Vehicle collision identification method, device, equipment and medium

Abstract

The invention provides a vehicle collision identification method and device, equipment and a medium. The method comprises the steps of obtaining vehicle data of a target vehicle in a preset time period; data preprocessing is carried out on the vehicle data to obtain a target feature map, the target feature map comprises time domain features and frequency domain features of the vehicle data, and the data preprocessing comprises time-frequency feature extraction processing based on continuous wavelet transform; and inputting the target feature map into a target model to obtain a collision recognition result of the target vehicle in the preset time period, the target model being a trained deep learning model. A target feature map comprises time domain features and frequency domain features of vehicle data; by inputting the target feature map into the target model, the target model can comprehensively analyze the time domain feature and the frequency domain feature of the vehicle data, so that the accuracy of the output vehicle collision recognition result can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle collision, and in particular relates to a vehicle collision identification method, device, equipment and medium. Background Art

[0002] In related technologies, identifying whether a vehicle has experienced a collision is primarily accomplished by determining and analyzing the time series characteristics of vehicle signals. However, these characteristics are relatively simple, representing only changes in vehicle signals over time. Therefore, performing collision recognition based solely on these characteristics results in low accuracy. Summary of the Invention

[0003] In view of the above-mentioned shortcomings of the prior art, an object of the present invention is to provide a vehicle collision identification method, device, equipment and medium for solving the above-mentioned problems.

[0004] The vehicle collision identification method provided by the present invention includes:

[0005] Obtain vehicle data of the target vehicle in a preset time period;

[0006] Performing data preprocessing on the vehicle data to obtain a target feature map, wherein the target feature map includes time domain features and frequency domain features of the vehicle data, and the data preprocessing includes performing time-frequency feature extraction processing based on continuous wavelet transform;

[0007] The target feature map is input into a target model to obtain a collision recognition result of the target vehicle in the preset time period, and the target model is a trained deep learning model.

[0008] In one embodiment of the present invention, the vehicle data includes time series data corresponding to a plurality of vehicle signals, and the data preprocessing of the vehicle data to obtain a target feature map includes:

[0009] Performing the time-frequency feature extraction process on the time series data corresponding to the multiple vehicle signals to obtain a plurality of two-dimensional feature maps, each of which includes the time domain features and frequency domain features of the corresponding vehicle signals;

[0010] The multiple two-dimensional feature maps are spliced together to obtain a three-dimensional feature map of the vehicle data, and the target feature map includes the three-dimensional feature map.

[0011] In one embodiment of the present invention, before splicing the multiple two-dimensional feature maps to obtain the three-dimensional feature map of the vehicle data, the method further includes:

[0012] Performing standardization processing on the multiple two-dimensional feature maps to obtain the multiple two-dimensional feature maps after standardization processing, wherein the standardization processing includes performing modulo processing and normalization processing on each feature in the two-dimensional feature map, and the normalization processing is used to map each feature after the modulo processing to a grayscale value within a preset range;

[0013] The step of splicing the multiple two-dimensional feature maps to obtain a three-dimensional feature map of the vehicle data includes:

[0014] The multiple two-dimensional feature maps after normalization are spliced together to obtain a three-dimensional feature map of the vehicle data.

[0015] In one embodiment of the present invention, the target model includes a feature extraction network, an attention network, a fully connected network, and a classification network. Inputting the target feature map into the target model to obtain a collision recognition result of the target vehicle in the preset time period includes:

[0016] Inputting the target feature map into a feature extraction network for feature extraction processing to obtain a first feature map, wherein the feature extraction processing includes a convolution operation;

[0017] Inputting the first feature map into the attention network to obtain a second feature map determined based on a target attention mechanism, wherein the target attention mechanism includes at least one of a channel attention mechanism and a spatial attention mechanism;

[0018] Inputting the second feature map into the fully connected network to obtain classification data;

[0019] The classification data is input into the classification network to obtain a collision recognition result of the target vehicle in the preset time period.

[0020] In one embodiment of the present invention, the feature extraction network is a multi-layer feature extraction network, and the multi-layer feature extraction network is used to implement multi-level feature extraction processing from low-level features to high-level features.

[0021] In one embodiment of the present invention, the feature extraction network includes a first feature extraction network, and the first feature extraction network includes:

[0022] a first feature extraction layer, the first feature extraction layer being used to extract features of input data of the first feature extraction network;

[0023] a first channel adjustment layer, wherein the number of output channels of the first channel adjustment layer is the same as the number of output channels of the first feature extraction network, and the first channel adjustment layer is used to adjust the channels of input data of the first feature extraction network;

[0024] The first feature extraction layer and the first channel adjustment layer use a residual connection to add the output data of the first feature extraction layer and the output data of the first channel adjustment layer.

[0025] In one embodiment of the present invention, the target model is trained by the following steps:

[0026] Obtain vehicle data of the target vehicle in different time periods;

[0027] Performing data preprocessing on the vehicle data of the different time periods to obtain characteristic graphs of the different time periods;

[0028] Determine a label of the feature graph of each time period in the feature graphs of different time periods;

[0029] Based on the feature graphs of the different time periods and the labels of the feature graphs of each time period, the deep learning model is trained to adjust the parameters of the deep learning model;

[0030] When the model training meets the preset requirements, the target model is obtained.

[0031] The vehicle collision recognition device provided by the present invention includes:

[0032] An acquisition module is used to acquire vehicle data of a target vehicle in a preset time period;

[0033] A data preprocessing module, configured to perform data preprocessing on the vehicle data to obtain a target feature map, wherein the target feature map includes time domain features and frequency domain features of the vehicle data, wherein the data preprocessing includes time-frequency feature extraction based on continuous wavelet transform;

[0034] A determination module is used to input the target feature map into a target model to obtain a collision recognition result of the target vehicle in the preset time period, and the target model is a trained deep learning model.

