Road classification detection method of BP neural network based on genetic algorithm optimization

Abstract

The invention discloses a road classification detection method based on a BP neural network optimized by a genetic algorithm, and belongs to the technical field of intelligent traffic. An accelerometer, a gyroscope, a GPS module, a camera and other devices are installed on a vehicle, and acceleration and speed data of the vehicle under different types of roads are collected; model input is determined, and a BP neural network is built; optimizing and improving the neural network model by adopting a genetic algorithm; and training the model, and performing road classification prediction by using the collected data. The method can be widely applied to the field of intelligent traffic technical equipment, and particularly has remarkable technical advantages and application prospects in the aspect of road environment understanding of an automatic driving system. Compared with other classification methods, the road classification detection method based on the sensor acquisition data and the neural network algorithm has the characteristics of wide application range and high accuracy, can provide accurate road condition data for traffic management departments, and helps to optimize road maintenance decisions.

Description

Technical Field

[0001] The invention belongs to the field of intelligent transportation technology, and in particular relates to a road classification detection method based on a BP neural network optimized by a genetic algorithm. Background Art

[0002] With the development of autonomous driving technology, autonomous driving systems need to be able to accurately identify the type of roads they are traveling on in order to adapt to different road environments. How to effectively collect and analyze road information has become one of the key technologies to improve traffic safety, road maintenance efficiency, and driving experience. Traditional methods of collecting road information mainly rely on ground sensors, cameras, and satellite remote sensing equipment. Although these technologies are mature, they have certain limitations, such as high cost, complex deployment, and poor real-time performance. Most road classification methods rely on manually annotated data, which is not only time-consuming and labor-intensive, but also prone to misjudgment.

[0003] However, the vision-based environmental perception technology in autonomous driving is easily disturbed by environmental factors such as lighting and alternating light and dark. Under strong light or backlight conditions, the image quality of the camera may deteriorate, causing the autonomous driving system to be unable to accurately identify the road conditions ahead. Especially when driving at night with poor road lighting conditions, the camera cannot accurately collect road information, and thus cannot effectively adjust driving parameters, affecting passenger comfort.

[0004] In recent years, using vehicle-mounted sensors to collect road information has become an innovative solution. As an important dynamic monitoring tool, vehicle acceleration sensors can capture the acceleration changes of vehicles in real time during driving. By analyzing these data, the condition of the road surface, the curvature of the curve, whether the road surface is flat, and other relevant information of the traffic environment can be judged. Compared with traditional methods, this sensor-based approach not only has lower costs, but also can achieve more accurate real-time monitoring and has broad application prospects. At present, there are no reports on road classification detection methods based on sensor collection data and neural network algorithms.

[0005] Therefore, a new technical solution is urgently needed in the prior art to solve this problem. Summary of the invention

[0006] The technical problem to be solved by the present invention is: to provide a road classification detection method based on a BP neural network optimized by a genetic algorithm to solve the problems of high cost, complex deployment and poor real-time performance of the current road information collection method; and there is no relevant report on the road classification detection method based on sensor data collection and a neural network algorithm.

[0007] A road classification detection method based on a BP neural network optimized by a genetic algorithm comprises the following steps, and the following steps are performed in sequence:

[0008] Step 1: Data collection

[0009] Collect the acceleration, three-axis rotational speed, driving speed, and environmental image data of the vehicle during driving on various types of roads at the set sampling interval, and remove abnormal data and invalid data from the collected data to obtain the preliminarily screened and sorted data;

[0010] Step 2: Data preprocessing

[0011] Smooth the acceleration, rotational speed, and speed data in the preliminarily screened and sorted data to remove noise and interference, obtain the three-axis acceleration, three-axis rotational speed, and driving speed data and store them in a combined form in a table, identify the road type in the image through the environmental image data and store it corresponding to the table data;

[0012] Step 3: Build a BP neural network model

[0013] The BP neural network model has an input layer, a hidden layer, and an output layer; select some or all of the three-axis acceleration, three-axis rotational speed, and driving speed as feature parameters, build a BP neural network model according to the extracted feature parameters, use the table data of the selected feature parameters as the input of the BP neural network model, and the corresponding road type as the output of the BP neural network model, and then determine the neural network topology;

[0014] Step 4: Optimize the initial weights and thresholds of the BP neural network model using a genetic algorithm;

