A Vehicle Behavior Recognition Method and System Based on Double-Layer GRU and SimAM

Abstract

The present invention provides a vehicle behavior recognition method and system based on double-layer GRU and SimAM, including: obtaining vehicle motion trajectory data and splitting it into trajectory segments through a sliding window; calculating the change in motion position and the change in motion direction for each trajectory segment to construct a new description vector; training the DGRU-SimAM network model according to sample data during the training phase; using the trained DGRU-SimAM network model to obtain the vehicle behavior category corresponding to the trajectory segment during the inference phase; and combining the behavior recognition results of each trajectory segment into a final behavior result sequence. The present invention uses a sliding window for trajectory splitting, realizes a secondary judgment of the splitting position area, and avoids the problem of discontinuous behavior states caused by improper splitting positions. The DGRU-SimAM network model is used to realize the recognition of vehicle behaviors. The double-layer GRU and SimAM modules used in the DGRU-SimAM network model can enable the network to focus on effective local information and improve the network's feature learning ability.

Description

Technical Field

[0001] The present invention relates to the fields of intelligent transportation and behavior recognition, and particularly to a vehicle behavior recognition method and system based on a double-layer GRU and SimAM. Background Art

[0002] With the increasing development of technology and the continuous improvement of people's living standards, the ownership of civilian cars in China has increased rapidly. Strengthening the behavior supervision of vehicles on the road is of great significance for improving the driving efficiency of vehicles and the safety of road traffic. The driving trajectory of a vehicle on the road contains the motion characteristics of the vehicle. Extracting and learning the characteristics of the driving trajectory corresponding to the vehicle behavior can discover the patterns of these fixed behaviors. Through the automatic recognition of vehicle behaviors, it can provide auxiliary decision-making support for traffic supervision and promote the development of intelligent transportation; it can provide clues for traffic accident investigation, conduct correlation reasoning with other moving targets or the surrounding environment, and achieve a more comprehensive understanding of the traffic scene.

[0003] The motion trajectory of a vehicle is a sequence of position changes generated by the vehicle under a certain spatio-temporal background. For the processing of time-series data, the recurrent neural network has more advantages than the traditional neural network. The traditional neural network model mainly consists of an input layer, a hidden layer, and an output layer. There are full connections between the neurons of each layer, and there is no connection between the nodes of each layer. By continuously learning and adjusting the weight parameters of the neurons in each layer of the network, the traditional neural network model can achieve one-to-one and one-to-many tasks. However, when the input data is continuous and the multi-step input is provided by sequence data, the traditional neural network model only processes each step of the input independently and is difficult to extract sequence features. At the same time, the output corresponding to the sequence data is not only related to the current input but also related to the previous input. The recurrent neural network consists of an input layer, a hidden layer, and an output layer. The hidden layer has a structure with shared weights, which can realize the transmission of the hidden layer state at time steps, making the neuron nodes between the hidden layers no longer unconnected to each other, and is more suitable for solving sequence problems.

[0004] When people observe an object, they often focus on the more discriminative feature parts. According to this visual attention characteristic of humans, the attention mechanism focuses on local feature information, assigns a larger weight to strong features, and a smaller weight to weak features in the neural network. The weight parameters of the attention mechanism are usually obtained through network training. During network inference, the neural network can focus on calculating useful feature vectors and ignore unimportant feature vectors. Applying the attention module to the recurrent neural network, by calculating the weights of the feature vectors output by the network hidden layer at each time step and assigning different weight coefficients to the hidden layer neurons at different time steps, the feature extraction and learning ability of the recurrent neural network can be improved. Summary of the Invention

[0005] To solve the above problems, the present invention mainly aims at problems such as the recognition and semantic description of vehicle behaviors in intelligent transportation and spatial data mining, and provides a vehicle behavior recognition method and system based on a double-layer GRU and SimAM. The trajectory feature extraction and learning are carried out through a network model composed of a double-layer GRU and SimAM module, and the behavior category corresponding to the vehicle trajectory is recognized in the inference stage. A vehicle behavior recognition method based on a double-layer GRU and SimAM mainly includes:

[0006] S1: Record the vehicle's motion trajectory, and split the vehicle's motion trajectory into multiple trajectory segments through a sliding window;

