A neural network road obstacle recognition method based on four-wheel turning radius

Abstract

A neural network road obstacle identification method based on four-wheel turning radius belongs to the technical field of vehicle traffic safety.The present application separates the data segment of vehicle roundabout movement from the total data set when the vehicle is running, and obtains the relative relationship between the turning radius of the vehicle and the inherent relationship between the length of the obstacle through the geometric relationship between the turning radius of the four wheels of the vehicle and the size of the obstacle when the vehicle is roundabout, and the BP neural network optimized by the chaotic mapping adaptive whale algorithm, realizes the accurate prediction of the size of the obstacle, and through the threshold judgment method, the abnormal data is screened, and the false alarm is effectively reduced. The prediction of the size of the obstacle reflects the abnormal situation of the road surface, reduces the traffic safety hidden danger caused by the road obstacle, and has important practical significance. The algorithm structure is clear, the reliability of the result information is high, and the cost is low, and the goal of detecting the road abnormal obstacle can be realized.

Description

Technical Field

[0001] The invention belongs to the technical field of vehicle traffic safety, and in particular relates to a neural network road obstacle recognition method based on four-wheel turning radius. Background Art

[0002] During daily commutes, obstacles on the road can obstruct vision, make the road bumpy, and even distract drivers, leading to traffic jams or accidents. While road authorities quickly address obstacles upon discovery, the lack of timely reporting of obstacles often delays their removal.

[0003] Therefore, a new technical solution is urgently needed in the existing technology to solve this problem. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a neural network road obstacle recognition method based on four-wheel turning radius to solve the technical problem in the prior art that road obstacles cannot be reported in a timely manner, resulting in untimely removal of road obstacles.

[0005] A neural network road obstacle recognition method based on four-wheel turning radius includes the following steps, which are performed in sequence:

[0006] Step 1: Using a driving simulator to simulate the vehicle's movement around obstacles of various sizes on the road, obtaining vehicle motion data when the vehicle moves around the corresponding obstacles and storing the data in a total data set, wherein the vehicle motion data includes vehicle speed, steering wheel angle, yaw angle, left front wheel speed, and right front wheel speed of the vehicle;

[0007] Step 2: Separate the data segments of the initial vehicle's circumnavigation from the total data set from the corresponding vehicle motion data whose steering wheel angles meet the set threshold conditions, and exclude the data segments of the vehicle's normal turning behavior from the data segments of the initial vehicle's circumnavigation to obtain the data segments of the vehicle's circumnavigation;

[0008] Step 3: Obtaining the length-to-width ratio of the obstacle according to a formula for estimating the length-to-width ratio of the obstacle circumvented based on the vehicle speed and steering wheel angle in the data segment of the vehicle's circumventing motion;

[0009] Step 4: Obtain the turning radius of the four wheels of the vehicle during the detour according to the left front wheel speed and the right front wheel speed of the vehicle;

[0010] Step five, the BP neural network optimized by the chaotic mapping adaptive whale optimization algorithm is established, the turning radii of the vehicle are respectively compared to describe the relationship between the turning radii of the four wheels during the turning, and the comparison is used as the input layer data of the training set of the BP neural network optimized by the chaotic mapping adaptive whale optimization algorithm, the length of the set obstacle is used as the output layer data of the training set, the BP neural network optimized by the chaotic mapping adaptive whale optimization algorithm is trained, and the trained BP neural network optimized by the chaotic mapping adaptive whale optimization algorithm is obtained;

[0011] Step six, the vehicle motion data of the vehicle in the actual driving process is obtained through the control system of the vehicle and is transmitted to the trained BP neural network optimized by the chaotic mapping adaptive whale optimization algorithm installed in the control system in real time, and an obstacle length prediction value is obtained;

[0012] The length-width ratio of the obstacle obtained in step three is used, and the obstacle length prediction value is obtained, and an obstacle width prediction value is obtained;

[0013] The obtained obstacle length prediction value and obstacle width prediction value are uploaded to a road management department control terminal in communication connection with the control system of the vehicle;