[0015] According to the topology of the BP neural network model, determine the coding rule, select the initial population according to the set genetic algorithm objective function, and judge whether the set termination condition is satisfied. If satisfied, proceed to Step 5; if not, perform genetic variation to form offspring, and then perform the objective function calculation until the termination condition is satisfied. The obtained optimal initial weights and thresholds are used as the initial weights and thresholds of the BP neural network model;

[0016] Step 5: Model training:

[0017] Use the data preprocessed in Step 2 to train the BP neural network model, continuously update the learning rate, weights, and thresholds until the set BP neural network model objective function value is lower than the set value, and obtain the trained BP neural network model;

[0018] Step 6: Road classification prediction:

[0019] Input the real-time acceleration, three-axis rotational speed, driving speed, and environmental image data collected by the vehicle into the trained BP neural network model, and output the road type on which the vehicle is driving.

[0020] The method for collecting the acceleration, speed, and environmental image data of the vehicle during driving on various types of roads in the first step is as follows:

[0021] Accelerometers are installed at both ends of the vehicle suspension to collect the acceleration waveforms of the x, y, and z axes. Gyroscopes are installed at the hubs of two wheels to collect the rotational speed information of the x, y, and z axes. A GPS module is installed on the in-vehicle center console to collect the vehicle driving speed data. A camera is installed inside the vehicle to collect the external environment information;

[0022] Select various types of roads and plan the driving routes. Let the driver drive the vehicle on the planned routes respectively, and record the acceleration, speed, and environmental image data during driving at the same time.

[0023] For the initial population in the fourth step, set its population parameters including size, crossover probability, and mutation probability.

[0024] The termination condition in the fourth step is that the set number of iterations is reached or the fitness function value of the genetic algorithm no longer increases. The fitness function of the genetic algorithm is obtained according to the set genetic algorithm objective function.

[0025] The genetic algorithm objective function in the fourth step is the loss function:

[0026]

[0027] In the formula, L(b) represents the loss function of the b-th chromosome sample, N is the number of samples of the tabular data, y cs is the true probability that the c-th tabular data sample belongs to the s-th type of road, is the predicted probability that the c-th tabular data sample is the s-th type of road, and R is the total number of road types;

[0028] The fitness function of the genetic algorithm is:

[0029]

[0030] Among them, g(b) represents the fitness of the b-th chromosome sample in the genetic algorithm.

[0031] The objective function L of the BP neural network model in the fifth step is:

[0032]

[0033] The learning rate formula in the fifth step is:

[0034]

[0035] Among them, η tis the learning rate for the t-th iteration, η 0 is the initial learning rate, which is determined by experiments, d

[0036] is the attenuation rate, taking values from 0.8 to 0.9, h is the step size, taking values from 500 to 1000;

[0037] By adjusting the learning rate, the weights and thresholds are updated.

[0038] The constraint condition that the objective function value of the BP neural network model in step five is lower than the set value is replaced by the fitness function value of the BP neural network model being higher than the set value. Among them, the fitness function formula of the BP neural network model is:

[0039]

[0040] Among them, N co is the number of table data samples with correct predictions.

[0041] Through the above design scheme, the present invention can bring the following beneficial effects:

[0042] The present invention collects acceleration and speed data of the vehicle under different types of roads by installing accelerometers, gyroscopes, GPS modules, and cameras on the vehicle; determines the model input and constructs a BP neural network; optimizes and improves the neural network model using a genetic algorithm; trains the model and uses the collected data for road classification prediction. The present invention can be widely applied to the field of intelligent transportation technology equipment, especially in the road environment understanding of autonomous driving systems, and has significant technical advantages and application prospects.

[0043] The road classification detection method of the present invention based on sensor-collected data and neural network algorithms has the characteristics of wide applicability and high accuracy compared with other classification methods. Through the implementation of this technology, it can provide accurate road condition data for traffic management departments, help optimize road maintenance decisions, and at the same time improve the driving safety and comfort of drivers. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The following further describes the present invention in conjunction with the drawings and specific embodiments:

[0045] Figure 1 is a schematic diagram of the installation positions of the accelerometer, gyroscope, GPS module, and camera in a road classification detection method of a BP neural network optimized by a genetic algorithm according to the present invention;

[0046] Figure 2 is a flowchart of a road classification detection method of a BP neural network optimized by a genetic algorithm according to the present invention;

[0047] Figure 3It is the flowchart of using genetic algorithm to select the optimal weights and thresholds in a road classification detection method based on a BP neural network optimized by genetic algorithm according to the present invention;

[0048] Figure 4 It is the flowchart of on-site test in an embodiment of a road classification detection method based on a BP neural network optimized by genetic algorithm according to the present invention;

[0049] Figure 5 It is the flowchart of training samples in an embodiment of a road classification detection method based on a BP neural network optimized by genetic algorithm according to the present invention.