[0007] S2: For each trajectory segment, use a sequence composed of vectors to describe, where is used to represent the position offset of the vehicle in the x direction during the time interval from the i-th moment to the (i - 1)-th moment, is used to represent the position offset of the vehicle in the y direction during the time interval from the i-th moment to the (i - 1)-th moment, is used to represent the change in the vehicle's motion direction during the time interval from the i-th moment to the (i - 1)-th moment;

[0008] S3: Construct a DGRU-SimAM network model, which includes stacked double-layer GRU units, SimAM modules, fully connected layers, and Softmax layers;

[0009] S4: Input the description vector sequence corresponding to each trajectory segment into the double-layer GRU unit in sequence to obtain output features. Concatenate the output features X at the time step, calculate the attention weights through the SimAM module, and input the output features after being weighted by the SimAM module into the fully connected layer and Softmax layer to obtain the output vector <cls, conf>, which represents the possibility of the trajectory segment corresponding to various vehicle behaviors. Among them, cls represents the behavior category, and conf represents the confidence;

[0010] S5: Use the trajectory segments with known behavior category labels corresponding to various vehicle behaviors to form sample data, and train the DGRU-SimAM network model according to this sample data to obtain a converged network model;

[0011] S6: Input the trajectory segments obtained by splitting the actually obtained vehicle motion trajectory into the trained DGRU-SimAM network model to obtain the corresponding vehicle behavior category;

[0012] S7: According to the order corresponding to each split trajectory segment, combine the obtained vehicle behavior category labels into a vehicle behavior sequence corresponding to the trajectory to obtain the semantic description of the behavior, that is, realize the recognition of vehicle behavior.

[0013] Further, in step S1, use in-vehicle sensors or visual target tracking technology to obtain the vehicle's motion trajectory, and use a sliding window with a certain window length and sliding step to split the trajectory, and intercept to generate multiple trajectory segments.

[0014] Further, in step S2, calculate using the following formula

[0015]

[0016]

[0017]

[0018] where, x i-1 , y i-1 respectively represent the x-position coordinate and y-position coordinate of the vehicle at the (i - 1)th moment, x i , y i respectively represent the x-position coordinate and y-position coordinate of the vehicle at the ith moment, x i-2 , y i-2 respectively represent the x-position coordinate and y-position coordinate of the vehicle at the (i - 2)th moment, i represents a positive integer greater than or equal to 2, x max and x min represent the maximum and minimum values of the x-position coordinate, y max and y min represent the maximum and minimum values of the y-position coordinate, arctan represents the arctangent function.

[0019] Further, in step S4, when calculating the attention weight using the SimAM module, use the following formula to calculate the minimum energy of neuron t at time step j

[0020]

[0021] where, λ is the regularization term parameter, and the empirical value is set to 0.00001, and are:

[0022]

[0023]

[0024] where, M represents the number of hidden layer neurons, T represents the time span of the sequence, t(j) and respectively represent the states of the target neuron and other neurons at time step j;

[0025] After calculating the minimum energy of M neurons over T time steps, the corresponding energy matrix E is constructed, and the output feature X is weighted with this to obtain the weighted result of the output feature

[0026]

[0027] where sigmoid is the activation function, and ⊙ represents the dot product operation of matrices.

[0028] Furthermore, in step S7, when obtaining the semantic description of the behavior, since there are overlapping parts between the trajectory segments obtained by splitting before, it is necessary to judge the behavior category labels of the overlapping parts and make them the behavior categories with higher confidence:

[0029] <cls i,j ,conf i,j > = max(<cls i ,conf i ,<cls j ,conf j )

[0030] where cls and conf represent the behavior category and the confidence corresponding to the category respectively. After determining the behavior category labels of the overlapping parts, the final behavior sequence result is formed to realize the semantic description of the behavior.

