A pedestrian trajectory prediction method based on SFM-LSTM neural network model

Abstract

The present invention relates to a method for predicting pedestrian trajectories based on an SFM-LSTM neural network model. The method comprises the following steps: step 1: obtaining motion state information, individual feature information, and human-vehicle interaction scene information of pedestrians crossing the street; step 2: performing data preprocessing and data enhancement to establish a pedestrian trajectory dataset; step 3: establishing and training an LSTM neural network model; step 4: obtaining predicted trajectories of pedestrians crossing the street through the trained LSTM neural network model; step 5: calibrating parameters of a social force model using a maximum likelihood estimation method; step 6: correcting the predicted trajectory according to the social force model and outputting the optimal predicted trajectory of the pedestrian crossing the street; and step 7: broadcasting the optimal predicted trajectory to nearby vehicles to assist intelligent connected vehicles in making decisions. Compared with the existing technology, the present invention has the advantages of improving the safety of pedestrians crossing the street, reducing the delay rate of vehicles, and improving the traffic capacity of roads.

Description

Technical Field

[0001] The present invention relates to the field of vehicle-road collaboration for intelligent connected vehicles, and in particular to a method for predicting the trajectory of pedestrians crossing the street based on an SFM-LSTM neural network model. Background Art

[0002] In recent years, with the rapid development of intelligent connected vehicle technology, more and more car models have different degrees of autonomous driving capabilities. Functions such as Lane Keeping Assistance System (LKAS), Forward Collision Warning System (FCWS), and Adaptive Cruise Control (ACC) in the Advanced Driver Assistant System (ADAS) have been installed on some mid-to-high-end brand models and have become their standard features.

[0003] Pedestrian safety is a crucial consideration for intelligent connected vehicles. However, relying solely on vehicle intelligence to ensure pedestrian safety consumes significant onboard computing resources and struggles to address single-vehicle perception blind spots, such as "ghosting" scenarios. Pedestrian trajectory prediction using roadside sensing and computing equipment can effectively address this issue. As key participants in traffic, pedestrians' movements are more complex and varied than those of vehicles, exhibiting significant flexibility and randomness. Understanding pedestrian movement patterns and predicting their trajectory poses significant challenges for intelligent connected vehicles.

[0004] Currently, autonomous driving research related to pedestrian protection primarily focuses on the environmental perception layer, including improving the accuracy of pedestrian recognition using image algorithms and pedestrian detection through multi-sensor fusion. In practical applications, pedestrian handling often involves using sensors to identify pedestrians and immediately stopping to avoid them when they appear within a certain safe distance in front of the vehicle. Alternatively, simple constant velocity (CV) and constant acceleration (CA) models are used to describe pedestrians' movement, defining them as moving obstacles for avoidance. However, pedestrians are highly mobile, and these approaches are either overly conservative, resulting in an unnatural driving process, impacting traffic efficiency and the passenger experience. Alternatively, they fail to adequately analyze pedestrians' movement, making accurate predictions impossible, resulting in inadequate pedestrian protection. In serious cases, the consequences can be disastrous.

[0005] Among pedestrian trajectory prediction methods, the Social Force Model (SFM) is a method for pedestrian microdynamics. In specific scenarios, given initial conditions, the Social Force Model can generate trajectories that conform to pedestrian walking patterns. Using these generated trajectories as prediction values is also an approach for pedestrian trajectory prediction. While the deep learning Long Short-Term Memory (LSTM) model is used to predict pedestrian trajectories, the prediction effect is good, but it only considers the trajectory prediction of a single pedestrian in isolation, without considering the impact of vehicles and surrounding pedestrians on the target pedestrian. Summary of the Invention

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a pedestrian trajectory prediction method based on the SFM-LSTM neural network model.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] A method for predicting pedestrian trajectories based on an SFM-LSTM neural network model, the method comprising the following steps:

[0009] Step 1: Collect data through multiple sensors to obtain pedestrian movement status information, individual characteristics information, and human-vehicle interaction scene information;

[0010] Step 2: Perform data preprocessing and data enhancement to establish a pedestrian trajectory dataset;

[0011] Step 3: Establish an LSTM neural network model and train it based on the pedestrian trajectory dataset;

[0012] Step 4: Obtain the predicted trajectory of pedestrians crossing the street within the first preset time period in the future through the trained LSTM neural network model;

[0013] Step 5: Use the maximum likelihood estimation method to calibrate the parameters of the social force model;

[0014] Step 6: Correct the predicted trajectory of the LSTM neural network model using the social force model, and output the optimal predicted trajectory of pedestrians crossing the street within the first preset time period;

[0015] Step 7: Broadcast the optimal predicted trajectory to nearby vehicles to assist intelligent connected vehicles in decision-making.