[0035] In one embodiment of the present invention, the vehicle data includes time series data corresponding to a plurality of vehicle signals, and the data preprocessing module includes:

[0036] a time-frequency feature extraction and processing unit, configured to perform the time-frequency feature extraction processing on the time series data corresponding to the multiple vehicle signals, respectively, to obtain a plurality of two-dimensional feature maps, each of which includes a time domain feature and a frequency domain feature of the corresponding vehicle signal;

[0037] A splicing unit is used to splice the multiple two-dimensional feature maps to obtain a three-dimensional feature map of the vehicle data, and the target feature map includes the three-dimensional feature map.

[0038] In one embodiment of the present invention, the data preprocessing module further includes:

[0039] a normalization processing unit, configured to perform normalization processing on the plurality of two-dimensional feature maps to obtain the plurality of normalized two-dimensional feature maps, wherein the normalization processing includes performing modulo processing and normalization processing on each feature in the two-dimensional feature map, and the normalization processing is configured to map each feature after the modulo processing to a grayscale value within a preset range;

[0040] The splicing unit is specifically used for:

[0041] The multiple two-dimensional feature maps after normalization are spliced together to obtain a three-dimensional feature map of the vehicle data.

[0042] In one embodiment of the present invention, the target model includes a feature extraction network, an attention network, a fully connected network, and a classification network, and the determination module is specifically used to:

[0043] Inputting the target feature map into a feature extraction network for feature extraction processing to obtain a first feature map, wherein the feature extraction processing includes a convolution operation;

[0044] Inputting the first feature map into the attention network to obtain a second feature map determined based on a target attention mechanism, wherein the target attention mechanism includes at least one of a channel attention mechanism and a spatial attention mechanism;

[0045] Inputting the second feature map into the fully connected network to obtain classification data;

[0046] The classification data is input into the classification network to obtain a collision recognition result of the target vehicle in the preset time period.

[0047] In one embodiment of the present invention, the feature extraction network is a multi-layer feature extraction network, and the multi-layer feature extraction network is used to implement multi-level feature extraction processing from low-level features to high-level features.

[0048] In one embodiment of the present invention, the feature extraction network includes a first feature extraction network, and the first feature extraction network includes:

[0049] a first feature extraction layer, the first feature extraction layer being used to extract features of input data of the first feature extraction network;

[0050] a first channel adjustment layer, wherein the number of output channels of the first channel adjustment layer is the same as the number of output channels of the first feature extraction network, and the first channel adjustment layer is used to adjust the channels of input data of the first feature extraction network;

[0051] The first feature extraction layer and the first channel adjustment layer use a residual connection to add the output data of the first feature extraction layer and the output data of the first channel adjustment layer.

[0052] In one embodiment of the present invention, the target model is trained by the following steps:

[0053] Obtain vehicle data of the target vehicle in different time periods;

[0054] Performing data preprocessing on the vehicle data of the different time periods to obtain characteristic graphs of the different time periods;

[0055] Determine a label of the feature graph of each time period in the feature graphs of different time periods;

[0056] Based on the feature graphs of the different time periods and the labels of the feature graphs of each time period, the deep learning model is trained to adjust the parameters of the deep learning model;

[0057] When the model training meets the preset requirements, the target model is obtained.

[0058] The electronic device provided by the present invention includes:

[0059] one or more processors;

[0060] The storage device is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the electronic device implements the vehicle collision recognition method.

[0061] The computer-readable storage medium provided by the present invention stores a computer program thereon. When the computer program is executed by a processor of a computer, the computer is enabled to execute the vehicle collision identification method.

[0062] The beneficial effects of the present invention are as follows: the present invention performs data preprocessing on the vehicle, including time-frequency feature extraction processing based on continuous wavelet transform to obtain a target feature map, which may include time domain features and frequency domain features of the vehicle data; by inputting the target feature map into the target model, the target model can comprehensively analyze the time domain features and frequency domain features of the vehicle data, thereby improving the accuracy of the output vehicle collision recognition results.

[0063] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings are incorporated into and constitute a part of the specification, illustrating embodiments consistent with the present application and, together with the specification, serving to explain the principles of the present application. It is obvious that the drawings described below are merely some embodiments of the present application, and a person of ordinary skill in the art can derive other drawings based on these drawings without inventive effort. In the drawings:

[0065] Figure 1 is a flow chart illustrating a vehicle collision identification method according to an exemplary embodiment of the present application;

[0066] Figure 2 is a schematic structural diagram of a target model and a deep learning model shown in an exemplary embodiment of the present application;

[0067] Figure 3 is a block diagram of a vehicle collision recognition device shown in an exemplary embodiment of the present application;

[0068] Figure 4 A schematic diagram of the structure of a computer system suitable for implementing an electronic device according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0069] The following describes the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art will readily appreciate the other advantages and benefits of the present invention from the disclosure herein. The present invention may also be implemented or applied through various other specific embodiments, and the various details in this specification may be modified or altered based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are intended only to illustrate the present invention and are not intended to limit the scope of protection of the present invention.

[0070] It should be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0071] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the embodiments of the present invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring the embodiments of the present invention.

[0072] See also Figure 1 , Figure 1FIG. 1 is a flow chart of a vehicle collision identification method shown in an exemplary embodiment of the present application. Figure 1 As shown, in an exemplary embodiment, the vehicle collision identification method includes at least steps S110 to S130, which are described in detail as follows:

[0073] Step S110: Acquire vehicle data of the target vehicle in a preset time period.

[0074] Step S120 , performing data preprocessing on the vehicle data to obtain a target feature map, wherein the target feature map includes time domain features and frequency domain features of the vehicle data. The data preprocessing includes performing time-frequency feature extraction processing based on continuous wavelet transform.

[0075] Step S130: input the target feature map into a target model to obtain a collision recognition result of the target vehicle in the preset time period, where the target model is a trained deep learning model.

[0076] In step S110, the vehicle data primarily includes data collected by vehicle body sensors. The type of vehicle data required for actual application of the target model can be consistent with the type of data required during the training process. For example, during training, data on N types of signals is obtained for subsequent model training. In actual application, data on the same N types of signals is also obtained for further vehicle collision identification, where N is a positive integer. The vehicle data for the preset time period can be obtained using a sliding window. The vehicle data is arranged in time sequence, and based on the window sliding, the data within a window, i.e., the vehicle data for the preset time period, is obtained.