[0031] A vehicle behavior recognition system based on a double - layer GRU and SimAM, the system includes:

[0032] A trajectory splitting module, used to obtain the vehicle motion trajectory and split it into multiple trajectory segments;

[0033] A sequence description module, used to describe each trajectory segment using a sequence composed of vectors, where is used to represent the position offset of the vehicle in the x - direction within the current time interval, is used to represent the position offset of the vehicle in the y - direction within the current time interval, is used to represent the change in the motion direction of the vehicle within the current time interval;

[0034] A network training module, which is used to construct a DGRU-SimAM network model and train the DGRU-SimAM network model until convergence; the DGRU-SimAM network model includes stacked double-layer GRU units, a SimAM module, a fully connected layer, and a Softmax layer;

[0035] A network inference module, which is used to obtain the behavior recognition result corresponding to the trajectory segment after inputting it into the DGRU-SimAM network model, that is, to obtain its corresponding vehicle behavior category;

[0036] A behavior sequence generation module, which is used to combine the behavior recognition results of each trajectory segment according to the order corresponding to each split trajectory segment, combine the obtained vehicle behavior category labels into the vehicle behavior sequence corresponding to the trajectory, and obtain the semantic description of the behavior, that is, to realize the recognition of vehicle behavior.

[0037] Furthermore, in the trajectory splitting module, the movement trajectory of the vehicle is obtained by using in-vehicle sensors or visual target tracking technology, and the trajectory is split by using a sliding window with a certain window length and a sliding step to intercept and generate multiple trajectory segments.

[0038] Furthermore, in the network training module, the following formula is used for calculation

[0039]

[0040]

[0041]

[0042] where x i-1 and y i-1 respectively represent the x-position coordinate and y-position coordinate of the vehicle at the (i - 1)th moment, x i and y i respectively represent the x-position coordinate and y-position coordinate of the vehicle at the ith moment, x i-2 and y i-2 respectively represent the x-position coordinate and y-position coordinate of the vehicle at the (i - 2)th moment, i represents a positive integer greater than or equal to 2, x max and x min represent the maximum and minimum values of the x-position coordinate, y max and y min represent the maximum and minimum values of the y-position coordinate, and arctan represents the arctangent function.

[0043] Furthermore, in the network training module, when calculating the attention weight using the SimAM module, the following formula is used to calculate the minimum energy of neuron t at time step j

[0044]

[0045] Among them, λ is the regularization term parameter, and the empirical value is set to 0.00001. and are:

[0046]

[0047]

[0048] Among them, M represents the number of neurons in the hidden layer, T represents the time span of the sequence, and t (j) and respectively represent the states of the target neuron and other neurons at time step j;

[0049] After calculating the minimum energy of M neurons at T time steps, the corresponding energy matrix E is constructed, and the output feature X is weighted with this to obtain the result after weighting the output feature

[0050]

[0051] Among them, sigmoid is the activation function, and ⊙ represents the dot product operation of matrices.

[0052] Furthermore, in the behavior sequence generation module, when obtaining the semantic description of the behavior, since there are overlapping parts between the trajectory segments obtained by previous splitting, it is necessary to label the behavior category labels of the overlapping parts separately, and make them the behavior categories with greater confidence:

[0053] <cls i,j ,conf i,j > = max(<cls i ,conf i >, <cls j ,conf j )

[0054] Among them, cls and conf respectively represent the behavior category and the confidence corresponding to the category. After determining the behavior category labels of the overlapping parts, the final behavior sequence result will be formed to realize the semantic description of the behavior.

[0055] The beneficial effects brought by the technical solution provided by the present invention are:

[0056] 1. In the DGRU-SimAM model, the double-layer GRU network stacked by GRU units can improve the network's feature learning ability for sequence data through stacking processing. Adding the SimAM module to the output of the network's hidden layer can make the output focus more on useful local information without increasing the network's parameter quantity.

[0057] 2. When using a sliding window to split the trajectory sequence, there will be redundant overlapping parts between each trajectory segment, thereby realizing a secondary judgment of the splitting position area and avoiding the problem of discontinuous behavior states caused by improper splitting positions. Brief Description of the Drawings

[0058] The present invention will be further described below in conjunction with the drawings and embodiments. In the drawings:

[0059] Figure 1 is a flowchart of a vehicle behavior recognition method based on double-layer GRU and SimAM in an embodiment of the present invention.

[0060] Figure 2 is a structural diagram of the DGRU-SimAM network model in an embodiment of the present invention.

[0061] Figure 3 is a schematic diagram of the generation of the behavior sequence result in an embodiment of the present invention. Detailed Embodiments

[0062] For a clearer understanding of the technical features, objectives, and effects of the present invention, the specific embodiments of the present invention will now be described in detail with reference to the drawings.

