Ultrasonic distance measurement alarm method and system for automobile door opening anti-collision

Abstract

This invention discloses an ultrasonic ranging alarm method and system for preventing collisions when a car door is opened. The alarm method relies on the position information of moving objects around the vehicle collected before the door is opened. It then uses a seven-layer deep learning neural network provided by this invention to predict the final position of these moving objects, determining whether opening the door will cause a dangerous situation. If there is a danger, the vehicle control system prevents the door from being opened, thus avoiding a door collision. If there is no danger, the vehicle control system allows the door to be opened. The alarm system includes hardware control modules, information storage modules, deep learning modules, and information sampling modules required to support this invention's ultrasonic ranging alarm method, effectively avoiding door collisions on the road without relying on human intervention.

Description

Technical Field

[0001] The invention relates to the technical field of ultrasonic measurement and traffic safety, and in particular to an ultrasonic distance measurement alarm method for automobile door opening anti-collision. Background Art

[0002] With the development of economy and the progress of society, the number of motor vehicles in my country is increasing. By the end of June 2024, the number of cars in China will reach 345 million, and the number of new energy vehicles will reach 24.72 million. These data show that with the rapid development of my country's economy, people's commuting has become more convenient. At the same time, with the increase in vehicles, traffic safety management and related traffic hazards have become more prominent. For example, the problem of door opening collision is caused by the fact that the passengers on the car do not carefully observe whether there are pedestrians or vehicles behind them when opening the door, and rashly open the door, resulting in pedestrians or vehicles passing by without time to react, causing collisions, and even causing casualties in serious cases. In recent years, with the increase in the number of vehicles, the number of door opening collisions caused by blind spots and omissions of the passengers on the car has increased, and the problem of door opening collision has become a difficult problem that needs to be solved in traffic management.

[0003] The present invention provides an ultrasonic ranging alarm method and system for preventing collision when opening a car door, aiming to solve the door opening collision problem from the vehicle control system end by adding an artificial intelligence method and system, without relying on the reaction and judgment of the people in the car. Summary of the invention

[0004] The purpose of the present invention is to solve the technical problems existing in the prior art and to provide an ultrasonic ranging alarm method and system for preventing collision when a car door is opened.

[0005] To achieve the above object, the technical solution provided by the present invention is: an ultrasonic ranging alarm method for automobile door opening anti-collision, the alarm method comprising the following steps: the step of obtaining the speed average position information matrix is ​​to first calculate the average speed according to the extracted position information of the moving object, then form a speed position information matrix with the position information and the average speed of each target at different times, and then further group them according to the speed, extract part of the data from the speed position information matrix according to the moving speed of the target to form a speed average position information matrix;

[0006] S1, training data collection stage: define the prediction area, determine the position of the moving object entering the prediction area through the ultrasonic wave of the automobile ultrasonic detection device, sample according to the moving speed-time interval, extract the position information of the moving object within the extraction time interval, and form a speed-averaged position information matrix;

[0007] S2, training data set determination stage: randomly extracting a training position information set and a training real position information set from the speed average position information matrix;

[0008] S3, test data set determination stage: delete the training position information set L from the speed average position information matrix tr-c-v According to the information of the corresponding moving object, a test position information set and a test real position information set are established;

[0009] S4, deep learning network establishment stage: according to the application scenario, a seven-layer deep learning network with good generalization ability and strong robustness suitable for mobile location prediction is established;

[0010] S5, deep learning network training stage: the training position information set in S2 is split according to the two-dimensional coordinates x and y axes, and is split into a training position x-axis information set and a training position y-axis information set; the training position x-axis information set and the training position y-axis information set are used to train the seven-layer deep learning network for mobile position prediction, and the positioning average accuracy threshold and the positioning maximum error threshold are set. When the prediction accuracy of the trained network meets the standard, the network training is stopped, and the corresponding hyperparameters and weight parameter information of the seven-layer deep learning network for mobile position prediction are output;

[0011] S6, mobile location prediction network testing phase: the test location information set in S3 is split according to the two-dimensional coordinates x and y axes, and is divided into a test location x-axis information set and a test location y-axis information set; the trained seven-layer deep learning network prediction positioning accuracy is tested by the test location x-axis information set, the test location y-axis information set and the test real location information set to see if it meets the standard;

[0012] S7, actual moving object position prediction stage: two modes are set: low-energy automobile ultrasonic ranging mode and low-latency automobile ultrasonic ranging mode. According to different choices of the occupants, the automobile ultrasonic detection device determines the position of the moving object entering the prediction area, extracts the online position information set of all moving objects in the automobile prediction area, substitutes it into the trained seven-layer deep learning network, and outputs the online prediction position information set of all moving objects;

[0013] S8, door opening collision safety control stage: set the car door opening collision danger zone, compare the final positions of all moving objects in the online predicted position information set of moving objects entering the prediction area with the car door opening collision danger zone, if there is a moving object in the car door opening collision danger zone within the predicted time length of door opening, it is determined to be a dangerous state, and the behavior of opening the door is prevented through the vehicle control system, thereby avoiding the occurrence of door opening collision traffic safety problems; if all moving objects are not in the car door opening collision danger zone within the predicted time length of door opening, it is determined to be a safe state, and the behavior of opening the door is allowed through the vehicle control system.

[0014] Preferably, the specific contents of the training data collection stage in S1 include: in the training data collection stage, the size of the prediction area is defined according to the extraction time interval, and the car extraction time interval is denoted as T a ,The extraction time interval includes the sampling time length and the prediction time length, which are denoted as T c , T p , there is the following relationship:

[0015] T=T+T

[0016] acp;

[0017] Then, the radius d of the prediction area defined by the car a for:

[0018] d a =(T a *V max )+d adj ;

[0019] Among them, V max is the preset maximum moving object speed, d adj is the preset environment distance, and the prediction area defined by the car is d a is the area within a circle of radius;

[0020] The position of the moving object is sampled using the automobile ultrasonic ranging device according to the moving speed and sampling time length within the extraction time interval: the sampling time interval is t, then the two-dimensional position coordinates of the moving object entering the designated area and the average speed of the moving object are extracted according to the sampling time length within the extraction time interval to form a speed position information matrix, denoted as L V :

[0021]

[0022] There are s moving objects entering the prediction area within the extraction time interval; the two-dimensional position information of each moving object is extracted with the sampling time interval t as the sampling period; (x,c) represents the two-dimensional position coordinates of the x-th moving object at the c-th time interval; x is 1, 2, 3...s; c is Sampling position information for each moving object Group, It represents the average moving speed of the moving object s, and the calculation formula is as follows:

[0023]

[0024] From the velocity position information matrix L V The velocity-averaged position information matrix L is further extracted according to the average moving speed of the moving object. ave-V ;

[0025] First, the average moving speed of the moving object is divided into four groups: Group 1, the actual average moving speed is not more than 8m / s; Group 2, the actual average moving speed is not more than 12m / s; Group 3, the actual average moving speed is not more than 20m / s; Group 4, the actual average speed is more than 20m / s; Then, according to the average speed of the moving object, the position speed information is extracted from the speed position information matrix with equal data volume to form the corresponding speed average position information matrix L ave-V :

[0026]

[0027] in, represents the average speed of the c1th moving object in group a1; It represents the two-dimensional position information of the c1th moving object in group a1 in the b1th sampling period; p represents the two-dimensional position information path of p moving objects in each group; by extracting the speed-averaged position information matrix and sampling the motion trajectories of objects with different moving speeds with equal priority, the comprehensiveness and reliability of the data are increased, and the robustness of the network after training is increased.