[0077] In step S120, the vehicle data is preprocessed, including time-frequency feature extraction based on continuous wavelet transform. Continuous Wavelet Transform (CWT) decomposes the signal at multiple scales (frequency) by introducing a wavelet function (wavelet), revealing the characteristics of the signal at different scales and time positions. The core of the continuous wavelet transform is to convolve the signal through a mother wavelet function. The wavelet function is a function with finite support or rapid decay, and has good locality. By translating and scaling the wavelet function, the local characteristics of the signal can be captured. Among them, translation refers to adjusting the position of the wavelet function, which is used to analyze the characteristics of the signal at different time positions; scaling refers to adjusting the scale of the wavelet function, which is used to analyze the high-frequency or low-frequency components of the signal. Therefore, the embodiment of the present application can reasonably set the above-mentioned translation parameters and scale parameters based on the continuous wavelet transform to obtain the time domain characteristics and frequency domain characteristics of the above-mentioned vehicle data.

[0078] In some embodiments, the time-frequency characteristics of the vehicle data can be directly extracted based on wavelet transform. In other embodiments, the data preprocessing includes first performing certain processing on the vehicle data and then extracting the time-frequency characteristics of the vehicle data, wherein the vehicle data is processed, for example, by normalization based on Z-score. Z-score normalization converts the data into a distribution with a mean of 0 and a standard deviation of 1 by subtracting the mean from the original data and then dividing it by the standard deviation. The specific formula for Z-score normalization is as follows:

[0079]

[0080] in, is a vector representing the normalized signal value of vehicle signal i in the nth time period (or time segment); is a vector representing the original signal value of vehicle signal i in the nth time period; μ i , the mean value calculated based on all historical data of vehicle signal i; σ i , the standard deviation calculated based on all historical data of vehicle signal i.

[0081] In step 130, the target feature map is input into the target model, which generates a collision recognition result for the target vehicle during the preset time period. The collision recognition result includes whether a collision occurred or not. The target model is a trained deep learning model that determines the vehicle collision recognition result based on the pre-processed vehicle data.

[0082] The present application performs data preprocessing on the vehicle, including time-frequency feature extraction based on continuous wavelet transform to obtain a target feature map, which may include the time domain features and frequency domain features of the vehicle data; by inputting the target feature map into the target model, the target model can comprehensively analyze the time domain features and frequency domain features of the vehicle data, thereby improving the accuracy of the output vehicle collision recognition results.

[0083] In one embodiment of the present invention, the vehicle data includes time series data corresponding to a plurality of vehicle signals, and the data preprocessing of the vehicle data to obtain a target feature map includes:

[0084] Performing the time-frequency feature extraction process on the time series data corresponding to the multiple vehicle signals to obtain a plurality of two-dimensional feature maps, each of which includes the time domain features and frequency domain features of the corresponding vehicle signals;

[0085] The multiple two-dimensional feature maps are spliced together to obtain a three-dimensional feature map of the vehicle data, and the target feature map includes the three-dimensional feature map.

[0086] In this embodiment, the vehicle data includes time series data of multiple vehicle signals. Time-frequency feature extraction can be performed on the time series data of each vehicle signal to obtain a two-dimensional feature map for each vehicle signal. The two-dimensional feature map is a feature map composed of time domain features (usually corresponding to columns) and frequency domain features (usually corresponding to rows). The time-frequency feature extraction process is the feature extraction process based on continuous wavelet transform, and the specific formula is as follows:

[0087]

[0088] Among them, in the above formula:

[0089] In the nth time period, vehicle signal i is at scale a m and pan b l The wavelet coefficient matrix at , the rows correspond to the scale a m (M, customizable); columns correspond to translations (a total of N time points, i.e., the data collection time points within the nth time period), with a total of M rows and N columns.

[0090] The normalized signal value in the above formula.

[0091] a m : The mth scale parameter, a m >0, the value range is determined by the scale range.

[0092] Minimum scale where f c is the center frequency of the wavelet. This embodiment uses the Mexican Hat Wavelet, f c ≈0.25HZ,f max is the maximum frequency of the signal in the window,

[0093] Minimum scale where f c is the center frequency of the wavelet. This embodiment uses the Mexican hat wavelet, f c ≈0.25HZ,f min is the minimum frequency of the signal in the window,

[0094] a m In the interval [a min , a max ]According to the arithmetic progression, take M=50.

[0095] b l : the lth translation parameter, l = 0, 1, 2…, N-1.

[0096] Δt: sampling interval, 1 second.

[0097] Values of the Mexican Hat wavelet function at corresponding scales and translations.

[0098] N: number of sampling points in time period n, 50.

[0099] After performing time-frequency feature extraction processing on the time series data of each of the above vehicle signals, multiple two-dimensional feature maps are obtained. By splicing the multiple two-dimensional feature maps and stacking them on the channel, a three-dimensional feature map can be obtained.

[0100] In the embodiment of the present application, the time-frequency images of multiple signals are fused by channel stacking, which can fully utilize the features of multiple signals and thus improve the accuracy of collision event recognition.

[0101] In one embodiment of the present invention, before splicing the multiple two-dimensional feature maps to obtain the three-dimensional feature map of the vehicle data, the method further includes:

[0102] Performing standardization processing on the multiple two-dimensional feature maps to obtain the multiple two-dimensional feature maps after standardization processing, wherein the standardization processing includes performing modulo processing and normalization processing on each feature in the two-dimensional feature map, and the normalization processing is used to map each feature after the modulo processing to a grayscale value within a preset range;

[0103] The step of splicing the multiple two-dimensional feature maps to obtain a three-dimensional feature map of the vehicle data includes:

[0104] The multiple two-dimensional feature maps after normalization are spliced together to obtain a three-dimensional feature map of the vehicle data.