[0063] The embodiments of the present invention provide a vehicle behavior recognition method and system based on double-layer GRU and SimAM.

[0064] Please refer to Figure 1 , Figure 1 is a flowchart of a vehicle behavior recognition method based on double-layer GRU and SimAM in an embodiment of the present invention, specifically including:

[0065] S1: Use in-vehicle sensors or vision target tracking technology to record the vehicle's motion trajectory, and split the vehicle's motion trajectory into multiple trajectory segments through a sliding window; wherein, the vehicle's motion trajectory is composed of a change sequence of the x coordinate value and y coordinate value of the vehicle on the motion plane. In this embodiment, the sampling frequency is set to 2.5 Hz, the sliding window size is set to 5 groups of data, and the sliding step is set to 2 groups of data.

[0066] S2: Describe each trajectory segment using a sequence composed of vectors, where Used to represent the position offset of the vehicle in the x - direction within the time interval from the i - th moment to the (i - 1)-th moment, Used to represent the position offset of the vehicle in the y - direction within the time interval from the i - th moment to the (i - 1)-th moment, Used to represent the change in the movement direction of the vehicle within the time interval from the i - th moment to the (i - 1)-th moment. The calculation formula is:

[0067]

[0068]

[0069]

[0070] Wherein, x i represents the x - position coordinate of the vehicle at the corresponding moment, y i represents the y - position coordinate of the vehicle at the corresponding moment, x max and x min represent the maximum and minimum values of the x - position coordinate, y max and y min represent the maximum and minimum values of the y - position coordinate, and arctan represents the arctangent function.

[0071] S3: Construct a DGRU - SimAM network model, the structure is as Figure 2 shown, including a two - layer GRU unit, a SimAM module, a fully - connected layer, and a Softmax layer;

[0072] Among them, the DGRU - SimAM model receives a sample sequence with a time - dimension length of 5. The two - layer GRU is stacked by single GRU units, and the number of hidden - layer nodes of each GRU unit is set to 64. The SimAM module receives a vector input of size 5×64, and this vector is formed by concatenating the output vectors of the two - layer GRU at time steps. In this embodiment, the input dimension of the fully - connected layer is set to 320, and the output dimension is set to 3, corresponding to three vehicle behavior categories of going straight, turning left, and turning right. After passing through the Softmax layer, a three - dimensional vector is output, representing the confidence levels corresponding to the three behavior categories respectively.

[0073] S4: Input the description - vector sequence corresponding to each trajectory segment into the two - layer GRU unit in sequence, obtain the output features, concatenate these output features at time steps, calculate the attention weights through the SimAM module, and input the output features after being weighted by the SimAM module into the fully - connected layer and the Softmax layer. The obtained output vector <cls, conf> represents the possibility of this trajectory segment corresponding to various vehicle behaviors, where cls represents the behavior category and conf represents the confidence level;

[0074] When calculating the attention weight using the SimAM module, the formula for the minimum energy of neuron t at time step j is:

[0075]

[0076] where λ is the regularization term parameter, and the empirical value is set to 0.00001, and are:

[0077]

[0078]

[0079] where M represents the number of neurons in the hidden layer, T represents the time span of the sequence, t (j) and are the states of the target neuron and other neurons at time step j, respectively. The minimum energies of M neurons at T time steps are calculated respectively and form the corresponding energy matrix E, and the features X are weighted with this to obtain the weighted result:

[0080]

[0081] where sigmoid is the activation function, and ⊙ represents the dot product operation of matrices. The weighted features are input into the fully connected layer, and then Softmax is used for activation. The formula for Softmax is:

[0082]

[0083] where z i represents the output of the i-th neuron in the fully connected layer, z c represents the output of the c-th neuron in the fully connected layer, and N is the number of neurons. The behavior recognition result takes the behavior category and confidence corresponding to the maximum value.

[0084] S5: The trajectory segments corresponding to the known behavior category labels of various vehicle behaviors are formed into sample data, and the DGRU-SimAM network model is trained according to this sample data; according to the recognition result output by the network and the actual behavior label, the mean square error MSE is used to calculate the loss value, and the Adam optimizer is used to adjust the parameters of the DGRU-SimAM network model according to the loss value, and then return to step S5 until the preset accuracy is reached;

[0085] where the formula for the mean square error MSE loss function is:

[0086]

[0087] where n represents the number of samples, y irepresents the true value, represents the predicted value.

