Feature fusion improved attention mechanism and LSTM (Long Short Term Memory) combined vehicle track prediction method

Abstract

The invention discloses a vehicle trajectory prediction method, and particularly relates to a feature fusion improved attention mechanism and LSTM combined vehicle trajectory prediction method. The system is mainly composed of two modules: a safe distance grading attention mechanism module and an interactive attention mechanism module. The method mainly comprises the steps that historical driving data of a target vehicle and surrounding interactive vehicles are acquired, the ratio of the actual distance to the theoretical safe distance of the target research vehicle and the surrounding interactive vehicles is calculated and divided into three grades, and a safe distance grading attention mechanism module processes information of each group through an encoder to acquire three groups of context vectors; the interactive attention mechanism module divides historical data into a target vehicle group and a surrounding interactive vehicle group, the input of the target vehicle group is converted into a query matrix Q through a linear layer, the surrounding interactive vehicle group is converted into a key matrix K and a value matrix V through two linear layers, a weighting result is obtained through dot product and Softmax operation, and finally the outputs of the two modules are fused to obtain a weighted result. And the data is input into the LSTM decoder.

Description

Technical Field

[0001] The present invention patent discloses a vehicle trajectory prediction method, and in particular relates to a feature fusion improved attention mechanism combined with LSTM vehicle trajectory prediction method. Background Art

[0002] Currently, intelligent driving vehicle trajectory prediction has attracted the attention of many researchers. It is crucial to pay attention to surrounding driving vehicles in dynamic driving scenarios and to incorporate the driving behaviors and states of surrounding vehicles into vehicle prediction modeling.

[0003] There are existing methods based on physical models, classical machine learning, and deep learning for the problem of trajectory prediction of intelligent driving vehicles. Trajectory prediction methods based on physical models simulate vehicle motion by directly applying the principles of dynamics and kinematics. They are suitable for quick and simple predictions of vehicle behavior, but may not be sufficient to capture complex interactions and the diversity of driving behaviors. Trajectory prediction methods based on classical machine learning use statistical models to learn trajectory patterns from historical data and can handle a variety of traffic scenarios, but are usually limited by data quality and model generalization capabilities, and require a large amount of labeled data for effective training. Trajectory prediction methods based on deep learning automatically extract spatiotemporal features from data through complex neural network structures. They are particularly good at processing large-scale and high-dimensional traffic data and generating accurate and multimodal prediction results.

[0004] The encoder-decoder structure based on various attention mechanisms is used to solve the problem of insufficient prediction accuracy of traditional physical models, and the insufficient utilization of global interactive trajectory information is also a problem worth solving. Therefore, it is necessary to improve the existing methods by weighted fusion of local features and global features to understand the dynamic behavior of vehicles from multiple dimensions. An accurate and reliable vehicle trajectory prediction method is urgently needed. Summary of the invention

[0005] To solve the above problems, the present invention provides a feature fusion improved attention mechanism combined with LSTM vehicle trajectory prediction method, which aims to improve driving efficiency and driving safety while meeting the requirements of double precision in vehicle trajectory prediction time and space.

[0006] In order to achieve the above object, the present invention adopts the following technical scheme:

[0007] An embodiment of the present invention provides a feature fusion improved attention mechanism combined with LSTM vehicle trajectory prediction method, comprising the following steps:

[0008] Step 1: Obtain historical driving data of the target vehicle and surrounding interactive vehicles;

[0009] Step 2: Calculate the ratio of the actual distance between the target vehicle and the surrounding interactive vehicles to the theoretical safe distance and divide it into three levels;

[0010] Step 3: Input the three-level safety distance into the safety distance attention mechanism module encoder to obtain three groups of context vectors; divide the target vehicle and the interactive vehicle into two groups and input them into the interactive attention mechanism module. The target vehicle group input is converted into a query matrix Q through a linear layer, and the surrounding interactive vehicle group is converted into a key matrix K and a value matrix V through two linear layers. After the linear operation, Q and K are dot-producted and passed to the Softmax layer for weighting;

[0011] Step 4: The context information Ct obtained by the safety distance hierarchical attention mechanism and the interactive attention mechanism is fused to obtain enhanced context information, which is then input into the LSTM decoder and finally the prediction result is output.

[0012] On the basis of the above technical solution, preferably, the historical driving data includes position information, speed, acceleration, timestamp, and lane information of the target vehicle and the interacting vehicle.

[0013] Based on the above technical solution, preferably, the calculation and classification of the safe distance between the target vehicle and the surrounding interactive vehicles divides the safe distance into three levels: close, safe and far. This classification is defined based on the ratio of the actual distance between the vehicles to the theoretical safe distance calculated by the following formula:

[0014]

[0015] Where v is the vehicle speed, T r is the reaction time, a is the target vehicle deceleration, w is the lane width, k is the safety factor, d l is the lateral distance between the target vehicle and the left interacting vehicle, d r is the lateral distance between the target vehicle and the vehicle on the right, d f is the distance between the target vehicle and the preceding vehicle, d b is the distance between the target vehicle and the following vehicle.

[0016]

[0017] Among them, min(d f ,d b ,d l ,d r ) is the actual distance between the vehicle and the interacting vehicle.