[0050] In the figure, 1 - accelerometer, 2 - gyroscope, 3 - GPS module, 4 - camera. Specific implementation mode

[0051] In order to better understand the purpose, structure and function of the present invention, the following further describes in detail a road classification detection method based on a BP neural network optimized by genetic algorithm according to the present invention in combination with the accompanying drawings and embodiments.

[0052] Embodiment:

[0053] A road classification detection method based on a BP neural network optimized by genetic algorithm, as Figure 2 shown, includes the following steps:

[0054] Step 1, obtain data and perform preprocessing:

[0055] 1) As Figure 1 shown, install the accelerometer 1 near the vehicle suspension, and at the same time use a data acquisition device adapted to the accelerometer 1 to collect the analog signal output by the accelerometer 1, and convert it into a digital signal and transmit it to the storage device. Use Kalman filtering for data preprocessing to obtain the x, y, z three-axis acceleration tabular data of the vehicle's driving; install the gyroscope 2 at the wheel hub, cooperate with the data acquisition device, collect the analog signal output by the gyroscope 2, and convert it into a digital signal and transmit it to the storage device. Use low-pass filtering for data smoothing processing to obtain the x, y, z three-axis rotation speed tabular data of the vehicle's driving; install the GPS module 3 on the in-vehicle center console, cooperate with programming, and use moving average filtering for data processing to obtain the driving speed data table; install the camera 4 in the vehicle to shoot videos to record the surrounding environment of the vehicle's driving.

[0056] The driving roads can be divided into multiple types. In this embodiment, they are only divided into three most common types: dirt roads, stone roads and asphalt roads.

[0057] Considering the influence of drivers' driving habits and different vehicle models on the experiment, three different drivers were asked to drive vehicles of three different brands on three different roads, and the acceleration waveforms, three-axis rotational speeds, and speed information during the driving were recorded.

[0058] 2) Set the acquisition sample points and process the image data, that is, merge the obtained three-axis acceleration table data, three-axis rotational speed table data, and driving speed data table according to the sampling time. The last column of the table is the road type in the corresponding image. Determining the road type based on the image data is a prior art and will not be elaborated here.

[0059] The data of vehicles of three different brands driving on three different roads can generate 9 data sets. Each data set contains eight types of data, namely the x, y, and z three-axis accelerations, x, y, and z three-axis rotational speeds, vehicle driving speed, and the corresponding road type at multiple different times.

[0060] 3) Then determine the detection parameters:

[0061] Among the 7 parameters collected in the experiment, not all of them can accurately reflect the changes in the road surface conditions. Observe the real-time changes of the collected acceleration data, rotational speed data, speed data, and the external environment video to compare and select the detection parameters that are more sensitive to the road surface type. In this embodiment, all the characteristic parameters, namely the three-axis acceleration, three-axis rotational speed, and driving speed, a total of 7 parameters, are selected as the input of the BP neural network model, and the road type is the output of the BP neural network model. Furthermore, determine the topological structure of the BP neural network.

[0062] Step 2: Establish a BP neural network model according to the determined topological structure of the BP neural network

[0063] 1) Initialize the BP neural network

[0064] a. Determine the topological structure of the BP neural network model:

[0065] Input: x[n] = [a x , a y , a z , n x , n y , n z , v]; If the selection of the detection parameters in Step 1 3) is reduced, the values in the input sequence will be reduced accordingly; where, a x , a y , a z represent the x, y, and z three-axis accelerations respectively, n x , n y , n z represent the x, y, and z three-axis rotational speeds respectively, and v represents the vehicle driving speed;

[0066] Output: o[3] = [o 1 , o 2 , o 3 , representing the road type; setting the output to [1, 0, 0] represents a dirt road, [0, 1, 0] represents a masonry road, and [0, 0, 1] represents an asphalt road.

[0067] Determine the number of nodes n in the input layer, the number of nodes m in the hidden layer, and the number of hidden layers is L. Then the output layer is the (L + 1)-th layer, and the number of nodes R in the output layer is taken as 3; if there are more than three road types, the number of nodes R in the output layer is adjusted accordingly.