[0014] Step seven, the road management department control terminal is provided with an obstacle length threshold value and an obstacle width threshold value, when one of the obstacle length prediction value and the obstacle width prediction value received by the road management department control terminal is greater than the corresponding threshold value, the road management department control terminal determines that it is abnormal data, and eliminates the abnormal state of the obstacle being too small due to the random driving parameter fluctuation of the vehicle in the normal driving process;

[0015] The road management department control terminal is also provided with a number threshold value of abnormal data received in a set time interval on the same road section, when the number of abnormal data received in the set time interval on the same road section exceeds the set threshold value, it is determined that the obstacle exists, and the received obstacle length prediction value and obstacle width prediction value are determined as the estimated length of the obstacle and the estimated width of the obstacle.

[0016] The threshold condition in step two is:

[0017] When the vehicle motion data at the i th sampling time is used as the data starting point of the turning motion of the vehicle;

[0018] When the vehicle motion data at the i th sampling time is used as the data ending point of the turning motion of the vehicle;

[0019] In the formula, is the steering wheel angle at the i th sampling time, is the steering wheel angle at the (i-5)th sampling moment, and m is the minimum change angle of the steering wheel angle.

[0020] The specific method of excluding the data segment of the normal turning behavior of the vehicle from the data segment of the initial vehicle circumnavigation in step 2 is:

[0021] When the vehicle is going around an obstacle, the integral of the steering wheel angle over time t is close to 0, while when the vehicle is turning normally, the integral of the steering wheel angle over time t is much greater than zero. Therefore, when a certain data segment in the data segment of the initial vehicle's rounding motion satisfies , it is determined that the data segment is a data segment of vehicle circling motion, where h is the threshold coefficient.

[0022] The formula for estimating the ratio of the length to the width of the obstacle in step 3 is:

[0023]

[0024] In the formula, AR is the ratio of the length to the width of the obstacle, also known as the axis ratio; a is the major axis of the ellipse; b is the minor axis of the ellipse; θ i is the yaw angle at time i.

[0025] The turning radius of the four wheels in step 4 are:

[0026]

[0027]

[0028] R k =R+B

[0029] Among them, R is the turning radius of the inner rear wheel of the vehicle, R i is the turning radius of the front wheel of the vehicle, R j is the turning radius of the vehicle's outer front wheel, R k is the turning radius of the vehicle's outer rear wheel, B is the distance between the left and right wheels, L is the vehicle's wheelbase, and k is the ratio of the distance traveled by the left and right front wheels.

[0030] Through the above design scheme, the present invention can bring the following beneficial effects:

[0031] The present invention separates the data segments of the vehicle's detour from the total data set during vehicle travel, and obtains the inherent relationship between the relative relationship between the vehicle's turning radius and the obstacle length through the geometric relationship between the vehicle's four-wheel turning radius and the obstacle size during the detour, as well as the BP neural network optimized by the chaotic mapping adaptive whale algorithm. This achieves accurate prediction of the obstacle size, and screens abnormal data through a threshold judgment method, effectively reducing the occurrence of false alarms. By predicting the size of obstacles, the abnormal conditions of the road surface are reflected, and the traffic safety hazards caused by road obstacles are reduced, which has relatively important practical significance. The algorithm has a clear structure, and the credibility of the result information is high and the cost is low, which can achieve the goal of detecting abnormal road obstacles.

[0032] The present invention can effectively eliminate interference caused by normal vehicle turning behavior and abnormal conditions where obstacles are too small due to random fluctuations in vehicle driving parameters during normal driving. It can effectively solve the problems of high data acquisition consumption, low acquisition efficiency and susceptibility to environmental changes in traditional road obstacle detection methods.