[0028] Preferably, the method of randomly extracting the training position information set and the training real position information set in S2 includes: c , from the velocity average position information matrix L ave-V Extract the training position information set, denoted as L tr-c-v ,but:

[0029]

[0030] in, represents the average speed of the c2th moving object randomly selected from group a2; p′ represents the two-dimensional position information path of p′ moving objects randomly selected from each group; the speed and position information of 4p′ moving objects are randomly selected from four groups with equal probability; the position information matrix L is divided from the speed according to the corresponding position of the moving object in the training position information set ave-V Extract the training real position information set, denoted as L tr-real-v ,but:

[0031]

[0032] Training real location information set L tr-real-v It represents the average speed and final position sampling information of the moving object corresponding to the training position information set; in order to facilitate the input of the subsequent training stage and the needs of the deep learning network, the training position information set L tr-c-vSplit the data according to the two-dimensional coordinates x and y axes: split into the training position x-axis information set Lx tr-c-v , training position y-axis information set Ly tr-c-v ;

[0033]

[0034] in, It represents the two-dimensional position information of the k1th moving object in group i1 in the j1th sampling period, and the corresponding x and y represent the corresponding horizontal coordinate information and vertical coordinate information.

[0035] Preferably, the method of stripping the predicted position information set, establishing the test position information set and predicting the real position information set in S3 includes: c , delete the training position information set L from the speed average position information matrix tr-c-v The corresponding moving object information in the training position information set is extracted, and one tenth of the moving object data in the total training position information set is used to form the corresponding test position information set, which is denoted as L test-c-v ,but:

[0036]

[0037] p" = p' / 10;

[0038] in, represents the average speed of the p'th moving object randomly selected from group 1. Each group extracts the two-dimensional position information path of p'th moving objects respectively. The four groups randomly extract the speed and position information of 4p'th moving objects with equal probability. According to the corresponding position of the moving object in the test position information set, the position information matrix L is divided from the speed average. ave-V Extract the real test location information set, denoted as L test-real-v ,but:

[0039]

[0040] Test real location information set L test-real-v It represents the average speed and final position sampling information of the moving object corresponding to the test position information set. In order to facilitate the input of the subsequent test phase and the needs of the deep learning network, the test position information set L test-c-v Split the data according to the two-dimensional coordinate xy axis into the test position x-axis information set Lx test-c-v , test position y-axis information set Ly test-c-v ;

[0041]

[0042]

[0043] in, It represents the two-dimensional position information of the k2th moving object in group i2 in the j2th sampling period, and the corresponding x and y represent the corresponding horizontal coordinate information and vertical coordinate information.

[0044] Preferably, the seven-layer deep learning network in the deep learning network establishment stage in S4: the first layer of the network is the input layer, and the input is two The matrices represent the positions of the x-axis and y-axis respectively. They are regarded as two channels, so the input shape is:

[0045]

[0046] Among them, C = 2 is the number of channels, representing the x-axis and y-axis position matrix, H = 4 is the height, representing the number of rows, is the width, indicating the number of columns;

[0047] The second layer of the network is the first convolutional layer. The convolutional layer uses a 3(T-7)×3(T-7) convolution kernel, a step size of 1, and a padding of 0. The output shape is: Conv1 output =(16,2,8);

[0048] The third layer of the network is the first pooling layer. The pooling layer uses a 2×2 maximum pooling with a step size of 2. The output shape is: Pool1 output =(16,1,4);

[0049] The fourth layer of the network is the second convolutional layer. The second convolutional layer uses 32 3×3 convolution kernels, with a step size of 1 and a padding of 1. The output shape is: Conv2 output =(32,1,4);

[0050] The fifth layer of the network is a flattening operation. Before entering the fully connected layer, the output of the convolutional layer needs to be flattened into a one-dimensional vector. The flattening operation turns (32, 1, 4) into 128 neurons.

[0051] The sixth layer of the network is the first fully connected layer, which has 128 neurons as input and 64 neurons as output;

[0052] The seventh layer of the network is the second fully connected layer, which is also the final output layer. The second fully connected layer has 64 neurons as input and 2 neurons as output, representing the two-dimensional coordinates (x, y) predicted by the final network.

[0053] Preferably, the deep learning network training stage in S5 includes: inputting the training position x-axis information set and the training position y-axis information set into the seven-layer deep learning network, training in batches, setting the minimum batch size to 100, extracting data by random sampling, and setting the average positioning accuracy threshold to and the maximum positioning error threshold L er-max The output two-dimensional coordinates are compared with the corresponding position information in the training real position information set. The loss function of the network is set to the mean square error function to measure the difference between the predicted coordinates and the real coordinates. The optimizer selected is the Adam optimizer, and the learning rate is set to 0.002 to minimize the loss function to update the network parameters and determine the optimal seven-layer deep learning network. When the average positioning accuracy and the maximum positioning error are both less than the threshold, the network training is stopped, and the deep learning network, the weights and hyperparameters in the network are output and saved.

[0054] Preferably, the mobile position prediction network test phase in S6 includes: substituting the test position x-axis information set and the test position y-axis information set as the input of the trained seven-layer deep learning network, and the output data is the final position prediction coordinates of the mobile object, and is compared with the corresponding real position information coordinates in the obtained test real position information set to obtain the test positioning average error value, which is recorded as L test-err :

[0055]

[0056] Among them, (x i ',y i ') represents the predicted position coordinates, (x i ,y i ) represents the real position information coordinates, |(x i ,y i ),(x i ',y i ')| represents the Euclidean distance between two points, A represents the total number of moving objects in the test real position information set; the average test positioning error value L test-err And the average positioning accuracy threshold set For comparison, if it is less than the average positioning accuracy threshold, the network meets the standard. If it is greater than or equal to the average positioning accuracy threshold, S6 needs to be repeated to train the network until the average positioning error value L is tested. test-err Less than the set positioning average accuracy threshold