[0105] In this embodiment, after obtaining the above-mentioned multiple two-dimensional feature maps, each feature in the two-dimensional feature map can be modulo processed and normalized in turn, and then the processed multiple two-dimensional feature maps can be spliced to obtain a three-dimensional feature map.

[0106] The formula for the above modulo processing is as follows:

[0107]

[0108] In the above formula, Represents the absolute value of the wavelet coefficient (non-negative real number); Represents wavelet coefficients, real or complex.

[0109] The wavelet coefficient matrix obtained by the above formula 2 is a matrix in which each element in the matrix obtains the modulus of the corresponding wavelet coefficient by taking the absolute value through formula 3.

[0110] In the above normalization process, the preset grayscale value range is usually [0, 255]. The normalization formula is as follows:

[0111]

[0112] in, Represents the normalized grayscale image matrix, ranging from [0, 255], with an image size of M×N, where M is the number of scales in the matrix and N is the number of sampling points in the time period; represents the absolute value of the wavelet coefficient; S min Represents the minimum value in the matrix; S max Represents the maximum value in the matrix.

[0113] After obtaining the above-mentioned grayscale image matrix, that is, after obtaining the standardized two-dimensional graphic matrix, it can be spliced to splice the grayscale image matrix arrays of multiple signals in the above-mentioned time period n into a three-dimensional feature map, where the shape of the three-dimensional feature map is Len×M×N height, Len is the number of signals, M is the number of scales, and N is the number of time points in the time period.

[0114] In the embodiment of the present application, the above-mentioned two-dimensional feature maps are sequentially modulo-processed and normalized, and then stacked in the channel dimension to form a three-dimensional feature map, which can make the input data of the target model more standardized, thereby improving the convergence speed of the model.

[0115] In one embodiment of the present invention, see Figure 2 The target model includes a feature extraction network, an attention network 240, a fully connected network 250, and a classification network 260. Inputting the target feature map into the target model to obtain a collision recognition result of the target vehicle in the preset time period includes:

[0116] Inputting the target feature map into a feature extraction network for feature extraction processing to obtain a first feature map, wherein the feature extraction processing includes a convolution operation;

[0117] Inputting the first feature map into the attention network 240 to obtain a second feature map determined based on a target attention mechanism, wherein the target attention mechanism includes at least one of a channel attention mechanism and a spatial attention mechanism;

[0118] Inputting the second feature map into the fully connected network 250 to obtain classification data;

[0119] The classification data is input into the classification network 260 to obtain a collision recognition result of the target vehicle in the preset time period.

[0120] During the model building phase, a deep learning model including the above-mentioned feature extraction network, attention network 240, fully connected network 250 and classification network 260 is constructed; during the model training phase, the parameters in the above-mentioned network are adjusted, and the deep learning model after the training is completed is the above-mentioned target model.

[0121] The input layer network of the target model is a feature extraction network, which is mainly used to extract the features of the input data of the target model. The feature extraction network includes a convolution layer, and the feature extraction network performs feature extraction processing, including convolution operations. In some embodiments, the feature extraction network may also include a nonlinear activation layer, which increases the expression ability of the entire network by introducing an activation function (such as a ReLU function), that is, introducing nonlinear characteristics. The output result of the convolution layer passes through the nonlinear activation layer, and the activation function converts the result after linear transformation into a nonlinear output, thereby increasing the expression ability of the network and enabling the network to learn and approximate various complex functional relationships.

[0122] The lower layer network of the feature extraction network is the attention network 240, which includes at least one of a network applying a channel attention mechanism and a network applying a spatial attention mechanism.

[0123] The channel attention mechanism focuses on the importance of each channel in the feature map. It learns the weight of each channel, giving higher weights to more meaningful channels and suppressing irrelevant or redundant channels.

[0124] Spatial Attention Mechanism,The spatial attention mechanism focuses on the importance of different spatial locations in the feature map.,It learns a spatial weight distribution to highlight the key areas at specific locations.

[0125] The lower layer of the attention network 240 is the fully connected network 250, which includes at least one fully connected layer. Its function is to receive the feature map from the attention network 240 and calculate the original score of each category (collision and no collision) through the weight matrix and bias term, that is, to obtain the above-mentioned classification data.

[0126] The underlying network of fully connected network 250 is classification network 260. Classification network 260 can apply a softmax function. The raw scores output by fully connected network 250 serve as the input to the softmax function, which normalizes these raw scores and other processes to obtain a probability distribution for each category. Once this probability distribution is obtained, the category with the highest probability is selected as the collision recognition result for the target model.

[0127] In an embodiment of the present application, during the model construction stage, the model is constructed including the attention network 240 that applies the above-mentioned channel attention and / or spatial attention, which is conducive to highlighting key features and improving model performance.

[0128] In one embodiment of the present invention, the feature extraction network is a multi-layer feature extraction network, and the multi-layer feature extraction network is used to implement multi-level feature extraction processing from low-level features to high-level features.

[0129] In this embodiment, each layer of the feature extraction network may include a corresponding convolutional layer for extracting features, and in each layer of the feature extraction network, the number of convolution kernels in the convolutional layer increases as the feature extraction becomes richer. Each layer of the feature extraction network may include a single convolutional layer for feature extraction, or multiple convolutional layers for feature extraction, which is not limited in this embodiment of the present application.

[0130] The above-mentioned multi-layer feature extraction network is used to realize multi-level feature extraction processing from low-level features to high-level features. The multi-layer feature extraction network may, for example, include a primary feature extraction network 210, an intermediate feature extraction network 220, and a high-level feature extraction network 230. In some embodiments, the primary feature extraction network 210 and the high-level feature extraction network 230 may also be directly included. The specific network structure can be constructed according to needs.