[0033] Furthermore, the present invention utilizes a BP neural network optimized by a chaotic map adaptive whale algorithm as a model for determining obstacle length and width, achieving high accuracy. Through big data analysis, data that meets threshold conditions is used as the final basis for judgment, effectively improving accuracy. Furthermore, all vehicles are connected to the road management department's control terminal, enabling 24 / 7 road monitoring. Furthermore, data is easily accessible and widely sourced, making it highly feasible. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0035] Figure 1 This is a schematic diagram of the wheel turning radius in a neural network road obstacle recognition method based on four-wheel turning radius of the present invention. DETAILED DESCRIPTION

[0036] A neural network road obstacle recognition method based on four-wheel turning radius uses simulation software to simulate the actual situation of a vehicle bypassing an obstacle to obtain the required vehicle speed, steering wheel angle, yaw angle, left front wheel speed, and right front wheel speed. The data of each vehicle bypassing the obstacle are separated from the total data set, and the four-wheel turning radius of the vehicle during the bypass, the geometric relationship between the turning radius and the obstacle size, and the ratio between the vehicle's steering angle and the turning radius are calculated. The obstacle size is predicted using a BP neural network optimized by a chaotic mapping adaptive whale algorithm, and the abnormal data obtained from each vehicle is regarded as an abnormal sample. When the total amount of abnormal samples in a certain road section within a certain period of time exceeds a threshold, it is ultimately determined to be an obstacle.

[0037] The specific steps are as follows:

[0038] Step 1: Use a driving simulator to simulate the situation of the vehicle bypassing obstacles of various sizes on the road, obtain the vehicle motion data when the vehicle bypasses the corresponding obstacle and store it in the total data set. The vehicle motion data includes vehicle speed, steering wheel angle, yaw angle, left front wheel speed of the vehicle, and right front wheel speed of the vehicle.

[0039] Step 2: Separate the corresponding vehicle motion data whose steering wheel angle meets the set threshold condition from the total data set to obtain the data segment of the initial vehicle's circumventing motion, exclude the data segment of the vehicle's normal turning behavior from the data segment of the initial vehicle's circumventing motion, and obtain the data segment of the vehicle's circumventing motion.

[0040] The data collected when the vehicle is driving contains a lot of meaningless data, so the data segments of the vehicle's orbital motion must be separated from it. The vehicle's orbital motion is directly related to the steering wheel angle, so the threshold conditions set are:

[0041] When , the vehicle motion data at the i-th sampling moment is used as the data starting point of the vehicle's circumnavigation motion;

[0042] When , the vehicle motion data at the i-th sampling moment is used as the data end point of the vehicle's circumnavigation motion;

[0043] In the formula, is the steering wheel angle at the i-th sampling moment, is the steering wheel angle at the (i-5)th sampling moment, and m is the minimum change angle of the steering wheel angle.

[0044] The details of eliminating interference from the data segment of normal vehicle turning behavior are as follows:

[0045] The data segment of the initial vehicle's detour movement includes not only the data of the vehicle detouring around obstacles, but also the data of the vehicle turning normally. Therefore, the influence of the vehicle's normal turning data must be excluded. Let h be the threshold coefficient. When the vehicle detours around obstacles, the integral of the steering wheel angle over time t is close to 0, while when the vehicle turns normally, the integral of the steering wheel angle over time t is much greater than zero. Therefore, , it is determined that the data segment is a data segment of vehicle circling motion.

[0046] Step 3: Obtain the length-to-width ratio of the obstacle using a formula for estimating the length-to-width ratio of the obstacle circumvented based on the vehicle speed and steering wheel angle in the data segment of the vehicle's circumventing motion.

[0047] When a vehicle circumvents an obstacle, its trajectory is similar to an ellipse. The size of the ellipse is closely related to the size of the obstacle. That is, the axis ratio of the ellipse is approximately the same as the length and width of the obstacle being circumvented. Therefore, the axis ratio is set to AR, which is the length and width of the obstacle. The major axis of the ellipse is a, the minor axis of the ellipse is b, and the lateral displacement of the vehicle is x. a , the longitudinal displacement of the vehicle is x b , the velocity of the center of mass at time i is v i , the yaw angle at time i is θ i but:

[0048]

[0049] The lateral displacement of the vehicle is x a and the vehicle longitudinal displacement is x b The velocity of the center of mass at time i is v i , the yaw angle at time i is θ i Obtain:

[0050]

[0051]

[0052] Then AR can be simplified as:

[0053]

[0054] Step 4: Obtain the turning radius of the four wheels of the vehicle during the detour according to the left front wheel speed and the right front wheel speed of the vehicle.