[0057] Preferably, in S7, when the energy efficiency of the car is good, the low-latency car ultrasonic ranging mode is selected: the low-latency car ultrasonic ranging mode focuses on the system being able to quickly make a judgment on whether to open the door or not, and requires the car ultrasonic device to collect the position information data set of the moving object entering the prediction area in real time. The car extraction time interval is T a ,The extraction time interval consists of two parts: the sampling time length T c and the prediction time length T p , extract the two-dimensional position information of the moving object within the sampling time length with the sampling time interval t as the sampling period. When the occupant makes a door opening request, the online position information set of the moving object entering the prediction area is immediately input into the trained seven-layer deep learning network for mobile position prediction, and the online prediction position information set L of all moving objects is output. pred-online , which is expressed as follows:

[0058] L pred-online =[l1 … l i … l n ];

[0059] Where n represents the total number of moving objects entering the prediction area; l n The two-dimensional coordinate information of the predicted position of the nth moving object entering the prediction area; the reaction delay is the calculation time of the seven-layer deep learning network, and the car can be predicted within the prediction time length, that is, T p After the time has passed, the door is opened or a danger alarm is issued;

[0060] In the case of poor energy efficiency of the car, choose the low-energy car ultrasonic ranging mode: the low-energy car ultrasonic ranging mode focuses on the low energy consumption of the system. When the occupants make a door opening request, the car ultrasonic device starts to collect the position information data set of the moving objects entering the prediction area. The car extraction time interval is T a ,The extraction time interval consists of two parts: the sampling time length T c and the prediction time length T p , extract the two-dimensional position information of the moving object within the sampling time length with the sampling time interval t as the sampling period, input the collected online position information set of the moving object into the trained seven-layer deep learning network for mobile position prediction, and output the online predicted position information set L of all moving objects pred-online , the reaction delay is the sum of the computation time and sampling time length of the seven-layer deep learning network. The car can extract the time interval, i.e., T a After the time has passed, the door is opened or a danger alarm is issued.

[0061] Preferably, the door opening collision danger zone in S8 includes: setting the door opening collision danger zone according to the size of the car door, where the door size is denoted as dcar , the environmental state parameter is denoted as d adj ; The corresponding door opening collision danger zone is the 90-degree fan-shaped area corresponding to the two doors, and the radius of the fan-shaped area is D dgr :

[0062]

[0063] The online prediction location information set L pred-online The predicted position information of the moving objects in the vehicle is compared with the door opening collision danger zone. If there is a moving object in the door opening collision danger zone within the predicted door opening time, it is determined to be a dangerous state. The vehicle control system prevents the door from being opened and issues a danger alarm to avoid the door opening collision traffic safety problem. If all moving objects are not in the door opening collision danger zone within the predicted door opening time, it is determined to be a safe state, and the vehicle control system allows the door to be opened.

[0064] The present invention also discloses an alarm system of the ultrasonic ranging alarm method for preventing collision of automobile door opening as described above, wherein the alarm system comprises an on-board system module, a danger alarm, an ultrasonic ranging module, an ultrasonic ranging sampling module, and a door opening trigger judgment module; the on-board system module comprises a deep learning module, an information storage module, and a central control module;

[0065] When the central control module receives the mode information, the position sampling information of the moving object entering the predicted area is stored in the information storage module through the ultrasonic ranging module and the ultrasonic ranging sampling module. After the central control module receives the door opening request from the occupant of the vehicle, the position sampling information of the moving object is extracted and transmitted to the deep learning module. The deep learning module extracts the deep learning network and the corresponding network weights and hyperparameter values ​​from the information storage module to obtain the corresponding predicted position information. The predicted position information and the door opening collision danger zone are judged through the door opening trigger judgment module, and the door opening command or danger alarm command is output to the danger alarm.

[0066] Beneficial effects of the present invention:

[0067] The present invention realizes high-precision monitoring of moving objects through ultrasonic ranging and deep learning technology. In the training stage, the speed-averaged position information matrix is ​​used to enhance the comprehensiveness and reliability of the data, thereby improving the robustness of the network. In the testing stage, a seven-layer deep learning network with strong generalization ability is established to ensure the positioning accuracy and stability of the system in complex environments. Finally, the system significantly improves driving safety by real-time monitoring of moving objects and judging the risk of door opening through a seven-layer deep learning network, effectively eliminating the traffic safety problem of door opening collision and ensuring the safety of the vehicle and passengers. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The drawings described herein are used to provide further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0069] Figure 1 It is a schematic diagram of the steps of the present invention;

[0070] Figure 2 It is a schematic diagram of the automobile area in the present invention;

[0071] Figure 3 is a schematic diagram of a seven-layer deep learning network for mobile position prediction in the present invention;

[0072] Figure 4 It is a schematic diagram of the terminal structure of the present invention. DETAILED DESCRIPTION

[0073] This section will describe in detail the specific embodiments of the present invention. The preferred embodiments of the present invention are shown in the accompanying drawings. The purpose of the accompanying drawings is to supplement the description of the text part of the specification with graphics, so that people can intuitively and vividly understand each technical feature and the overall technical solution of the present invention, but it cannot be understood as a limitation on the scope of protection of the present invention.

[0074] In the description of the present invention, "several" means one or more, "more" means more than two, "greater than", "less than", "exceed" etc. are understood as not including the number itself, and "above", "below", "within" etc. are understood as including the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0075] The present invention proposes an ultrasonic ranging alarm method for preventing collision when opening a car door. The alarm method relies on the position information of moving objects around the vehicle collected before opening the door, and predicts the final position information of the surrounding moving objects through the seven-layer deep learning neural network for mobile position prediction provided by the present invention, to determine whether opening the door will cause a dangerous situation. If there is a danger, the vehicle control system is used to stop the behavior of opening the door, thereby avoiding the occurrence of traffic safety problems such as door-opening collisions; if there is no danger, the vehicle control system is used to allow the behavior of opening the door. The present invention also proposes an ultrasonic ranging alarm system for preventing collision when opening a car door, comprising the hardware control module, information storage module, deep learning module and information sampling module required by the ultrasonic ranging alarm method for preventing collision when opening a car door, so as to effectively avoid the problem of door-opening collisions in road safety without relying on subjective human control.