[0088] S6: Input the trajectory segments obtained by splitting the actually acquired vehicle motion trajectory into the trained DGRU-SimAM network model to obtain the corresponding vehicle behavior categories;

[0089] S7: According to the order corresponding to each trajectory segment after splitting, combine the obtained vehicle behavior category labels into the vehicle behavior sequence corresponding to the trajectory to obtain the semantic description of the behavior, that is, realize the recognition of vehicle behavior.

[0090] When obtaining the semantic description of the behavior, since there are overlapping parts between the previously split trajectory segments, it is necessary to judge the behavior category labels of the overlapping parts, as Figure 3 shown, let it be the behavior category with a higher confidence:

[0091] <cls i,j ,conf i,j > = max(<cls i ,conf i , <cls j ,conf j )

[0092] where cls and conf represent the behavior category and the confidence corresponding to the category respectively. After determining the behavior category labels of the overlapping parts, the final behavior sequence result will be formed to realize the semantic description of the behavior.

[0093] The present invention provides a vehicle behavior recognition system based on a double-layer GRU and SimAM, including the following modules:

[0094] A trajectory splitting module, which uses on-vehicle sensors or visual target tracking technology to obtain the vehicle's motion trajectory, and adopts a sliding window with a certain window length and sliding step to split the trajectory and intercept and generate multiple trajectory segments;

[0095] A sequence description module, which is used to describe the sequence composed of vectors, where is used to represent the position offset of the vehicle in the x direction within the current time interval, is used to represent the position offset of the vehicle in the y direction within the current time interval, is used to represent the change in the vehicle's motion direction within the current time interval;

[0096] Calculate using the formula shown below

[0097]

[0098]

[0099]

[0100] Among them, x i-1 and y i-1 respectively represent the x-position coordinate and y-position coordinate of the vehicle at time i-1, x i and y i respectively represent the x-position coordinate and y-position coordinate of the vehicle at time i, x i-2 and y i-2 respectively represent the x-position coordinate and y-position coordinate of the vehicle at time i-2, i represents a positive integer greater than or equal to 2, x max and x min represent the maximum and minimum values of the x-position coordinate, y max and y min represent the maximum and minimum values of the y-position coordinate, arctan represents the arctangent function.

[0101] The network training module is used to construct a DGRU-SimAM network model and train the DGRU-SimAM network model until convergence; the DGRU-SimAM network model includes stacked double-layer GRU units, a SimAM module, a fully connected layer, and a Softmax layer;

[0102] When calculating the attention weight using the SimAM module, the minimum energy of neuron t at time step j is calculated using the following formula

[0103]

[0104] Among them, λ is the regularization term parameter, and the empirical value is set to 0.00001, and are:

[0105]

[0106]

[0107] Among them, M represents the number of neurons in the hidden layer, T represents the time span of the sequence, t (j) and respectively represent the states of the target neuron and other neurons at time step j;

[0108] After calculating the minimum energy of M neurons at T time steps, the corresponding energy matrix E is formed, and the output feature X is weighted with this to obtain the weighted result of the output feature

[0109]

[0110] Among them, sigmoid is the activation function, and ⊙ represents the dot product operation of matrices.

[0111] The network inference module is used to obtain the behavior recognition result corresponding to the trajectory segment after inputting the DGRU-SimAM network model, that is, to obtain its corresponding vehicle behavior category.

[0112] The behavior sequence generation module is used to combine the behavior recognition results of each trajectory segment according to the order corresponding to each split trajectory segment, combine the obtained vehicle behavior category labels into the vehicle behavior sequence corresponding to the trajectory, and obtain the semantic description of the behavior, that is, to realize the recognition of vehicle behavior.

[0113] When obtaining the semantic description of the behavior, since there are overlapping parts between the previously split trajectory segments, it is necessary to label the behavior category labels of the overlapping parts separately, and make it the behavior category with a higher confidence:

[0114] <cls i,j ,conf i,j > = max(<cls i ,conf i , <cls j ,conf j )

[0115] Among them, cls and conf respectively represent the behavior category and the confidence corresponding to the category. After determining the behavior category labels of the overlapping parts, the final behavior sequence result will be formed to realize the semantic description of the behavior.