[0068] b. Activation function: Select the ReLU function as the activation function f(x).

[0069] c. Obtain weights and thresholds: The weights between the input layer and the hidden layer, and between the hidden layers are denoted by w ij and the threshold b j , and the weights w j between the hidden layer and the output layer and the threshold b s are initialized with the optimal weights and thresholds calculated by the genetic algorithm in step three below.

[0070] 2) Calculation of the hidden layer and the output layer:

[0071] For the hidden layer H, the j-th neuron in the l-th layer has the following output representation:

[0072]

[0073] where n (l-1) is the number of nodes in the (l - 1)-th layer, is the weight from the i-th neuron in the (l - 1)-th layer to the j-th neuron in the l-th layer, is the threshold of the j-th neuron in the l-th layer, is the output of the i-th neuron in the (l - 1)-th layer, and l ≤ L.

[0074] For the k-th neuron O in the output layer O k , it is represented as follows:

[0075]

[0076] where, is the output of the j-th neuron in the last hidden layer, which is the (L + 1)-th layer.

[0077] 3) Objective function of the BP neural network model:

[0078] Take the loss function L of the BP neural network model as the objective function of the BP neural network model:

[0079]

[0080] Among them, N is the number of samples of the tabular data, and y cs is the true probability that the c-th tabular data sample belongs to the s-th type of road. is the predicted probability that the c-th tabular data sample is the s-th type of road. R is the total number of road types. Since three road types are taken in this embodiment, the loss function is expressed as:

[0081]

[0082] 4) Update the weights and thresholds by adjusting the learning rate;

[0083] Set the dynamic learning rate:

[0084]

[0085] Among them, η t is the learning rate of the t-th iteration, η 0 is the initial learning rate, which is determined by experiments. d is the decay rate, taking 0.8 - 0.9, and h is the step size, taking 500 - 1000.

[0086] The weight update formula is as follows:

[0087]

[0088] Among them is the weight from the i-th neuron in the (l - 1)-th layer to the j-th neuron in the l-th layer at the (t + 1)-th iteration.

[0089] The threshold update formula is as follows:

[0090]

[0091] Among them, is the threshold of the j-th neuron in the l-th layer at the (t + 1)-th iteration.

[0092] 5) Set the end condition;

[0093] When the loss value L predicted by the neural network is less than the set threshold, stop the iteration.

[0094] Step Three: Use the genetic algorithm to select the best initial weights and thresholds of the BP neural network, as Figure 3 shown.

[0095] 1) Determine the coding rule:

[0096] The parameters to be optimized by the genetic algorithm are encoded into chromosome strings according to certain rules, that is, the connection weights and thresholds of the BP neural network are encoded. The weights and thresholds correspond to each other. Since the connection weights and thresholds of the BP neural network are within the interval (0,1), real number encoding is used, and the chromosome length Le is determined by the topological structure of the neural network.

[0097] 2) Select the initial population according to the set objective function, and determine whether it meets the set termination condition. If it meets the condition, proceed to step 4. If it does not meet the condition, perform genetic mutation to form offspring, and then calculate the objective function until the termination condition is met. The optimal initial weights and thresholds are used as the initial weights and thresholds of the BP neural network model to improve the training efficiency of the model.

[0098] The objective function of the genetic algorithm is also set as a loss function. The loss function of the genetic algorithm in this embodiment is expressed as follows:

[0099]

[0100] In the formula, L(b) represents the loss function of the b-th chromosome sample. According to the objective function, the fitness function of the genetic algorithm is determined, that is, the fitness of the b-th sample is:

[0101]

[0102] Among them, generating the initial population is specifically as follows:

[0103] a. Determine the population size Q and the length of the chromosome Le;

[0104] b. Generate Le random numbers s in the interval [0,1] a (a=1、2、3、…、Le) constitute chromosome S b ={s 1 ,s 2 ,s 3 ,…,s Le}(b=1, 2, 3, …, Q).

[0105] c. Repeat step b. Each repetition generates Le random numbers, which is equivalent to generating one chromosome each time. Repeat Q times, and Q chromosomes are generated. The population initialization is now complete. The initial population is S (0) ={S 1 ,S 2 ,…,S b}.