[0076] Reference Figure 1-Figure 4, a preferred embodiment of the present invention, an ultrasonic ranging alarm method for automobile door opening anti-collision, the alarm method comprises the following steps:

[0077] S1, training data collection stage: delineate the prediction area, determine the position of the moving object entering the prediction area through the ultrasonic wave of the automobile ultrasonic detection device, sample according to the moving speed-time interval, extract the position information of the moving object within the extraction time interval, and form a speed average position information matrix; the step of obtaining the speed average position information matrix is ​​to first calculate the average speed according to the extracted position information of the moving object, and then form a speed position information matrix with the position information and average speed of each target at different times, and then further group them according to the speed, and extract part of the data from the speed position information matrix according to the moving speed of the target to form a speed average position information matrix;

[0078] Specifically, in the training data collection stage, the prediction area size is defined according to the extraction time interval, and the car extraction time interval is denoted as T a ,The extraction time interval includes the sampling time length and the prediction time length, which are denoted as T c , T p , there is the following relationship:

[0079] T a =T c +T p

[0080] Then, the radius d of the prediction area defined by the car a for:

[0081] d a =(T a *V max )+d adj ;

[0082] Among them, V max is the preset maximum moving object speed, d adj is the preset environment distance, and the prediction area defined by the car is d a is the area within a circle of radius;

[0083] The position of the moving object is sampled using the automobile ultrasonic ranging device according to the moving speed and sampling time length within the extraction time interval: the sampling time interval is t, then the two-dimensional position coordinates of the moving object entering the designated area and the average speed of the moving object are extracted according to the sampling time length within the extraction time interval to form a speed position information matrix, denoted as L V :

[0084]

[0085] There are s moving objects entering the prediction area within the extraction time interval; the two-dimensional position information of each moving object is extracted with the sampling time interval t as the sampling period; (x,c) represents the two-dimensional position coordinates of the x-th moving object at the c-th time interval; x is 1, 2, 3...s; c is Sampling position information for each moving object Group, It represents the average moving speed of the moving object s, and the calculation formula is as follows:

[0086]

[0087] From the velocity position information matrix L V The velocity-averaged position information matrix L is further extracted according to the average moving speed of the moving object. ave-V ;

[0088] First, the average moving speed of the moving object is divided into four groups: Group 1, the actual average moving speed is not more than 8m / s; Group 2, the actual average moving speed is not more than 12m / s; Group 3, the actual average moving speed is not more than 20m / s; Group 4, the actual average speed is more than 20m / s; Then, according to the average speed of the moving object, the position speed information is extracted from the speed position information matrix with equal data volume to form the corresponding speed average position information matrix L ave-V :

[0089]

[0090] in, represents the average speed of the c1th moving object in group a1; It represents the two-dimensional position information of the c1th moving object in group a1 in the b1th sampling period; p represents the two-dimensional position information path of p moving objects in each group; by extracting the speed-averaged position information matrix and sampling the motion trajectories of objects with different moving speeds with equal priority, the comprehensiveness and reliability of the data are increased, and the robustness of the network after training is increased.

[0091] S2, training data set determination stage: randomly extract the training position information set and the training real position information set from the speed average position information matrix; specifically, according to the sampling time length T c , from the velocity average position information matrix L ave-V Extract the training position information set, denoted as L tr-c-v ,but:

[0092]

[0093] in, represents the average speed of the c2th moving object randomly selected from group a2; p′ represents the two-dimensional position information path of p' moving objects randomly selected from each group; the speed and position information of 4p' moving objects are randomly selected from four groups with equal probability; the position information matrix L is divided from the speed according to the corresponding position of the moving object in the training position information set ave-V Extract the training real position information set, denoted as L tr-real-v ,but:

[0094]

[0095] Training real location information set L tr-real-v It represents the average speed and final position sampling information of the moving object corresponding to the training position information set; in order to facilitate the input of the subsequent training stage and the needs of the deep learning network, the training position information set L tr-c-v Split the data according to the two-dimensional coordinates x and y axes: split into the training position x-axis information set Lx tr-c-v , training position y-axis information set Ly tr-c-v ;

[0096]

[0097] in, It represents the two-dimensional position information of the k1th moving object in group i1 in the j1th sampling period, and the corresponding x and y represent the corresponding horizontal coordinate information and vertical coordinate information.

[0098] S3, test data set determination stage: delete the training position information set L from the speed average position information matrix tr-c-v The information of the corresponding moving object in the test position information set and the test real position information set are established;

[0099] Specifically, according to the sampling time length T c , delete the training position information set L from the speed average position information matrix tr-c-v The corresponding moving object information in the training position information set is extracted, and one tenth of the moving object data in the total training position information set is used to form the corresponding test position information set, which is denoted as L test-c-v ,but:

[0100]

[0101] p" = p' / 10;

[0102] in, represents the average speed of the p'th moving object randomly selected from group 1. Each group extracts the two-dimensional position information path of p'th moving objects respectively. The four groups randomly extract the speed and position information of 4p'th moving objects with equal probability. According to the corresponding position of the moving object in the test position information set, the position information matrix L is divided from the speed average. ave-V Extract the real test location information set, denoted as L test-real-v ,but:

[0103]

[0104] Test real location information set L test-real-v It represents the average speed and final position sampling information of the moving object corresponding to the test position information set. In order to facilitate the input of the subsequent test phase and the needs of the deep learning network, the test position information set L test-c-v Split the data according to the two-dimensional coordinate xy axis into the test position x-axis information set Lx test-c-v , test position y-axis information set Ly test-c-v ;

[0105]

[0106]

[0107] in, It represents the two-dimensional position information of the k2th moving object in group i2 in the j2th sampling period, and the corresponding x and y represent the corresponding horizontal coordinate information and vertical coordinate information.

[0108] S4, deep learning network establishment stage: according to the application scenario, a seven-layer deep learning network with good generalization ability and strong robustness suitable for mobile location prediction is established;

[0109] Specifically, the seven-layer deep learning network is as follows Figure 3 As shown, according to the outdoor open road scene, the first layer of the network is the input layer, and the input is two The matrices represent the positions of the x-axis and y-axis respectively. They are regarded as two channels, so the input shape is:

[0110]

[0111] Among them, C = 2 is the number of channels, representing the x-axis and y-axis position matrix, H = 4 is the height, representing the number of rows, is the width, indicating the number of columns;

[0112] The second layer of the network is the first convolutional layer. The convolutional layer uses a 3(T-7)×3(T-7) convolution kernel, a step size of 1, and a padding of 0. The output shape is: Conv1 output=(16,2,8);

[0113] The third layer of the network is the first pooling layer. The pooling layer uses a 2×2 maximum pooling with a step size of 2. The output shape is: Pool1 output =(16,1,4);

[0114] The fourth layer of the network is the second convolutional layer. The second convolutional layer uses 32 3×3 convolution kernels, with a step size of 1 and a padding of 1. The output shape is: Conv2 output =(32,1,4);

[0115] The fifth layer of the network is a flattening operation. Before entering the fully connected layer, the output of the convolutional layer needs to be flattened into a one-dimensional vector. The flattening operation turns (32, 1, 4) into 128 neurons.

[0116] The sixth layer of the network is the first fully connected layer, which has 128 neurons as input and 64 neurons as output;

[0117] The seventh layer of the network is the second fully connected layer, which is also the final output layer. The second fully connected layer has 64 neurons as input and 2 neurons as output, representing the two-dimensional coordinates (x, y) predicted by the final network.