[0116] The beneficial effects of the present invention are:

[0117] 1. In the DGRU-SimAM model, the double-layer GRU network stacked by GRU units can improve the feature learning ability of the network for sequence data through stacking processing. Adding the SimAM module to the output of the hidden layer of the network can make the output more focused on useful local information without increasing the number of network parameters.

[0118] 2. When using a sliding window to split the trajectory sequence, there will be redundant overlapping parts between each trajectory segment, so as to realize the secondary judgment of the split position area and avoid the problem of discontinuous behavior states caused by improper split positions.

[0119] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A vehicle behavior recognition method based on double-layer GRU and SimAM, characterized in that: Including: S1: Record the vehicle's motion trajectory and split the vehicle's motion trajectory into multiple trajectory segments through a sliding window; S2: Describe each trajectory segment using a sequence of vectors, where represents the position offset of the vehicle in the x - direction during the time interval between the i - th moment and the (i - 1)-th moment, represents the position offset of the vehicle in the y - direction during the time interval between the i - th moment and the (i - 1)-th moment, represents the change in the movement direction of the vehicle during the time interval between the i - th moment and the (i - 1)-th moment; S3: Construct a DGRU-SimAM network model, which includes stacked double-layer GRU units, a SimAM module, a fully connected layer, and a Softmax layer; S4: Sequentially input the description vector sequences corresponding to each trajectory segment into the double-layer GRU unit to obtain output features. Concatenate the output features X at time steps, calculate the attention weights through the SimAM module, and input the output features after weighting by the SimAM module into the fully connected layer and the Softmax layer to obtain the output vector <cls, conf>, which represents the possibility of the trajectory segment corresponding to various vehicle behaviors. Among them, cls represents the behavior category and conf represents the confidence; S5: Use the trajectory segments corresponding to the known behavior category labels of various vehicle behaviors to form sample data, and train the DGRU-SimAM network model according to this sample data to obtain a converged network model; S6: Input the trajectory segments obtained by splitting the actually obtained vehicle motion trajectory into the trained DGRU-SimAM network model to obtain the corresponding vehicle behavior category; S7: Combine the obtained vehicle behavior category labels into the vehicle behavior sequence corresponding to the trajectory in the order corresponding to each trajectory segment after splitting to obtain the semantic description of the behavior, that is, realize the recognition of vehicle behavior.

2. The vehicle behavior recognition method based on double-layer GRU and SimAM according to claim 1, wherein: In step S1, the vehicle's motion trajectory is obtained by using an in-vehicle sensor or visual target tracking technology, and the trajectory is split by a sliding window with a certain window length and sliding step to intercept and generate multiple trajectory segments.

3. The vehicle behavior recognition method based on double-layer GRU and SimAM according to claim 1, wherein: In step S2, calculate using the formula shown below Among them, x i-1 , y i-1 respectively represent the x-position coordinate and y-position coordinate of the vehicle at the (i - 1)th moment, x i , y i respectively represent the x-position coordinate and y-position coordinate of the vehicle at the ith moment, x i-2 , y i-2 respectively represent the x-position coordinate and y-position coordinate of the vehicle at the (i - 2)th moment, i represents a positive integer greater than or equal to 2, x max and x min represent the maximum and minimum values of the x-position coordinate, y max and y min represent the maximum and minimum values of the y-position coordinate, and arctan represents the arctangent function.

4. The vehicle behavior recognition method based on double-layer GRU and SimAM according to claim 1, characterized in that: In step S4, when calculating the attention weights using the SimAM module, the minimum energy of neuron t at time step j is calculated using the following formula where λ is the regularization term parameter, and its empirical value is set to 0.00001, and is: where M represents the number of hidden layer neurons, T represents the time span of the sequence, and t (j) and represent the states of the target neuron and other neurons at time step j, respectively; After calculating the minimum energy of M neurons over T time steps, the corresponding energy matrix E is formed, and the output feature X is weighted with this to obtain the weighted result of the output feature. Among them, sigmoid is the activation function, and ⊙ represents the dot product operation of matrices.