[0131] For example, a multi-layer feature extraction network may include a primary feature extraction network 210, a mid-level feature extraction network 220, and a high-level feature extraction network 230. Figure 2 The primary feature extraction network 210 may include a first convolution layer and a second convolution layer (the specific number is constructed according to the extraction depth and computing resources, etc.), the first convolution layer and the second convolution layer are both used for feature extraction, both include 64 convolution kernels, and the convolution kernel size may be 3×3. The intermediate feature extraction network 220 may include a third convolution layer and a fourth convolution layer, the third convolution layer and the fourth convolution layer are both used for feature extraction, both include 128 convolution kernels, and the convolution kernel size may be 3×3. The advanced feature extraction network 230 may include a fifth convolution layer and a sixth convolution layer, the fifth convolution layer and the sixth convolution layer are both used for feature extraction, both include 256 convolution kernels, and the convolution kernel size may be 3×3.

[0132] In the embodiment of the present application, the multi-layer feature extraction network is beneficial to enrich feature extraction.

[0133] In one embodiment of the present invention, the feature extraction network includes a first feature extraction network, and the first feature extraction network includes:

[0134] a first feature extraction layer, the first feature extraction layer being used to extract features of input data of the first feature extraction network;

[0135] a first channel adjustment layer, wherein the number of output channels of the first channel adjustment layer is the same as the number of output channels of the first feature extraction network, and the first channel adjustment layer is used to adjust the channels of input data of the first feature extraction network;

[0136] The first feature extraction layer and the first channel adjustment layer use a residual connection to add the output data of the first feature extraction layer and the output data of the first channel adjustment layer.

[0137] In the case where the feature extraction network includes multiple layers of feature extraction networks, the first feature extraction network may be any layer of the feature extraction network in the multiple layers of the feature extraction network.

[0138] The first feature extraction network includes a first feature extraction layer, and the first feature extraction layer includes a convolution layer for extracting features of input data. For example, the first feature extraction network can be a primary feature extraction network 210, and the first feature extraction layer can be Figure 2 The first convolution layer and the second convolution layer in the network; For example, the first feature extraction network may be an intermediate feature extraction network 220, and the first feature extraction layer may be Figure 2 The third and fourth convolutional layers in .

[0139] The first channel adjustment layer is used to adjust the channels of the input data of the first feature extraction network. The number of output channels of the first channel adjustment layer is the same as the number of output channels of the first feature extraction layer. The first channel adjustment layer can be implemented by 1×1 convolution, that is, the convolution kernel size of the convolution layer in the first channel adjustment layer is set to 1×1, and the number is the same as the number of convolution kernels of the convolution layer in the first feature extraction layer, so that the output channels are the same. For example, Figure 2 As shown, in the primary feature extraction network 210, the convolution kernel size in the convolution layer used for feature extraction is 3×3, and the number is 64; the convolution kernel size in the convolution layer used for channel adjustment is 1×1, and the number is 64.

[0140] The first feature extraction network and the first channel adjustment layer use a residual connection, so that the output data of the first feature extraction network and the output data of the first channel adjustment layer can be added at the element level. In this embodiment of the application, by performing channel adjustment on the input data of the feature extraction network and applying a residual connection, it is beneficial to improve the training efficiency of the model during the training process.

[0141] In some embodiments, the output of the convolutional layer used for feature extraction may be sequentially subjected to a ReLU activation function and batch normalization. The activation function introduces nonlinearity, enabling the network to learn complex feature representations and decision boundaries; batch normalization stabilizes the network's input distribution.

[0142] In some embodiments, the feature map after residual connection can also pass through the ReLU activation function and then be input into the next layer.

[0143] In some embodiments, the output layer of the feature extraction network may include a maximum pooling layer. After the residual connection feature map passes through the ReLU activation function, it passes through the maximum pooling layer, which reduces the size of the feature before inputting it into the next layer of feature extraction network or attention network 240.

[0144] In one embodiment of the present invention, the target model is trained by the following steps:

[0145] Obtain vehicle data of the target vehicle in different time periods;

[0146] Performing data preprocessing on the vehicle data of the different time periods to obtain characteristic graphs of the different time periods;

[0147] Determine a label of the feature graph of each time period in the feature graphs of different time periods;

[0148] Based on the feature graphs of the different time periods and the labels of the feature graphs of each time period, the deep learning model is trained to adjust the parameters of the deep learning model;

[0149] When the model training meets the preset requirements, the target model is obtained.

[0150] The vehicle data during the training process is of the same type as the vehicle data within the above-mentioned preset time period, and the vehicle data includes time series data of at least one vehicle signal, such as the longitudinal acceleration and lateral acceleration of the vehicle.

[0151] Acquiring vehicle data in different time periods can be achieved through a sliding window, that is, sliding the window in the time series data of each vehicle signal respectively, and each window corresponds to the sampling data of a vehicle signal in a time period (or time segment).

[0152] By performing the aforementioned data preprocessing on the vehicle data for the different time periods, feature maps for the different time periods can be obtained. The feature maps for each time period serve as the input data for model training. The data preprocessing process can be referenced to the aforementioned process for processing vehicle data for the preset time periods. To avoid repetition, this process will not be further described.

[0153] Label the feature graph for each time period. For example, if a collision occurs within a time period, the sample can be labeled as a positive sample, with the corresponding feature graph label being 1. If no collision occurs within a time period, the sample can be labeled as a negative sample, with the corresponding feature graph label being 0.

[0154] Based on the feature maps of the above-mentioned different time periods and the labels of the feature maps of each time period, the deep learning model can be trained. When the training meets the preset requirements (for example, the training time meets the requirements or the loss function converges, etc.), the above-mentioned target model can be obtained.

[0155] In order to more clearly understand the technical solutions of the embodiments of the present application, the model training is described in detail below with reference to specific examples.

[0156] a. Get time segments

[0157] a1. Obtain data from the cloud, including various vehicle signals such as longitudinal acceleration and lateral acceleration. In this example, there are 30 types of signals.

[0158] a2. For each signal, use a sliding window approach to obtain multiple time segments.

[0159] b. Normalize the signal within the window with reference to the above formula 1, where the vehicle signal i in the nth time period in the formula annotation can be understood as the nth window of the vehicle signal i in this example.

[0160] c. Calculate the wavelet coefficient matrix of the sampled data in each vehicle signal window according to the above formula 2, where the number of sampling points N in the window is 50 and the scale a m The number M is 50, b l The number of sampling points in the window is 50.