[0118] S5, deep learning network training stage: the training position information set in S2 is split according to the two-dimensional coordinates x and y axes, and is split into a training position x-axis information set and a training position y-axis information set; the training position x-axis information set and the training position y-axis information set are used to train the seven-layer deep learning network for mobile position prediction, and the positioning average accuracy threshold and the positioning maximum error threshold are set. When the prediction accuracy of the trained network meets the standard, the network training is stopped, and the corresponding hyperparameters and weight parameter information of the seven-layer deep learning network for mobile position prediction are output;

[0119] Specifically, the training position x-axis information set and the training position y-axis information set are input into the seven-layer deep learning network, and training is performed in batches. The minimum batch size is set to 100, and data is extracted by random sampling. The average positioning accuracy threshold is set to and the maximum positioning error threshold L er-max The output two-dimensional coordinates are compared with the corresponding position information in the training real position information set. The loss function of the network is set to the mean square error function to measure the difference between the predicted coordinates and the real coordinates. The optimizer selected is the Adam optimizer, and the learning rate is set to 0.002 to minimize the loss function to update the network parameters and determine the optimal seven-layer deep learning network. When the average positioning accuracy and the maximum positioning error are both less than the threshold, the network training is stopped, and the deep learning network, the weights and hyperparameters in the network are output and saved.

[0120] S6, mobile location prediction network testing phase: the test location information set in S3 is split according to the two-dimensional coordinates x and y axes, and is divided into a test location x-axis information set and a test location y-axis information set; the trained seven-layer deep learning network prediction positioning accuracy is tested by the test location x-axis information set, the test location y-axis information set and the test real location information set to see if it meets the standard;

[0121] Specifically, the test position x-axis information set and the test position y-axis information set are substituted as the input of the trained seven-layer deep learning network, and the output data is the final position prediction coordinates of the moving object, which are compared with the corresponding real position information coordinates in the obtained test real position information set to obtain the average error value of the test positioning, denoted as L test-err :

[0122]

[0123] Among them, (x i ',y i ') represents the predicted position coordinates, (x i ,y i ) represents the real position information coordinates, |(x i ,y i ),(x i ',y i ')| represents the Euclidean distance between two points, A represents the total number of moving objects in the test real position information set; the average test positioning error value L test-err And the average positioning accuracy threshold set For comparison, if it is less than the average positioning accuracy threshold, the network meets the standard. If it is greater than or equal to the average positioning accuracy threshold, S6 needs to be repeated to train the network until the average positioning error value L is tested. test-err Less than the set positioning average accuracy threshold

[0124] Specific implementation examples Figure 2 As shown in the figure, the method and system verification embodiment is carried out on a certain open road section, wherein the positioning average accuracy threshold The predicted position coordinates obtained by the trained seven-layer deep learning network and the corresponding real position coordinates in the test real position information set obtained by S3 are sampled as shown in Table 1:

[0125] Table 1: Sampling table of position prediction coordinates and real position coordinates of seven-layer deep learning network

[0126]

[0127] The sampling group data is 5 groups, and the average error value of the test positioning is L test-errIt is 0.5709m, which is less than the preset average accuracy threshold of 0.6m, demonstrating the excellent performance and generalization ability of the designed and trained seven-layer deep learning network.

[0128] S7, actual moving object position prediction stage: two modes are set: low-energy automobile ultrasonic ranging mode and low-latency automobile ultrasonic ranging mode. According to different choices of the occupants, the automobile ultrasonic detection device determines the position of the moving object entering the prediction area, extracts the online position information set of all moving objects in the automobile prediction area, substitutes it into the trained seven-layer deep learning network, and outputs the online prediction position information set of all moving objects;

[0129] Specifically, two modes are set according to the actual energy efficiency of the car: When the car's energy efficiency is good, the low-latency car ultrasonic ranging mode is selected: The low-latency car ultrasonic ranging mode focuses on the system's ability to quickly determine whether to open the door or not, and requires the car ultrasonic device to collect the position information data set of the moving objects entering the prediction area in real time. The car extraction time interval is T a ,The extraction time interval consists of two parts: the sampling time length T c and the prediction time length T p , extract the two-dimensional position information of the moving object within the sampling time length with the sampling time interval t as the sampling period. When the occupant makes a door opening request, the online position information set of the moving object entering the prediction area is immediately input into the trained seven-layer deep learning network for mobile position prediction, and the online prediction position information set L of all moving objects is output. pred-online , which is expressed as follows:

[0130] L pred-online =[l1…l i …l n ];

[0131] Where n represents the total number of moving objects entering the prediction area; l n The two-dimensional coordinate information of the predicted position of the nth moving object entering the prediction area; the reaction delay is the calculation time of the seven-layer deep learning network, and the car can be predicted within the prediction time length, that is, T p After the time has passed, the door is opened or a danger alarm is issued;

[0132] In the case of poor energy efficiency of the car, choose the low-energy car ultrasonic ranging mode: the low-energy car ultrasonic ranging mode focuses on the low energy consumption of the system. When the occupants make a door opening request, the car ultrasonic device starts to collect the position information data set of the moving objects entering the prediction area. The car extraction time interval is T a ,The extraction time interval consists of two parts: the sampling time length T c and the prediction time length Tp , extract the two-dimensional position information of the moving object within the sampling time length with the sampling time interval t as the sampling period, input the collected online position information set of the moving object into the trained seven-layer deep learning network for mobile position prediction, and output the online predicted position information set L of all moving objects pred-online , the reaction delay is the sum of the computation time and sampling time length of the seven-layer deep learning network. The car can extract the time interval, i.e., T a After a certain time, the door opening action is completed or a danger alarm prediction is issued; according to the energy efficiency requirements of the car and the driving delay requirements of the occupants, the present invention provides two tendency modes, which are more humane and reasonable.

[0133] Specific implementation examples Figure 2 As shown in the figure, the method and system verification example is carried out on an open road section in Nanchang, wherein the vehicle extraction time interval T is preset. a The sampling time length is 3s. c is 2s and the prediction time length is T p The two-dimensional position information of the moving object is extracted within the sampling time interval and the sampling time interval t is 0.4s, as shown in Table 2:

[0134] Table 2: 2D position sampling information table of moving objects in the prediction area

[0135]

[0136] The above is the two-dimensional position sampling information of the five moving objects entering the prediction area in the sampling time interval, which is input into the trained seven-layer deep learning network for mobile position prediction, and the online predicted position information of all moving objects is output as shown in Table 3 below:

[0137] Table 3: Prediction information table of two-dimensional position of moving objects in the prediction area

[0138] serial number 1 2 3 4 5 Predicted location (27.89,30.89) (40.25,32.00) (-22.00,12.50) (31.30,-13.30) (1.20,1.10)

[0139] The above is the online predicted position information of the five moving objects entering the prediction area obtained through the seven-layer deep network.