[0055] The wheel radius diagram of a vehicle turning is as follows Figure 1 As shown. The vehicle's orbit is an ellipse, but in a very short time, the vehicle's orbit can be approximated as a perfect circle. The distances traveled by the two front wheels are:

[0056] S i =R i *θ

[0057] S j =R j *θ

[0058] Among them S i is the distance traveled by the front wheels of the vehicle, R i is the turning radius of the front wheel of the vehicle, and θ is the steering angle. j is the distance traveled by the outer front wheel of the vehicle, R j is the turning radius of the vehicle's outer front wheel.

[0059] R is the turning radius of the vehicle's rear wheels, B is the distance between the left and right wheels, and L is the vehicle's wheelbase, so:

[0060]

[0061] k is the ratio of the distances traveled by the left and right front wheels, then:

[0062]

[0063] The turning radius R of the inner rear wheel of the vehicle can be extracted by transformation:

[0064]

[0065] At this time, the turning radius of the inner rear wheel of the vehicle can be calculated based on the ratio of the distances traveled by the two front wheels.

[0066] The turning radius of the vehicle's four wheels can be calculated based on the turning radius of the vehicle's inner rear wheel:

[0067]

[0068] R=R,R k =R+B

[0069] Among them, R k is the turning radius of the vehicle's inner rear wheel.

[0070] The ratio between the turning radii is:

[0071]

[0072] Step 5. Establish a BP neural network optimized by the chaotic mapping adaptive whale algorithm. Compare the turning radii of the vehicles to describe the relationship between the turning radii of the four wheels when detouring. Use the comparison as the input layer data of the training set of the BP neural network optimized by the chaotic mapping adaptive whale algorithm. Use the length of the obstacle as the output layer data of the training set to train the BP neural network optimized by the chaotic mapping adaptive whale algorithm to obtain the trained BP neural network optimized by the chaotic mapping adaptive whale algorithm.

[0073] Since traditional BP neural networks are prone to falling into local minima, the use of the whale algorithm to optimize BP neural networks can effectively improve this problem. At the same time, we have also made some optimizations to the traditional whale algorithm.

[0074] The position vector is initialized using the Cubic map chaotic mapping sequence. The chaotic sequence formula is as follows:

[0075] leader_pos=ρ*(1-leader_pos1^2);

[0076] leader_pos is the position vector matrix for initializing the chaotic map, leader_pos1 is a random number matrix from 0 to 1, and ρ is the adjustment coefficient.

[0077] Traditional whale optimization algorithms use pseudo-random numbers to initialize position vectors. However, the resulting position vectors cannot effectively simulate the random distribution of humpback whales. To address this issue, the present invention uses chaotic sequences, commonly used in the communications field, instead of pseudo-random numbers to generate initial position vectors. This has been shown to achieve better results than using pseudo-random numbers.

[0078] Set the adaptive weight coefficient. The formula is as follows:

[0079] w=w min +(w max -w min )*mm*e (-t / maxgen)

[0080] Among them, w is the adaptive weight coefficient, mm is the adjustment coefficient, w min is the minimum coefficient of adaptive weight, w max is the maximum coefficient of adaptive weight, w min and w max Can be adjusted by yourself, e (-t / maxgen) Where t is the evolution generation, maxgen is the maximum evolution generation, and e is a mathematical constant. min =0,w max =1.

[0081] As the number of generations increases, the adaptive coefficient also increases exponentially, which can give the whale a higher exploration ability, so that the algorithm can obtain a more refined optimal solution.

[0082] The whale algorithm is used to optimize the weights from the input layer to the middle layer, the thresholds of the neurons in the middle layers, the weights from the middle layer to the output layer, and the neurons in the output layer. That is, the chaotic map adaptive whale optimization algorithm is used to optimize the overall structure of the neural network.