[0161] d. Referring to the above formula 3, each element in the wavelet coefficient matrix is taken to determine the modulus of the corresponding wavelet coefficient by taking the absolute value.

[0162] e. Referring to the above formula 4, for each wavelet coefficient matrix Each element is normalized and mapped to a grayscale value of [0,255].

[0163] f. Construct a multi-channel input matrix

[0164] f1. For each sliding window n: concatenate the wavelet coefficient matrices of multiple signals into a tensor with a shape of Len × M × N height: Len is the number of signals, M is the number of scales, and N is the number of time points in the sliding window.

[0165] f2. Sample label assignment. Positive sample (label 1): If a collision event occurs within the time range of the sliding window, the sample is marked as a positive sample. Negative sample (label 0): If no collision event occurs within the sliding window, it is marked as a negative sample.

[0166] g. Build a deep learning model

[0167] This example takes as input a three-dimensional tensor of size 50×50×30, where the height and width are 50, representing the scale and time dimensions, respectively, and the number of channels is 30, representing the number of vehicle signal types.

[0168] like Figure 2 As shown, the deep learning model includes a primary feature extraction network 210, a mid-level feature extraction network 220, a high-level feature extraction network 230, an attention network 240, a fully connected network 250 and a classification network 260.

[0169] g1. Initial feature extraction

[0170] First, the input data passes through the first convolutional layer, which uses a 3×3 kernel, 64 kernels, a stride of 1, and "same" padding to maintain the feature map size. The input shape is 50×50×30, and the output shape is 50×50×64. The layer then passes through a ReLU activation function. The output then undergoes batch normalization to stabilize the training process.

[0171] Next, after the second convolutional layer, we use the same 3×3 kernel size, 64 kernels, stride 1, “same” padding, input and output shapes are both 50×50×64, ReLU activation function, and batch normalization again.

[0172] To implement the residual connection, a 1×1 convolution is performed on the original input on the residual path, adjusting the number of channels from 30 to 64, with an input shape of 50×50×30 and an output shape of 50×50×64. The output of the residual path is then added to the output of the second convolutional layer at the element level to obtain a feature map of shape 50×50×64, which is then activated by the ReLU function.

[0173] Subsequently, after a maximum pooling layer with a size of 2×2 and a stride of 2, the feature map size is reduced to 25×25×64, completing the initial feature extraction.

[0174] g2. Intermediate feature extraction

[0175] The pooled feature map is fed into the third convolutional layer, using a 3×3 kernel, 128 kernels, a stride of 1, and "same" padding, resulting in an output shape of 25×25×128. It is then activated by a ReLU function and batch normalized. The fourth convolutional layer has the same structure as the third convolutional layer, with both input and output shapes of 25×25×128, also using the ReLU activation function and batch normalization.

[0176] In the residual connection, a 1×1 convolution is performed on the output of the initial feature extraction to adjust the number of channels from 64 to 128, with an input shape of 25×25×64 and an output shape of 25×25×128. The output of the residual path is added to the output of the fourth convolutional layer to obtain a feature map of shape 25×25×128, which is then activated by the ReLU function.

[0177] Next, the feature map size is reduced to 12×12×128 through a maximum pooling layer with a size of 2×2 and a stride of 2 to complete the extraction of intermediate features.

[0178] g3. Advanced feature extraction

[0179] The pooled feature map is input into the fifth convolutional layer, using a convolution kernel of size 3×3, 256 convolution kernels, a stride of 1, “same” padding, an output shape of 12×12×256, and a ReLU activation function, followed by batch normalization.

[0180] Then, after the sixth convolutional layer, the structure is the same as the fifth convolutional layer, with input and output shapes of 12×12×256, and the same ReLU activation function and batch normalization are used.

[0181] In the residual connection, a 1×1 convolution is performed on the output of the mid-level feature extraction, adjusting the number of channels from 128 to 256, with an input shape of 12×12×128 and an output shape of 12×12×256. The output of the shortcut path is added to the output of the sixth convolutional layer to obtain a feature map of shape 12×12×256, which is then activated by the ReLU function.

[0182] Subsequently, the feature map size is reduced to 6×6×256 through a maximum pooling layer with a size of 2×2 and a stride of 2 to complete the extraction of high-level features.

[0183] g4. Attention Network 240

[0184] After high-level feature extraction, an attention mechanism is introduced to enhance the model's focus on important features. First, a channel-wise attention mechanism (SE module) is applied to the feature map, performing global average pooling in the spatial dimension to produce a feature vector of shape 1×1×256. Next, two fully connected layers are applied. The first layer reduces the number of channels from 256 to 16 (compression ratio r=16) using the ReLU activation function; the second layer increases the number of channels from 16 to 256 using the Sigmoid activation function to obtain channel weights. The channel weights are then multiplied by the original feature map channel by channel to achieve channel recalibration.

[0185] Next, the spatial attention mechanism is applied. The feature maps that have undergone channel attention are subjected to global max pooling and average pooling in the channel dimension, resulting in two feature maps of shape 6×6×1. These two feature maps are concatenated in the channel dimension to form a tensor of shape 6×6×2. This is then passed through a convolutional layer with 1 kernel, size 3×3, stride 1, and "same" padding, using a sigmoid activation function, resulting in spatial weights of shape 6×6×1. The spatial weights are element-wise multiplied with the feature map to complete the spatial attention mechanism.

[0186] g5. Fully connected network 250 and classification network 260

[0187] The feature map after the attention mechanism is flattened into a one-dimensional vector with a shape of 6×6×256=9216. Next, it enters the first fully connected layer with 4096 neurons, the ReLU activation function, and dropout regularization with a dropout rate of 0.5. Next, it enters the second fully connected layer with 1024 neurons, the ReLU activation function, and dropout regularization with a dropout rate of 0.5. Next, it enters the third fully connected layer with 2 neurons and the ReLU activation function.

[0188] Finally, classification is performed through the softmax layer (classification network 260, also known as the output layer).