[0018] Compared with the prior art, the present invention has the following beneficial technical effects:

[0019] 1. The safety distance graded attention module encodes the safety distance grade information of the target vehicle and the surrounding interactive vehicles separately, capturing the feature differences at different safety distances, so as to more accurately understand the status of the vehicle in various environments;

[0020] 2. The feature fusion of the safety distance hierarchical attention module and the interactive attention module can effectively capture the complex relationship between the global and local environment;

[0021] 3. Use the LSTM decoder to predict the fused context information, which can update and accurately predict the vehicle trajectory in real time, improving the system's real-time response capability and prediction accuracy; BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solution of the embodiment of the present invention, the specific implementation of the present invention is further described in detail below with reference to the accompanying drawings, wherein:

[0023] Figure 1 It is a flow chart of the vehicle trajectory prediction method combining feature fusion improved attention mechanism with LSTM in the present invention;

[0024] Figure 2 It is a schematic diagram of the safety distance graded attention mechanism module of the present invention;

[0025] Figure 3 It is a schematic diagram of the interactive attention mechanism module of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention more clear, the present invention is further described in detail. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] See also Figure 1 The embodiment of the present invention discloses a method for predicting vehicle trajectories by combining a feature fusion improved attention mechanism with an LSTM, comprising the following steps:

[0028] Step 1: Obtain historical driving data of the target vehicle and surrounding interactive vehicles;

[0029] Step 2: Calculate the safe distance between the target vehicle and the surrounding interactive vehicles and divide them into three levels. The safe distance formula is:

[0030]

[0031] Where v is the vehicle speed, T r is the reaction time, a is the target vehicle deceleration, w is the lane width, k is the safety factor, dl is the lateral distance between the target vehicle and the left interacting vehicle, d r is the lateral distance between the target vehicle and the vehicle on the right, d f is the distance between the target vehicle and the preceding vehicle, d b is the distance between the target vehicle and the following vehicle.

[0032] The formula for grading standards is:

[0033]

[0034] Among them, min(d f ,d b ,d l ,d r ) is the actual distance between the vehicle and the interacting vehicle.

[0035] Step 3: Input the three-level safety distance into the safety distance attention mechanism module encoder to obtain three sets of context vectors. The safety distance attention mechanism module is as follows: Figure 2 As shown in Figure 2, the information of the target vehicle and the surrounding interactive vehicles are respectively input into the interactive attention mechanism module. The interactive attention mechanism module is as follows: Figure 3 As shown in the figure, the information of the target vehicle is converted into a query matrix Q through a linear layer, and the information of the surrounding interactive vehicles is converted into a key matrix K and a value matrix V through two linear layers. The query matrix Q is combined with the key matrix K through a dot product operation, and weighted through a Softmax layer to calculate the weight of the interactive relationship between the target vehicle and the surrounding vehicles;

[0036] Step 4: Combine the context information C obtained by the safety distance hierarchical attention mechanism and the interactive attention mechanism t Fusion is performed to obtain enhanced context information, and the fused context information C t As the input of the current time step of the LSTM decoder, the hidden state h0 and cell state c0 of the LSTM are initialized. For each time step t, the following operations are performed:

[0037] Forget gate update: The forget gate determines which information should be forgotten from the cell state, using the sigmoid activation function and the forget gate weight W f and the bias term b f Calculate the forget gate output f t .

[0038] Input gate update: The input gate controls the writing of new information and calculates the input gate output i t and candidate value vector C~ t , using sigmoid activation function and tanh activation function respectively.

[0039] Cell state update: Cell state C t is the old cell state C t-1 and the new candidate value vector C~ t A combination of , where the old cell state is forgotten according to the output of the forget gate, using a sigmoid activation function and an output gate weight W o and the bias term b o Calculate the output gate output o t .

[0040] Hidden state update: Output hidden state h at the current time step t .

[0041] Output prediction: hidden state h of the LSTM cell t is used as input to predict the next position,

[0042] Sequence iteration and trajectory generation: The above steps are repeated until all time step sequences are processed and the output is converted into trajectory points to form the future position of the vehicle.

Claims

1. A feature fusion improved attention mechanism combined with LSTM vehicle trajectory prediction method, characterized in that: include: Step 1: Obtain historical driving data of the target vehicle and surrounding interactive vehicles; Step 2: Calculate the ratio of the actual distance between the target vehicle and the surrounding interactive vehicles to the theoretical safe distance and divide it into three levels; Step 3: Input the three-level safety distance into the safety distance attention mechanism module encoder to obtain three groups of context vectors; divide the target vehicle and the interactive vehicle into two groups and input them into the interactive attention mechanism module. The target vehicle group input is converted into a query matrix Q through a linear layer, and the surrounding interactive vehicle group is converted into a key matrix K and a value matrix V through two linear layers. After the linear operation, Q and K are dot-producted and passed to the Softmax layer for weighting; Step 4: The context information Ct obtained by the safety distance hierarchical attention mechanism and the interactive attention mechanism is fused to obtain enhanced context information, which is then input into the LSTM decoder and finally the prediction result is output.

2. The feature fusion improved attention mechanism combined with LSTM vehicle trajectory prediction method as described in claim 1 is characterized by: The historical driving data in step 1 includes the location information, speed, acceleration, timestamp, and lane information of the target vehicle and the interacting vehicle.

3. The feature fusion improved attention mechanism combined with LSTM vehicle trajectory prediction method as described in claim 1 is characterized by: In step 2, the theoretical safe distance between the target vehicle and the surrounding interactive vehicles is calculated comprehensively by vehicle speed, reaction time, acceleration of the target vehicle, lane width, safety factor, lateral distance between the target vehicle and the left interactive vehicle, lateral distance between the target vehicle and the right vehicle, distance between the target vehicle and the front vehicle, and distance between the target vehicle and the rear vehicle.

4. The feature fusion improved attention mechanism combined with LSTM vehicle trajectory prediction method as described in claim 1 is characterized in that: The three levels in step 2 are: close, safe, and far.

Images