromosome coding of the model using a differential evolution algorithm. The vehicle chromosome coding is the vehicle control parameter of the neural network in the differential evolution neural network model. A processing module is used to input the newly acquired Euclidean distances of the vehicle from the obstacle at different angles into the differential evolution neural network model, assign the iteratively optimized vehicle control parameters to the weight parameters and bias value parameters of the neurons in each layer of the model neural network, and calculate and output the steering angle of the vehicle according to the input Euclidean distance; use the Euclidean distances of the vehicle from the obstacle at different angles as input data, calculate through each layer of neurons in the input layer of the neural network, multiply the input data with the weight parameters of the neurons, and consider the bias value parameters, and obtain the output data one of each layer of the input layer by the activation function; input the output data one of each layer of the input layer to the hidden layer of the neural network, calculate through each layer of neurons in the hidden layer of the neural network, use the output data one input to each layer of the hidden layer as the input data of the corresponding layer, and calculate to obtain the output data two of the neurons in each layer of the hidden layer; calculate the output data two through the neurons in the output layer of the neural network to obtain output data three, and the output data three is the steering angle of the vehicle; The control module is used to control the vehicle to avoid collision based on the output steering angle.

10. A car collision avoidance device based on differential evolution neural network, characterized in that: It comprises a processor and a memory, wherein when the processor executes the program data stored in the memory, the automobile collision avoidance method based on a differential evolution neural network as described in any one of claims 1 to 8 is implemented.

Images