[0140] S8, door opening collision safety control stage: set the car door opening collision danger zone, compare the final positions of all moving objects in the online predicted position information set of moving objects entering the prediction area with the car door opening collision danger zone, if there is a moving object in the car door opening collision danger zone within the predicted time interval of door opening, it is determined to be a dangerous state, and the behavior of opening the door is stopped through the vehicle control system, thereby avoiding the occurrence of door opening collision traffic safety problems; if all moving objects are not in the car door opening collision danger zone within the predicted time interval of door opening, it is determined to be a safe state, and the behavior of opening the door is allowed through the vehicle control system.

[0141] Specifically, the door opening collision danger zone is set according to the size of the car door, and the door size is denoted as d car , the environmental state parameter is denoted as d adj ; The corresponding door opening collision danger zone is the 90-degree fan-shaped area corresponding to the two doors, and the radius of the fan-shaped area is D dgr :

[0142]

[0143] The online prediction location information set L pred-online The predicted position information of the moving objects in the vehicle is compared with the door opening collision danger zone. If there is a moving object in the door opening collision danger zone within the predicted door opening time, it is determined to be a dangerous state. The vehicle control system prevents the door from being opened and issues a danger alarm to avoid the door opening collision traffic safety problem. If all moving objects are not in the door opening collision danger zone within the predicted door opening time, it is determined to be a safe state, and the vehicle control system allows the door to be opened.

[0144] Specific implementation examples Figure 2 As shown in the figure, the method and system verification embodiment is carried out on an open road section in Nanchang, where the door size d car It is 1.5m. From the data in Table 3, it can be seen that the online predicted position of the moving object No. 5 will fall in the preset door opening collision danger zone, which is a dangerous predicted position. Therefore, although the online predicted positions of the moving objects No. 1 to No. 4 fall in the non-door opening collision danger zone, the moving object No. 5 has a safety hazard and is determined to be in a dangerous state. The vehicle control system stops the behavior of opening the door and issues a danger alarm to avoid the occurrence of door opening collision traffic safety problems.

[0145] like Figure 4As shown, the present invention also discloses an alarm system according to the above-mentioned ultrasonic ranging alarm method for automobile door opening anti-collision, the alarm system comprising a vehicle system module, a hazard alarm, an ultrasonic ranging module, an ultrasonic ranging sampling module, and a door opening trigger judgment module;

[0146] On-board system module: includes deep learning module, information storage module and central control module; among them, the deep learning module has two functions: on the one hand, it is used to receive the position sampling information of real-time moving objects, and output the corresponding predicted position information of moving objects through the trained deep learning network; on the other hand, it is used to receive the position sampling information of moving objects in the training stage, and output the neural network and corresponding network weights and hyperparameter values ​​that have reached the accuracy through training; the information storage module is used to store the position sampling information of moving objects entering the prediction area, the deep learning network, the network weights and hyperparameter values ​​trained by the neural network module, and all size information parameters; the central control module has two functions: on the one hand, it is used to receive mode selection information and control the operation of the ultrasonic module; on the other hand, after receiving the door opening request, it extracts the position sampling information of the moving object and transmits it to the deep learning module to obtain the predicted position information of the moving object, controls the transmission of this information to the door opening trigger judgment module, receives the corresponding judgment information, outputs the corresponding door opening command or outputs the danger command to the danger alarm module;

[0147] Danger alarm module: receives the danger command from the central control module, rejects the operator's request to open the door, and sends out a danger alarm message to inform the occupants that the door is prohibited from opening at this time;

[0148] Ultrasonic module: used to select the mode of the central control module, perform sampling according to the extraction time interval, and complete the distance and position extraction of the moving object entering the vehicle prediction area through the ultrasonic ranging principle;

[0149] Door opening trigger judgment module: used to receive the predicted position information of the corresponding moving objects obtained by the deep learning module, and make a positional relationship judgment between the predicted position information and the door opening collision danger zone: if the predicted positions of all moving objects are not in the door opening collision danger zone, the door opening command is output; otherwise, the danger alarm command is output to the danger alarm module;

[0150] Preferably, a computer storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned ultrasonic ranging alarm method for preventing collision when opening a car door;

[0151] Preferably, a computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned ultrasonic ranging alarm method for automobile door opening anti-collision when executing the computer program.

[0152] The present invention realizes high-precision monitoring of moving objects through ultrasonic ranging and deep learning technology. In the training stage, the speed-averaged position information matrix is ​​used to enhance the comprehensiveness and reliability of the data, thereby improving the robustness of the network. In the testing stage, a seven-layer deep learning network with strong generalization ability is established to ensure the positioning accuracy and stability of the system in complex environments. Finally, the system significantly improves driving safety by real-time monitoring of moving objects and judging the risk of door opening through a seven-layer deep learning network, effectively eliminating the traffic safety problem of door opening collision and ensuring the safety of the vehicle and passengers.

[0153] Under the premise that no conflict occurs, those skilled in the art may freely combine and superimpose the above-mentioned additional technical features.

[0154] The above description is only a preferred implementation mode of the present invention. Any technical solution that achieves the purpose of the present invention by substantially the same means belongs to the protection scope of the present invention.

Claims

1. An ultrasonic ranging alarm method for preventing collision when opening a car door, characterized in that: The alarm method comprises the following steps: S1, training data collection stage: delineate the prediction area, determine the position of the moving object entering the prediction area through the ultrasonic wave of the automobile ultrasonic detection device, sample according to the moving speed-time interval, extract the position information of the moving object within the extraction time interval, and form a speed average position information matrix; the step of obtaining the speed average position information matrix is ​​to first calculate the average speed according to the extracted position information of the moving object, and then form a speed position information matrix with the position information and average speed of each target at different times, and then further group them according to the speed, and extract part of the data from the speed position information matrix according to the moving speed of the target to form a speed average position information matrix; S2, training data set determination stage: randomly extracting a training position information set and a training real position information set from the speed average position information matrix; S3, test data set determination stage: delete the training position information set L from the speed average position information matrix tr-c-v According to the information of the corresponding moving object, a test position information set and a test real position information set are established; S4, deep learning network establishment stage: according to the application scenario, a seven-layer deep learning network with good generalization ability and strong robustness suitable for mobile location prediction is established; S5, deep learning network training stage: the training position information set in S2 is split according to the two-dimensional coordinates x and y axes, and is split into a training position x-axis information set and a training position y-axis information set; the training position x-axis information set and the training position y-axis information set are used to train the seven-layer deep learning network for mobile position prediction, and the positioning average accuracy threshold and the positioning maximum error threshold are set. When the prediction accuracy of the trained network meets the standard, the network training is stopped, and the corresponding hyperparameters and weight parameter information of the seven-layer deep learning network for mobile position prediction are output; S6, mobile location prediction network testing phase: the test location information set in S3 is split according to the two-dimensional coordinates x and y axes, and is divided into a test location x-axis information set and a test location y-axis information set; the trained seven-layer deep learning network prediction positioning accuracy is tested by the test location x-axis information set, the test location y-axis information set and the test real location information set to see if it meets the standard; S7, actual moving object position prediction stage: two modes are set: low-energy automobile ultrasonic ranging mode and low-latency automobile ultrasonic ranging mode. According to different choices of the occupants, the automobile ultrasonic detection device determines the position of the moving object entering the prediction area, extracts the online position information set of all moving objects in the automobile prediction area, substitutes it into the trained seven-layer deep learning network, and outputs the online prediction position information set of all moving objects; S8, door opening collision safety control stage: set the car door opening collision danger zone, compare the final positions of all moving objects in the online predicted position information set of moving objects entering the prediction area with the car door opening collision danger zone, if there is a moving object in the car door opening collision danger zone within the predicted time length of door opening, it is determined to be a dangerous state, and the behavior of opening the door is prevented through the vehicle control system, thereby avoiding the occurrence of door opening collision traffic safety problems; if all moving objects are not in the car door opening collision danger zone within the predicted time length of door opening, it is determined to be a safe state, and the behavior of opening the door is allowed through the vehicle control system.