[0016] In step 1, the process of collecting data through multiple sensors is specifically as follows:

[0017] The multi-sensor system includes roadside lidar, millimeter-wave radar and cameras. A preliminary survey is conducted on the zebra crossing area where pedestrians and vehicles are in free flow conditions. The roadside lidar and millimeter-wave radar are used to collect the trajectory coordinates of pedestrians and vehicles crossing the street. At the same time, cameras are set up at the intersection to synchronously shoot pedestrian walking videos at a head-on angle to calibrate the age and gender of pedestrians. Based on the multi-sensor information fusion algorithm, the motion status information of pedestrians crossing the street within the vehicle safety envelope, the individual characteristics of pedestrians and the human-vehicle interaction scene information are obtained.

[0018] The movement status information of the pedestrian crossing the street includes the pedestrian's current position and speed;

[0019] The human-vehicle interaction scene information includes the vehicle's current location, speed, and vehicle model;

[0020] The individual characteristic information of pedestrians crossing the street includes the age, gender, and the ideal speed v of pedestrians of different ages and genders when crossing the street. d , maximum speed v max and reaction time τ;

[0021] The ideal speed v d The average walking speed of pedestrians of different ages and genders when crossing the street without any external interference;

[0022] The maximum speed v max The maximum speed that pedestrians of different ages and genders can reach when crossing a zebra crossing when the traffic light is flashing and about to change from green to red;

[0023] The reaction time τ is the time it takes for pedestrians of different ages and genders to change from their current speed to the ideal speed v d The time required.

[0024] In step 3, the data preprocessing process is specifically as follows:

[0025] Data preprocessing is performed on the acquired pedestrian motion status information, individual characteristics, and human-vehicle interaction information. The video is segmented into images at 0.2-second intervals to achieve the purpose of discretizing the trajectory. The coordinates of the pedestrian's pixel location are obtained from the image. The actual trajectory of the pedestrian crossing the street under initial conditions is obtained through the direct linear transformation method to calibrate the parameters of the social force model.

[0026] The specific process of data enhancement is as follows:

[0027] The trajectory rotation method is used for data enhancement. A random point in the line connecting the starting point and the end point of the original real trajectory is used as the rotation center, and the rotation operation is performed at fixed angles to expand the data.

[0028] In step 2, the pedestrian trajectory data set includes the actual trajectory of pedestrians crossing the street and pedestrian attribute data, and the pedestrian attribute data includes age and gender.

[0029] In step 3, the LSTM neural network model uses the pedestrian observation sequence as input features, adopts an encoder-decoder framework, and introduces an attention mechanism into the LSTM neural network model to mine hidden factors in the observation sequence that affect the future trajectory. The structure of the LSTM neural network model includes:

[0030] Hidden layer: The dimension is set to 256 and consists of LSTM units. The LSTM unit includes three control gates: input gate, forget gate, and output gate, which are used to control the relationship between input, output, and internal state.

[0031] Input layer: includes multiple input units, each of which uses the tanh activation function, and each input unit corresponds to the input feature;

[0032] Output layer: includes multiple output units, each input unit uses a tanh activation function, and the corresponding outputs are the X-direction movement position and the Y-direction movement position of the pedestrian crossing the street within the first preset time length.

[0033] The pedestrian observation sequences are pedestrian speed in the X direction, pedestrian speed in the Y direction, pedestrian position in the X direction, pedestrian position in the Y direction, age, gender, vehicle speed in the X direction, vehicle speed in the Y direction, vehicle position in the X direction, vehicle position in the Y direction, and vehicle type.