[0083] For example, 120 sets of vehicle turning radius ratios are used as training set input data, obstacle lengths are used as training set output data, the number of training times is set to 1000, the learning rate is 0.01, the minimum error of the training target is set to 0.0001, and the maximum number of failures is set to 6. The BP neural network optimized by the chaotic map adaptive whale algorithm is trained.

[0084] After training the BP neural network optimized by the chaotic mapping adaptive whale algorithm, the accuracy of the BP neural network optimized by the chaotic mapping adaptive whale algorithm was tested using the test set data. The result showed that the error was within an acceptable range, and the training of the BP neural network optimized by the chaotic mapping adaptive whale algorithm was completed.

[0085] By transmitting the data of the vehicle's actual driving process to the BP neural network optimized by the trained chaotic mapping adaptive whale algorithm, the obstacle length can be predicted.

[0086] The AR obtained in step 3 can be used to obtain the width of the obstacle. By setting the obstacle length threshold and obstacle width threshold, we can eliminate abnormal conditions where the obstacle is too small due to random driving parameter fluctuations during normal driving, thereby avoiding false alarms and other problems.

[0087] Step 6: The vehicle motion data of the vehicle during actual driving is obtained through the vehicle control system and transmitted in real time to the BP neural network optimized by the trained chaotic mapping adaptive whale algorithm installed in the control system to obtain the obstacle length prediction value;

[0088] Obtain the predicted obstacle width using the ratio of the obstacle length to the obstacle width obtained in step 3 and the predicted obstacle length.

[0089] Uploading the obtained obstacle length prediction value and obstacle width prediction value to a road management department control terminal that is communicatively connected to the vehicle control system;

[0090] Step 7: The road management department control terminal is set with an obstacle length threshold and an obstacle width threshold. When either the obstacle length prediction value or the obstacle width prediction value received by the road management department control terminal is greater than the corresponding threshold, the road management department control terminal determines that it is abnormal data, and eliminates abnormal states of obstacles that are too small due to random driving parameter fluctuations during normal driving.

[0091] The road management department control terminal is also provided with a threshold value for the number of abnormal data received on the same road section within a set time interval. When the number of abnormal data received on the same road section within the set time interval exceeds the set threshold value, it is determined that an obstacle exists, and the received obstacle length prediction value and obstacle width prediction value are determined as the estimated length and estimated width of the obstacle.

[0092] The method of the present invention innovatively considers the relationship between the vehicle's detour trajectory and the size of road obstacles, as well as the relationship between the change in the four-wheel turning radius of the vehicle when detouring and the obstacle, and predicts the size of the obstacle through a neural network. This effectively compensates for the current situation in which existing road obstacle detection methods only focus on extracting and processing obstacle information while ignoring the vehicle's own motion data. At the same time, the present invention also has the characteristics of low data acquisition cost, high acquisition efficiency and low environmental impact. The final multi-sample judgment also effectively improves the accuracy rate. These are of great practical significance and can effectively detect road problems and reduce traffic safety hazards caused by road abnormalities.