2. The ultrasonic ranging alarm method for preventing collision when opening a car door according to claim 1 is characterized in that: The specific contents of the training data collection stage in S1 include: In the training data collection stage, the prediction area size is demarcated according to the extraction time interval, and the car extraction time interval is denoted as T a ,The extraction time interval includes the sampling time length and the prediction time length, which are denoted as T c , T p , there is the following relationship: T a =T c +T p ; Then, the radius d of the prediction area defined by the car a for: d a =(T a *V max )+d adj ; Among them, V max is the preset maximum moving object speed, d adj is the preset environment distance, and the prediction area defined by the car is d a is the area within a circle of radius; The position of the moving object is sampled using the automobile ultrasonic ranging device according to the moving speed and sampling time length within the extraction time interval: the sampling time interval is t, then the two-dimensional position coordinates of the moving object entering the designated area and the average speed of the moving object are extracted according to the sampling time length within the extraction time interval to form a speed position information matrix, denoted as L V : There are s moving objects entering the prediction area within the extraction time interval; the two-dimensional position information of each moving object is extracted with the sampling time interval t as the sampling period; (x,c) represents the two-dimensional position coordinates of the x-th moving object at the c-th time interval; x is 1, 2, 3...s; c is Sampling position information for each moving object Group, It represents the average moving speed of the moving object s, and the calculation formula is as follows: From the velocity position information matrix L V The velocity-averaged position information matrix L is further extracted according to the average moving speed of the moving object. ave-V ; First, the average moving speed of the moving object is divided into four groups: Group 1, the actual average moving speed is not more than 8m / s; Group 2, the actual average moving speed is not more than 12m / s; Group 3, the actual average moving speed is not more than 20m / s; Group 4, the actual average speed is more than 20m / s; Then, according to the average speed of the moving object, the position speed information is extracted from the speed position information matrix with equal data volume to form the corresponding speed average position information matrix L ave-V : in, represents the average speed of the c1th moving object in group a1; It represents the two-dimensional position information of the c1th moving object in group a1 in the b1th sampling period; p represents the two-dimensional position information path of p moving objects in each group; by extracting the speed-averaged position information matrix and sampling the motion trajectories of objects with different moving speeds with equal priority, the comprehensiveness and reliability of the data are increased, and the robustness of the network after training is increased.

3. The ultrasonic ranging alarm method for preventing collision when opening a car door according to claim 2 is characterized in that: The method of randomly extracting the training position information set and the training real position information set in S2 includes: c , from the velocity average position information matrix L ave-V Extract the training position information set, denoted as L tr-c-v ,but: in, represents the average speed of the c2th moving object randomly selected from group a2; p′ represents the two-dimensional position information path of p′ moving objects randomly selected from each group; the speed and position information of 4p′ moving objects are randomly selected from four groups with equal probability; the position information matrix L is divided from the speed according to the corresponding position of the moving object in the training position information set ave-V Extract the training real position information set, denoted as L tr-real-v ,but: Training real location information set L tr-real-v It represents the average speed and final position sampling information of the moving object corresponding to the training position information set; in order to facilitate the input of the subsequent training stage and the needs of the deep learning network, the training position information set L tr-c-v Split the data according to the two-dimensional coordinates x and y axes: split into the training position x-axis information set Lx tr-c-v , training position y-axis information set Ly tr-c-v ; in, It represents the two-dimensional position information of the k1th moving object in group i1 in the j1th sampling period, and the corresponding x and y represent the corresponding horizontal coordinate information and vertical coordinate information.

4. The ultrasonic distance measurement alarm method for automobile door opening anti-collision according to claim 3 is characterized in that: The method of stripping the predicted position information set, establishing the test position information set and predicting the real position information set in S3 includes: c , delete the training position information set L from the speed average position information matrix tr-c-v The corresponding moving object information in the training position information set is extracted, and one tenth of the moving object data in the total training position information set is used to form the corresponding test position information set, which is denoted as L test-c-v ,but: p" = p' / 10; in, represents the average speed of the p'th moving object randomly selected from group 1. Each group extracts the two-dimensional position information path of p'th moving objects respectively. The four groups randomly extract the speed and position information of 4p'th moving objects with equal probability. According to the corresponding position of the moving object in the test position information set, the position information matrix L is divided from the speed average. ave-V Extract the real test location information set, denoted as L test-real-v ,but: Test real location information set L test-real-v It represents the average speed and final position sampling information of the moving object corresponding to the test position information set. In order to facilitate the input of the subsequent test phase and the needs of the deep learning network, the test position information set L test-c-v Split the data according to the two-dimensional coordinate xy axis into the test position x-axis information set Lx test-c-v , test position y-axis information set Ly test-c-v ; in, It represents the two-dimensional position information of the k2th moving object in group i2 in the j2th sampling period, and the corresponding x and y represent the corresponding horizontal coordinate information and vertical coordinate information.