[0034] In step 3, the process of training the LSTM neural network model specifically includes the following steps:

[0035] Step 301: Based on the relationship between input, output, and internal state, the expression of the LSTM model is obtained as follows:

[0036] i t =σ(W i ·[h t-1 , X t ]+b i )

[0037] f t =σ(W f ·[h t-1 , X t ]+b f )

[0038] O t =σ(W o ·[h t-1 , Xt ]+b o )

[0039] C t =f t *C t-1 +i t *tanh(W c ·[h i-1 , X t ]+b c )

[0040] h t =O t *tanh(C t )

[0041] Among them, i t represents the input gate, f t represents the forget gate, O t represents the output gate, C t represents the unit state at the current time t, h t represents the hidden state at the current time t, C t-1 Indicates the unit state at the previous moment, h t-1 Indicates the hidden state of the previous moment, X t Represents the input vector at the current time t, W i Represents the weight matrix of the input gate, W f Represents the weight matrix of the forget gate, W o Represents the weight matrix of the output gate, W c The weight matrix representing the unit state, b i represents the bias term of the input gate, b f represents the bias term of the forget gate, b o represents the bias term of the output gate, b c Represents the bias term of the unit state, σ represents the sigmoid function, and tanh represents the tanh function;

[0042] Step 302: Obtain the movement status information, individual characteristic information, and human-vehicle interaction information of pedestrians crossing the street through roadside perception sensors, perform data preprocessing and data enhancement on them, and then import them into the LSTM neural network model to train the structural weights and bias parameters of the LSTM neural network model.

[0043] In step 4, the process of obtaining the predicted trajectory of pedestrians crossing the street within the first preset time period in the future based on the LSTM neural network model specifically includes the following steps:

[0044] Step 401: Acquire 11 current input features of vehicles and pedestrians through multiple sensors;

[0045] Step 402: performing data normalization on the acquired input features and importing them into the trained LSTM neural network model;

[0046] Step 403: Output the predicted data and denormalize it to obtain the movement position of the pedestrian crossing the street within the first preset time period in the future, that is, the predicted trajectory of the pedestrian crossing the street within the first preset time period in the future.

[0047] In step 5, the process of calibrating the parameters of the social force model using the maximum likelihood estimation method specifically includes the following steps:

[0048] Step 501: obtaining the actual trajectory of pedestrians crossing the street under initial conditions in the pedestrian trajectory dataset;

[0049] Step 502: Set the same initial conditions, give the social force model parameters an initial value, and use the social force model to correct the predicted trajectory of the LSTM neural network model;

[0050] Step 503: Construct a maximum likelihood function based on the difference between the true trajectory and the predicted trajectory. The maximum likelihood function of the two-dimensional normal distribution is:

[0051]

[0052] Among them, θ is the parameter vector, A b Represents the zebra crossing boundary force-strength coefficient, B b Represents the zebra crossing boundary force-distance coefficient, A p represents the force-intensity coefficient between pedestrians, B p represents the force-distance coefficient between pedestrians, A v Indicates the influence of vehicles on pedestrians - intensity coefficient, B v Indicates the influence of vehicles on pedestrians - distance coefficient, represents the x-direction trajectory correction weight, Indicates the y-direction trajectory correction weight, p i (t) is the predicted trajectory of the i-th pedestrian after correction by the social force model, Δp i (t) = p i (t+1)-p i (t) is the predicted displacement of the pedestrian crossing the street from time t to time t+1, is the mean vector, ∑ p is the covariance matrix, p i (t), Δp i (t) and All about function;

[0053] Step 504: Assume that there are N pedestrians crossing the street in all the real trajectories used for parameter calibration, and the time length of each pedestrian crossing trajectory is M. Then the expression of the total maximum likelihood function is:

[0054]

[0055] Among them, M is the preset time length, and N is the number of pedestrians crossing the street;

[0056] Step 505: Continuously iterate through the maximum likelihood estimation method to finally complete the calibration of the social force model parameters.