Claims

1. A neural network road obstacle recognition method based on four-wheel turning radius, characterized by: The process includes the following steps, which are performed in sequence: Step 1: Using a driving simulator to simulate the vehicle's movement around obstacles of various sizes on the road, obtaining vehicle motion data when the vehicle moves around the corresponding obstacles and storing the data in a total data set, wherein the vehicle motion data includes vehicle speed, steering wheel angle, yaw angle, left front wheel speed, and right front wheel speed of the vehicle; Step 2: Separate the data segments of the initial vehicle's circumnavigation from the total data set from the corresponding vehicle motion data whose steering wheel angles meet the set threshold conditions, and exclude the data segments of the vehicle's normal turning behavior from the data segments of the initial vehicle's circumnavigation to obtain the data segments of the vehicle's circumnavigation; Step 3: Obtaining the length-to-width ratio of the obstacle according to a formula for estimating the length-to-width ratio of the obstacle circumvented based on the vehicle speed and steering wheel angle in the data segment of the vehicle's circumventing motion; Step 4: Obtain the turning radius of the four wheels of the vehicle during the detour according to the left front wheel speed and the right front wheel speed of the vehicle; Step 5: Establish a BP neural network optimized by the chaotic mapping adaptive whale algorithm. Compare the turning radii of the vehicles to describe the relationship between the turning radii of the four wheels during detours. The comparison is used as the input layer data of the training set of the BP neural network optimized by the chaotic mapping adaptive whale algorithm. The length of the obstacle is used as the output layer data of the training set. The BP neural network optimized by the chaotic mapping adaptive whale algorithm is trained to obtain the trained BP neural network optimized by the chaotic mapping adaptive whale algorithm. Step 6: The vehicle motion data of the vehicle during actual driving is obtained through the vehicle control system and transmitted in real time to the BP neural network optimized by the trained chaotic mapping adaptive whale algorithm installed in the control system to obtain the obstacle length prediction value; Obtain the predicted obstacle width using the ratio of the obstacle length to the obstacle width obtained in step 3 and the predicted obstacle length. Uploading the obtained obstacle length prediction value and obstacle width prediction value to a road management department control terminal that is communicatively connected to the vehicle control system; Step 7: The road management department control terminal is set with an obstacle length threshold and an obstacle width threshold. When either the obstacle length prediction value or the obstacle width prediction value received by the road management department control terminal is greater than the corresponding threshold, the road management department control terminal determines that it is abnormal data, and eliminates abnormal states of obstacles that are too small due to random driving parameter fluctuations during normal driving. The road management department control terminal is also provided with a threshold value for the number of abnormal data received on the same road section within a set time interval. When the number of abnormal data received on the same road section within the set time interval exceeds the set threshold value, it is determined that an obstacle exists, and the received obstacle length prediction value and obstacle width prediction value are determined as the estimated length and estimated width of the obstacle.

2. The neural network road obstacle recognition method based on four-wheel turning radius according to claim 1 is characterized by: The threshold condition in step 2 is: When , the vehicle motion data at the i-th sampling moment is used as the data starting point of the vehicle's circumnavigation motion; When , the vehicle motion data at the i-th sampling moment is used as the data end point of the vehicle's circumnavigation motion; In the formula, is the steering wheel angle at the i-th sampling moment, is the steering wheel angle at the (i-5)th sampling moment, and m is the minimum change angle of the steering wheel angle.

3. The neural network road obstacle recognition method based on four-wheel turning radius according to claim 1 is characterized by: The specific method of excluding the data segment of the normal turning behavior of the vehicle from the data segment of the initial vehicle circumnavigation in step 2 is: When the vehicle is going around an obstacle, the integral of the steering wheel angle over time t is close to 0, while when the vehicle is turning normally, the integral of the steering wheel angle over time t is much greater than zero. Therefore, when a certain data segment in the data segment of the initial vehicle's going around motion satisfies , it is determined that the data segment is a data segment of vehicle circling motion, where h is the threshold coefficient.

4. The neural network road obstacle recognition method based on four-wheel turning radius according to claim 1 is characterized by: The formula for estimating the ratio of the length to the width of the obstacle in step 3 is: In the formula, AR is the ratio of the length to the width of the obstacle, also known as the axis ratio; a is the major axis of the ellipse; b is the minor axis of the ellipse; θ i is the yaw angle at time i.

5. The neural network road obstacle recognition method based on four-wheel turning radius according to claim 1 is characterized by: The turning radius of the four wheels in step 4 are: R k =R+B Among them, R is the turning radius of the inner rear wheel of the vehicle, R i is the turning radius of the front wheel of the vehicle, R j is the turning radius of the vehicle's outer front wheel, R k is the turning radius of the vehicle's outer rear wheel, B is the distance between the left and right wheels, L is the vehicle's wheelbase, and k is the ratio of the distances traveled by the left and right front wheels.

Images