[0189] The target model can be obtained by training the above deep learning model based on samples consisting of feature maps and labels.

[0190] This application converts one-dimensional vehicle sensor time series signals into two-dimensional time-frequency images based on continuous wavelet transform, which can capture the time-frequency characteristics of the signals, especially the transient characteristics of collision events. This application stacks the time-frequency images of multiple sensor signals in the channel dimension, which can be applied to multivariate time series classification tasks and is conducive to providing characteristic information of the target model. This application combines residual networks, channel attention, and spatial attention mechanisms to improve the training efficiency and performance of the model.

[0191] It should be understood that the order of execution of the steps in the above embodiments does not necessarily mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0192] Figure 3 FIG. 1 is a block diagram of a vehicle collision recognition device according to an exemplary embodiment of the present application. Figure 3 As shown, the vehicle collision recognition device includes:

[0193] An acquisition module 310 is used to acquire vehicle data of a target vehicle in a preset time period;

[0194] A data preprocessing module 320 is configured to perform data preprocessing on the vehicle data to obtain a target feature map, wherein the target feature map includes time domain features and frequency domain features of the vehicle data. The data preprocessing includes time-frequency feature extraction based on continuous wavelet transform.

[0195] The determination module 330 is used to input the target feature map into the target model to obtain the collision recognition result of the target vehicle in the preset time period, and the target model is a trained deep learning model.

[0196] In one embodiment of the present invention, the vehicle data includes time series data corresponding to a plurality of vehicle signals, and the data preprocessing module 310 includes:

[0197] a time-frequency feature extraction and processing unit, configured to perform the time-frequency feature extraction processing on the time series data corresponding to the multiple vehicle signals, respectively, to obtain a plurality of two-dimensional feature maps, each of which includes a time domain feature and a frequency domain feature of the corresponding vehicle signal;

[0198] A splicing unit is used to splice the multiple two-dimensional feature maps to obtain a three-dimensional feature map of the vehicle data, and the target feature map includes the three-dimensional feature map.

[0199] In one embodiment of the present invention, the data pre-processing module 310 further includes:

[0200] a normalization processing unit, configured to perform normalization processing on the plurality of two-dimensional feature maps to obtain the plurality of normalized two-dimensional feature maps, wherein the normalization processing includes performing modulo processing and normalization processing on each feature in the two-dimensional feature map, and the normalization processing is configured to map each feature after the modulo processing to a grayscale value within a preset range;

[0201] The splicing unit is specifically used for:

[0202] The multiple two-dimensional feature maps after normalization are spliced together to obtain a three-dimensional feature map of the vehicle data.

[0203] In one embodiment of the present invention, the target model includes a feature extraction network, an attention network, a fully connected network, and a classification network, and the determination module 330 is specifically configured to:

[0204] Inputting the target feature map into a feature extraction network for feature extraction processing to obtain a first feature map, wherein the feature extraction processing includes a convolution operation;

[0205] Inputting the first feature map into the attention network to obtain a second feature map determined based on a target attention mechanism, wherein the target attention mechanism includes at least one of a channel attention mechanism and a spatial attention mechanism;

[0206] Inputting the second feature map into the fully connected network to obtain classification data;

[0207] The classification data is input into the classification network to obtain a collision recognition result of the target vehicle in the preset time period.

[0208] In one embodiment of the present invention, the feature extraction network is a multi-layer feature extraction network, and the multi-layer feature extraction network is used to implement multi-level feature extraction processing from low-level features to high-level features.

[0209] In one embodiment of the present invention, the feature extraction network includes a first feature extraction network, and the first feature extraction network includes:

[0210] a first feature extraction layer, the first feature extraction layer being used to extract features of input data of the first feature extraction network;

[0211] a first channel adjustment layer, wherein the number of output channels of the first channel adjustment layer is the same as the number of output channels of the first feature extraction network, and the first channel adjustment layer is used to adjust the channels of input data of the first feature extraction network;

[0212] The first feature extraction layer and the first channel adjustment layer use a residual connection to add the output data of the first feature extraction layer and the output data of the first channel adjustment layer.

[0213] In one embodiment of the present invention, the target model is trained by the following steps:

[0214] Obtain vehicle data of the target vehicle in different time periods;

[0215] Performing data preprocessing on the vehicle data of the different time periods to obtain characteristic graphs of the different time periods;

[0216] Determine a label of the feature graph of each time period in the feature graphs of different time periods;

[0217] Based on the feature graphs of the different time periods and the labels of the feature graphs of each time period, the deep learning model is trained to adjust the parameters of the deep learning model;

[0218] When the model training meets the preset requirements, the target model is obtained.

[0219] It should be noted that the vehicle collision recognition device provided in the above-described embodiment and the vehicle collision recognition method provided in the above-described embodiment are based on the same concept. The specific manner in which the various modules and units perform their operations has been described in detail in the method embodiments and will not be repeated here. In actual applications, the vehicle collision recognition device provided in the above-described embodiment can, as needed, allocate the aforementioned functions to different functional modules, i.e., divide the internal structure of the device into different functional modules to perform all or part of the functions described above. This is not a limitation herein.

[0220] An embodiment of the present application also provides an electronic device, comprising: one or more processors; a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the electronic device implements the vehicle collision recognition method provided in the above-mentioned embodiments.

[0221] Figure 4 The following is a schematic diagram showing the structure of a computer system suitable for implementing an electronic device according to an embodiment of the present application. Figure 4 The computer system 400 of the electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0222] like Figure 4 As shown, the computer system 400 includes a central processing unit (CPU) 401, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 402 or the program loaded from the storage part 408 into the random access memory (RAM) 403, such as executing the method described in the above embodiment. Various programs and data required for system operation are also stored in the RAM 403. The CPU 401, ROM 402 and RAM 403 are connected to each other via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0223] The following components are connected to the I / O interface 405: an input section 406 including a keyboard, a mouse, and the like; an output section 407 including devices such as a cathode ray tube (CRT), a liquid crystal display (LCD), and a speaker; a storage section 408 including a hard disk and the like; and a communication section 409 including a network interface card such as a LAN (Local Area Network) card or a modem. The communication section 409 performs communication processing via a network such as the Internet. A drive 410 is also connected to the I / O interface 405 as needed. Removable media 411, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is installed in the drive 410 as needed, so that computer programs read therefrom can be installed into the storage section 408 as needed.