5. The ultrasonic distance measurement alarm method for automobile door opening anti-collision according to claim 4, characterized in that: The seven-layer deep learning network in the deep learning network establishment phase in S4: The first layer of the network is the input layer, and the input is two The matrices represent the positions of the x-axis and y-axis respectively. They are regarded as two channels, so the input shape is: Among them, C = 2 is the number of channels, representing the x-axis and y-axis position matrix, H = 4 is the height, representing the number of rows, is the width, indicating the number of columns; The second layer of the network is the first convolutional layer. The convolutional layer uses a 3(T-7)×3(T-7) convolution kernel, a step size of 1, and a padding of 0. The output shape is: Conv1 output =(16,2,8); The third layer of the network is the first pooling layer. The pooling layer uses a 2×2 maximum pooling with a step size of 2. The output shape is: Pool1 output =(16,1,4); The fourth layer of the network is the second convolutional layer. The second convolutional layer uses 32 3×3 convolution kernels, with a step size of 1 and a padding of 1. The output shape is: Conv2 output =(32,1,4); The fifth layer of the network is a flattening operation. Before entering the fully connected layer, the output of the convolutional layer needs to be flattened into a one-dimensional vector. The flattening operation turns (32, 1, 4) into 128 neurons. The sixth layer of the network is the first fully connected layer, which has 128 neurons as input and 64 neurons as output; The seventh layer of the network is the second fully connected layer, which is also the final output layer. The second fully connected layer has 64 neurons as input and 2 neurons as output, representing the two-dimensional coordinates (x, y) predicted by the final network.

6. The ultrasonic distance measurement alarm method for automobile door opening anti-collision according to claim 5, characterized in that: The deep learning network training phase in S5 includes: inputting the training position x-axis information set and the training position y-axis information set into the seven-layer deep learning network, training in batches, with the minimum batch set to 100, extracting data using random sampling, and setting the positioning average accuracy threshold to and the maximum positioning error threshold L er-max The output two-dimensional coordinates are compared with the corresponding position information in the training real position information set. The loss function of the network is set to the mean square error function to measure the difference between the predicted coordinates and the real coordinates. The optimizer selected is the Adam optimizer, and the learning rate is set to 0.002 to minimize the loss function to update the network parameters and determine the optimal seven-layer deep learning network. When the average positioning accuracy and the maximum positioning error are both less than the threshold, the network training is stopped, and the deep learning network, the weights and hyperparameters in the network are output and saved.

7. The ultrasonic distance measurement alarm method for preventing collision when opening a car door according to claim 6 is characterized in that: The mobile position prediction network test phase in S6 includes: substituting the test position x-axis information set and the test position y-axis information set as the input of the trained seven-layer deep learning network, and the output data is the final position prediction coordinates of the mobile object, and is compared with the corresponding real position information coordinates in the test real position information set to obtain the test positioning average error value, which is recorded as L test-err : Among them, (x i ',y i ') represents the predicted position coordinates, (x i ,y i ) represents the real position information coordinates, |(x i ,y i ),(x i ',y i ')| represents the Euclidean distance between two points, A represents the total number of moving objects in the test real position information set; the average test positioning error value L test-err And the average positioning accuracy threshold set For comparison, if it is less than the average positioning accuracy threshold, the network meets the standard. If it is greater than or equal to the average positioning accuracy threshold, S6 needs to be repeated to train the network until the average positioning error value L is tested. test-err Less than the set average positioning accuracy threshold 8. The ultrasonic distance measurement alarm method for automobile door opening anti-collision according to claim 7, characterized in that: In S7, when the car energy efficiency is good, the low-latency car ultrasonic ranging mode is selected: the low-latency car ultrasonic ranging mode focuses on the system to quickly make a judgment on whether to open the door or not. The car ultrasonic device is required to collect the position information data set of the moving objects entering the prediction area in real time. The car extraction time interval is T a ,The extraction time interval consists of two parts: the sampling time length T c and the prediction time length T p , extract the two-dimensional position information of the moving object within the sampling time length with the sampling time interval t as the sampling period. When the occupant makes a door opening request, the online position information set of the moving object entering the prediction area is immediately input into the trained seven-layer deep learning network for mobile position prediction, and the online prediction position information set L of all moving objects is output. pred-online , which is expressed as follows: L pred-online ＝[l1 … l i … l n ]； Where n represents the total number of moving objects entering the prediction area; l n The two-dimensional coordinate information of the predicted position of the nth moving object entering the prediction area; the reaction delay is the calculation time of the seven-layer deep learning network, and the car can be predicted within the prediction time length, that is, T p After the time has passed, the door is opened or a danger alarm is issued; In the case of poor energy efficiency of the car, choose the low-energy car ultrasonic ranging mode: the low-energy car ultrasonic ranging mode focuses on the low energy consumption of the system. When the occupants make a door opening request, the car ultrasonic device starts to collect the position information data set of the moving objects entering the prediction area. The car extraction time interval is T a ,The extraction time interval consists of two parts: the sampling time length T c and the prediction time length T p , extract the two-dimensional position information of the moving object within the sampling time length with the sampling time interval t as the sampling period, input the collected online position information set of the moving object into the trained seven-layer deep learning network for mobile position prediction, and output the online predicted position information set L of all moving objects pred-online , the reaction delay is the sum of the computation time and sampling time length of the seven-layer deep learning network. The car can extract the time interval, i.e., T a After the time has passed, the door is opened or a danger alarm is issued.

9. The ultrasonic distance measurement alarm method for automobile door opening anti-collision according to claim 8, characterized in that: The car door opening collision danger zone in S8 includes: setting the door opening collision danger zone according to the size of the car door, and the door size is denoted as d car , the environmental state parameter is denoted as d adj ; The corresponding door opening collision danger zone is the 90-degree fan-shaped area corresponding to the two doors, and the radius of the fan-shaped area is D dgr : The online prediction location information set L pred-online The predicted position information of the moving objects in the vehicle is compared with the door opening collision danger zone. If there is a moving object in the door opening collision danger zone within the predicted door opening time, it is determined to be a dangerous state. The vehicle control system prevents the door from being opened and issues a danger alarm to avoid the door opening collision traffic safety problem. If all moving objects are not in the door opening collision danger zone within the predicted door opening time, it is determined to be a safe state, and the vehicle control system allows the door to be opened.

10. An alarm system of an ultrasonic ranging alarm method for preventing collision when opening a car door according to any one of claims 1 to 9, characterized in that: The alarm system includes a vehicle-mounted system module, a hazard alarm, an ultrasonic ranging module, an ultrasonic ranging sampling module, and a door opening trigger judgment module; the vehicle-mounted system module includes a deep learning module, an information storage module, and a central control module; When the central control module receives the mode information, the position sampling information of the moving object entering the predicted area is stored in the information storage module through the ultrasonic ranging module and the ultrasonic ranging sampling module. After the central control module receives the door opening request from the occupant of the vehicle, the position sampling information of the moving object is extracted and transmitted to the deep learning module. The deep learning module extracts the deep learning network and the corresponding network weights and hyperparameter values ​​from the information storage module to obtain the corresponding predicted position information. The predicted position information and the door opening collision danger zone are judged through the door opening trigger judgment module, and the door opening command or danger alarm command is output to the danger alarm.

Images