5. The vehicle behavior recognition method based on double-layer GRU and SimAM according to claim 1, characterized in that: In step S7, when obtaining the semantic description of the behavior, since there are overlapping parts between the previously split trajectory segments, it is necessary to judge the behavior category labels of the overlapping parts and make them the behavior category with a higher confidence: <cls i,j ,conf i,j > = max(<cls i ,conf i >, <cls j ,conf j ) Among them, cls and conf represent the behavior category and the confidence corresponding to the category respectively. After determining the behavior category labels of the overlapping parts, the final behavior sequence result will be formed to realize the semantic description of the behavior.

6. A vehicle behavior recognition system based on double-layer GRU and SimAM, which is used to implement a vehicle behavior recognition method based on double-layer GRU and SimAM according to any one of claims 1-5, and is characterized in that: The system includes: A trajectory splitting module, which is used to obtain the vehicle motion trajectory and split it into multiple trajectory segments; A sequence description module for describing the sequences formed by vectors, where is used to represent the position offset of the vehicle in the x direction within the current time interval, is used to represent the position offset of the vehicle in the y direction within the current time interval, is used to represent the change in the movement direction of the vehicle within the current time interval; A network training module, which is used to construct a DGRU-SimAM network model and train the DGRU-SimAM network model until it converges; the DGRU-SimAM network model includes stacked double-layer GRU units, a SimAM module, a fully connected layer, and a Softmax layer; A network inference module, which is used to obtain the behavior recognition result corresponding to the trajectory segment after inputting into the DGRU-SimAM network model, that is, obtain the corresponding vehicle behavior category; The behavior sequence generation module is used to combine the behavior recognition results of each trajectory segment according to the order corresponding to each split trajectory segment, and combine the obtained vehicle behavior category labels into the vehicle behavior sequence corresponding to the trajectory, so as to obtain the semantic description of the behavior, that is, to realize the recognition of vehicle behavior.

7. The vehicle behavior recognition system based on double-layer GRU and SimAM according to claim 6, characterized in that: In the trajectory splitting module, the vehicle's motion trajectory is obtained by using in-vehicle sensors or visual target tracking technology, and a sliding window with a certain window length and sliding step is used to split the trajectory, and multiple trajectory segments are intercepted and generated.

8. The vehicle behavior recognition system based on double-layer GRU and SimAM according to claim 6, wherein: In the network training module, the following formula is used for calculation where x i-1 and y i-1 represent the x-position coordinate and y-position coordinate of the vehicle at the (i - 1)th moment respectively, x i and y i represent the x-position coordinate and y-position coordinate of the vehicle at the ith moment respectively, x i-2 and y i-2 represent the x-position coordinate and y-position coordinate of the vehicle at the (i - 2)th moment respectively, i represents a positive integer greater than or equal to 2, x max and x min represent the maximum value and minimum value of the x-position coordinate, y max and y min represent the maximum value and minimum value of the y-position coordinate, arctan represents the arctangent function.

9. The vehicle behavior recognition system based on double-layer GRU and SimAM according to claim 6, characterized in that: In the network training module, when calculating the attention weights using the SimAM module, the minimum energy of neuron t at time step j is calculated using the following formula where λ is the regularization term parameter, and its empirical value is set to 0.00001, and is as follows: where M represents the number of neurons in the hidden layer, T represents the time span of the sequence, and t (j) and represent the states of the target neuron and other neurons at time step j, respectively; After calculating the minimum energy of M neurons over T time steps, the corresponding energy matrix E is constructed to weight the output feature X, resulting in the weighted output feature. Among them, sigmoid is the activation function, and ⊙ represents the dot product operation of matrices.

10. A vehicle behavior recognition system based on double-layer GRU and SimAM according to claim 6, characterized in that: In the behavior sequence generation module, when obtaining the semantic description of the behavior, since there are overlapping parts between the previously split trajectory segments, it is necessary to label the behavior category labels of the overlapping parts separately, and make them the behavior categories with higher confidence: <cls i,j ,conf i,j > = max(<cls i ,conf i >, <cls j ,conf j >) Among them, cls and conf represent the behavior category and the confidence corresponding to the category respectively. After determining the behavior category labels of the overlapping parts, the final behavior sequence result will be formed to realize the semantic description of the behavior.

Images