[0224] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 409, and / or installed from a removable medium 411. When the computer program is executed by the central processing unit (CPU) 401, the various functions defined in the system of the present application are executed.

[0225] It should be noted that the computer-readable medium shown in the embodiments of the present application can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media can include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, which carries a computer-readable computer program. This propagated data signal can take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. A computer program embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, or any suitable combination thereof.

[0226] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. Among them, each box in the flowchart or block diagram can represent a module, program segment, or part of the code, and the above-mentioned module, program segment, or part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0227] The units involved in the embodiments described in this application may be implemented by software or hardware, and the units described may also be set in a processor. In some cases, the names of these units do not constitute limitations on the units themselves.

[0228] Another aspect of the present application provides a computer-readable storage medium having a computer program stored thereon. When executed by a computer processor, the computer program causes the computer to perform the vehicle collision identification method described above. The computer-readable storage medium may be included in the electronic device described in the above embodiments, or may exist independently and not be incorporated into the electronic device.

[0229] Another aspect of the present application provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the vehicle collision identification method provided in each of the above embodiments.

[0230] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, any equivalent modifications or alterations made by one of ordinary skill in the art without departing from the spirit and technical principles disclosed herein are intended to be covered by the claims of the present invention.

Claims

1. A vehicle collision recognition method, characterized in that: include: Obtain vehicle data of the target vehicle in a preset time period; Performing data preprocessing on the vehicle data to obtain a target feature map, wherein the target feature map includes time domain features and frequency domain features of the vehicle data, and the data preprocessing includes performing time-frequency feature extraction processing based on continuous wavelet transform; The target feature map is input into a target model to obtain a collision recognition result of the target vehicle in the preset time period, and the target model is a trained deep learning model.

2. The method according to claim 1, characterized in that The vehicle data includes time series data corresponding to a plurality of vehicle signals. The data preprocessing is performed on the vehicle data to obtain a target feature map, including: Performing the time-frequency feature extraction process on the time series data corresponding to the multiple vehicle signals to obtain a plurality of two-dimensional feature maps, each of which includes the time domain features and frequency domain features of the corresponding vehicle signals; The multiple two-dimensional feature maps are spliced together to obtain a three-dimensional feature map of the vehicle data, and the target feature map includes the three-dimensional feature map.

3. The method according to claim 2, characterized in that Before splicing the multiple two-dimensional feature maps to obtain the three-dimensional feature map of the vehicle data, the method further includes: Performing standardization processing on the multiple two-dimensional feature maps to obtain the multiple two-dimensional feature maps after standardization processing, wherein the standardization processing includes performing modulo processing and normalization processing on each feature in the two-dimensional feature map, and the normalization processing is used to map each feature after the modulo processing to a grayscale value within a preset range; The step of splicing the multiple two-dimensional feature maps to obtain a three-dimensional feature map of the vehicle data includes: The multiple two-dimensional feature maps after normalization are spliced together to obtain a three-dimensional feature map of the vehicle data.

4. The method according to claim 1, wherein The target model includes a feature extraction network, an attention network, a fully connected network, and a classification network. Inputting the target feature map into the target model to obtain a collision recognition result of the target vehicle in the preset time period includes: Inputting the target feature map into a feature extraction network for feature extraction processing to obtain a first feature map, wherein the feature extraction processing includes a convolution operation; Inputting the first feature map into the attention network to obtain a second feature map determined based on a target attention mechanism, wherein the target attention mechanism includes at least one of a channel attention mechanism and a spatial attention mechanism; Inputting the second feature map into the fully connected network to obtain classification data; The classification data is input into the classification network to obtain a collision recognition result of the target vehicle in the preset time period.

5. The method according to claim 4, characterized in that The feature extraction network is a multi-layer feature extraction network, which is used to implement multi-level feature extraction processing from low-level features to high-level features.

6. The method according to claim 4 or 5, characterized in that The feature extraction network includes a first feature extraction network, and the first feature extraction network includes: a first feature extraction layer, the first feature extraction layer being used to extract features of input data of the first feature extraction network; a first channel adjustment layer, wherein the number of output channels of the first channel adjustment layer is the same as the number of output channels of the first feature extraction network, and the first channel adjustment layer is used to adjust the channels of input data of the first feature extraction network; The first feature extraction layer and the first channel adjustment layer use a residual connection to add the output data of the first feature extraction layer and the output data of the first channel adjustment layer.

7. The method according to claim 1, characterized in that The target model is trained through the following steps: Obtain vehicle data of the target vehicle in different time periods; Performing data preprocessing on the vehicle data of the different time periods to obtain characteristic graphs of the different time periods; Determine a label of the feature graph of each time period in the feature graphs of different time periods; Based on the feature graphs of the different time periods and the labels of the feature graphs of each time period, the deep learning model is trained to adjust the parameters of the deep learning model; When the model training meets the preset requirements, the target model is obtained.

8. A vehicle collision recognition device, characterized in that: include: An acquisition module is used to acquire vehicle data of a target vehicle in a preset time period; A data preprocessing module, configured to perform data preprocessing on the vehicle data to obtain a target feature map, wherein the target feature map includes time domain features and frequency domain features of the vehicle data, wherein the data preprocessing includes time-frequency feature extraction based on continuous wavelet transform; A determination module is used to input the target feature map into a target model to obtain a collision recognition result of the target vehicle in the preset time period, and the target model is a trained deep learning model.

9. A device, characterized in that include: one or more processors and memory, A computer program is stored in the memory, and when the one or more processors execute the computer program, the device executes the vehicle collision identification method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon, which, when executed by one or more processors, enables the device to perform the vehicle collision identification method according to any one of claims 1 to 7.

Images