[0106] 3) Determine the optimization end condition of the genetic algorithm:

[0107] Get the maximum number of iterations T and output error μ;

[0108] If the algorithm reaches the maximum number of iterations T, stop the iteration; or if the objective function is less than or equal to the output error μ, stop the iteration.

[0109] If either condition is met, stop the iteration and output the weights and threshold corresponding to the maximum value of the current objective function.

[0110] 4) Genetic variation to form offspring is as follows:

[0111] Use the roulette wheel selection method to screen the population. The currently screened population is the parent generation, and perform crossover and mutation operations:

[0112] a. Calculate the probability P(b) that the b-th chromosome individual is inherited to the next generation:

[0113]

[0114] b. Calculate the cumulative probability Q of the b-th chromosome individual: b :

[0115]

[0116] In the formula, k = 1, 2... b;

[0117] c. Generate a uniformly distributed random number r in the range [0, 1];

[0118] d. If r ≤ Q 1 , then S 1 is selected; if Q b-1 < r ≤ Q b , then S b is selected; S 1 is the first chromosome individual, and S b is the b-th chromosome individual;

[0119] e. Repeat steps c and d, repeat Q times, that is, select Q parent individuals (allowing repetition) from S (0) to form a mating pool;

[0120] f. Perform crossover and mutation operations on the individuals in the mating pool as follows:

[0121] Determine the crossover rate P C , then the number of crossovers Q C is:

[0122] Q C = Q·P C ;

[0123] Among them, P C takes values from 0.6 to 0.95.

[0124] Determine the mutation rate P M, then the number of mutations Q M is:

[0125] Q M = Q · P M ;

[0126] where P M takes values from 0.001 to 0.1.

[0127] Crossover: Randomly select two individuals from the parental population as parental crossover individuals. Randomly select a value E in the interval [1, Le]. Copy the genes from the (E + 1)-th to the Le-th of the first individual to the head of the second individual, and copy the genes from the (E + 1)-th to the Le-th of the second individual to the tail of the first individual to obtain the offspring crossover individuals.

[0128] Mutation: Randomly select an individual from the parental population as the parental mutation individual. Randomly select a value E in the interval [1, Le]. Generate a uniformly distributed pseudo-random number e in the interval [0, 1]. Replace the numerical value represented by the E-th gene of the mutation individual with e to obtain the offspring mutation individual.

[0129] For example: Suppose there are the following individuals in the parental population (taking binary coding as an example, and Le = 5):

[0130] Individual 1: [0, 1, 0, 1, 1];

[0131] Individual 2: [1, 0, 1, 0, 0];

[0132] Individual 3: [0, 0, 1, 1, 0];

[0133] Randomly select Individual 2 as the parental mutation individual.

[0134] Randomly select 3 in the interval [1, 5].

[0135] Generate a uniformly distributed pseudo-random number r = 0.7 in the interval.

[0136] Replace the 3rd gene (value 1) of Individual 2 with 0.7 to obtain the offspring mutation individual: [1, 0, 0.7, 0, 0].

[0137] g. Repeat step f until there are Q offspring samples, and execute step 3) to determine the optimization end condition of the genetic algorithm.

[0138] Step Four. Field test

[0139] Such as Figure 4As shown, the accelerometer 1, gyroscope 2, and camera 4 are installed on the experimental vehicle, and the GPS module 3 is installed on the in-vehicle console. A road is selected where the total length of dirt roads, brick roads, and asphalt roads accounts for more than 80%. The real-time acceleration data, rotational speed data, speed data, and real-time changes in the external environment video and other data are imported into the trained BP neural network model to judge the road surface type and input it into the autonomous driving system to guide the behavior of autonomous driving.

[0140] In addition, the end condition in step 2) 5) can also be set to another one, that is, the fitness function value of the BP neural network model is higher than the set value. The fitness function of the BP neural network model is set artificially and is used to evaluate the prediction performance of the neural network. The fitness function is set according to indicators such as prediction error and accuracy.

[0141] The fitness function of the BP neural network model is:

[0142]

[0143] where N co is the number of correctly predicted tabular data samples. Its training sample flow chart is as Figure 5 shown.

Claims

1. A road classification detection method based on BP neural network optimized by genetic algorithm, characterized by: The process includes the following steps, which are performed in sequence: Step 1: Data Collection The acceleration, three-axis rotation speed, driving speed and environmental image data of the vehicle when driving on various types of roads are collected according to the set sampling interval, and abnormal and invalid data are removed from the collected data to obtain the data after preliminary screening and sorting; Step 2: Data preprocessing Smoothing the acceleration, rotation speed and speed data in the preliminary screened and sorted data to remove noise and interference, obtaining the three-axis acceleration, three-axis rotation speed and driving speed data and merging and storing them in a table form, identifying the type of path in the image through the environmental image data and storing them in correspondence with the table data; Step 3: Build a BP neural network model The BP neural network model has an input layer, a hidden layer and an output layer; select part or all of the three-axis acceleration, the three-axis rotation speed and the driving speed as characteristic parameters, build the BP neural network model according to the extracted characteristic parameters, use the table data of the selected characteristic parameters as the input of the BP neural network model, and use the corresponding road type as the output of the BP neural network model, thereby determining the neural network topology; Step 4: Using genetic algorithm to optimize the initial weights and thresholds of the BP neural network model; According to the topological structure of the BP neural network model, the coding rules are determined, the initial population is selected according to the set genetic algorithm objective function, and it is judged whether the set termination condition is met. If it is met, step five is carried out. If it is not met, genetic mutation is carried out to form offspring, and then the objective function is calculated until the termination condition is met. The optimal initial weights and thresholds are used as the initial weights and thresholds of the BP neural network model. Step 5: Model training: The BP neural network model is trained using the data preprocessed in step 2, and the learning rate, weight and threshold are continuously updated until the set BP neural network model objective function value is lower than the set value, thereby obtaining a trained BP neural network model; Step 6: Road classification prediction: The real-time acceleration, three-axis rotation speed, driving speed and environmental image data collected by the vehicle are input into the trained BP neural network model to output the type of road the vehicle is traveling on.

2. The road classification detection method based on BP neural network optimized by genetic algorithm according to claim 1 is characterized by: The method for collecting acceleration, speed and environmental image data of the vehicle when driving on various types of roads in step 1 is as follows: Accelerometers are installed at both ends of the vehicle suspension to collect x, y, and z three-axis acceleration waveforms. Gyroscopes are installed at the hubs of the two wheels to collect x, y, and z three-axis speed information. A GPS module is installed on the center console in the vehicle to collect vehicle speed data. A camera is installed in the vehicle to collect external environment information. Select various types of roads and plan driving routes, and let the driver drive the vehicle on the planned routes respectively, while recording acceleration, speed and environmental image data during driving.

3. The road classification detection method based on BP neural network optimized by genetic algorithm according to claim 1 is characterized by: The initial population in step 4 sets its population parameters including size, crossover probability and mutation probability.

4. The road classification detection method based on BP neural network optimized by genetic algorithm according to claim 1 is characterized by: The termination condition in step 4 is that the set number of iterations is reached or the fitness function value of the genetic algorithm no longer increases, and the fitness function of the genetic algorithm is obtained according to the set genetic algorithm objective function.

5. A road classification detection method based on BP neural network optimized by genetic algorithm according to claim 1 or 4, characterized in that: The objective function of the genetic algorithm in step 4 is a loss function: Where L(b) represents the loss function of the b-th chromosome sample, N is the number of samples of the table data, and y cs is the true probability that the cth table data sample belongs to the sth type of road, To predict the probability that the cth table data sample is the sth type of road, R is the total number of road types; 6. The road classification detection method based on BP neural network optimized by genetic algorithm according to claim 5 is characterized by: The fitness function of the genetic algorithm is: Among them, g(b) represents the fitness of the b-th chromosome sample in the genetic algorithm.

7. The road classification detection method based on BP neural network optimized by genetic algorithm according to claim 1 is characterized by: The objective function L of the BP neural network model in step 5 is:

8. The road classification detection method based on BP neural network optimized by genetic algorithm according to claim 1 is characterized by: The learning rate formula in step 5 is: Among them, η t is the learning rate of the tth iteration, η0 is the initial learning rate, which is determined by experiments, d is the decay rate, which is 0.8-0.9, and h is the step size, which is 500-1000; Update the weights and thresholds by adjusting the learning rate.

9. The road classification detection method based on BP neural network optimized by genetic algorithm according to claim 1 is characterized by: In step 5, the constraint condition that the objective function value of the BP neural network model is lower than the set value is replaced by the fitness function value of the BP neural network model being higher than the set value, wherein the fitness function formula of the BP neural network model is: Among them, N co is the number of tabular data samples for which predictions are correct.

Images