[0057] Compared with the prior art, the present invention has the following advantages:

[0058] 1. This invention fully considers the individual differences of pedestrians and proposes a new prediction method that integrates social forces and long-short-term memory networks. It applies the optimal prediction trajectory to the intelligent decision-making field of self-driving cars, which can improve the safety of pedestrians crossing the street, reduce vehicle delays, and improve road capacity.

[0059] 2. The present invention can rely on roadside sensing equipment to collect information such as pedestrian locations and attributes, accurately predict the trajectories of pedestrians crossing the street, effectively reduce the error compared to the actual trajectories of pedestrians, and broadcast the predicted trajectories to nearby intelligent connected vehicles to assist them in decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0061] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0062] Example

[0063] The present invention provides a method for predicting pedestrian trajectories based on the SFM-LSTM neural network model. The method can accurately predict the trajectories of pedestrians crossing the street by collecting information such as the pedestrian's location and attributes through roadside sensing equipment, effectively reducing the error between the predicted trajectory and the pedestrian's actual trajectory, and broadcasting the predicted trajectory to nearby intelligent connected vehicles to assist them in decision-making.

[0064] The method comprises the following steps:

[0065] Step 1: Obtain information on the movement status, individual characteristics, and interaction scenarios of pedestrians crossing the street. Preliminary investigations are conducted on zebra crossings where pedestrians and vehicles are in free flow. Multiple sensors, including roadside lidar, cameras, and millimeter-wave radar, are used to obtain a variety of information. A multi-sensor information fusion algorithm is used to obtain information on the movement status, individual characteristics, and interaction scenarios of pedestrians crossing the street within the vehicle safety envelope.

[0066] Individual characteristics of pedestrians crossing the street include their age and gender, and the ideal speed v for pedestrians of different ages and genders when crossing the street. d , maximum speed v max And reaction time τ, ideal speed v d is the average speed of pedestrians of different ages and genders walking when crossing the street without any external interference, and the maximum speed v max is the maximum speed that pedestrians of different ages and genders can reach when crossing the zebra crossing when the traffic light flashes from green to red. The reaction time τ is the time it takes for pedestrians of different ages and genders to change from the current speed to the ideal speed v d the time required;

[0067] The movement status information of pedestrians crossing the street includes the pedestrian’s current position and speed;

[0068] The human-vehicle interaction scene information includes the vehicle's current location, speed, and model.

[0069] Step 2: Establish an LSTM neural network model and train the LSTM neural network model. The process of training the LSTM neural network model specifically includes the following steps:

[0070] The LSTM neural network model has a network structure with one hidden layer, an input layer, and an output layer. The input layer includes multiple input units, and the output layer includes multiple output units. The features of the input layer are pedestrian speed (including X and Y directions), position (including X and Y directions), age, gender, vehicle speed (including X and Y directions), position (including X and Y directions), and vehicle type. The gate control unit in the hidden layer uses the sigmoid activation function, and the input and output units use the tanh activation function. The dimension of the hidden layer is set to 256. The output information of the output unit is the movement position (including X and Y directions) of the pedestrian crossing the street within the first preset time length.

[0071] The LSTM unit consists of three control gates: input gate, forget gate, and output gate, which are used to control the relationship between input, output, and internal state. The relevant calculations are as follows:

[0072] i t =σ(Wi ·[h t-1 , X t ]+b i )

[0073] f t =σ(W f ·[h t-1 , X t ]+b f )

[0074] O t =σ(W o ·[h t-1 , X t ]+b o )

[0075] C t =f t *C t-1 +i t *tanh(W c ·[h t-1 , X t ]+b c )

[0076] h t =O t *tanh(C t )

[0077] Among them, i t represents the input gate, f t represents the forget gate, O t represents the output gate, C t represents the unit state at the current time t, h t represents the hidden state at the current time t, C t-1 Indicates the unit state at the previous moment, h t-1 Indicates the hidden state of the previous moment, X t Represents the input vector at the current time t, W i Represents the weight matrix of the input gate, W f Represents the weight matrix of the forget gate, W o Represents the weight matrix of the output gate, W c The weight matrix representing the unit state, b i represents the bias term of the input gate, b f represents the bias term of the forget gate, b o represents the bias term of the output gate, b crepresents the bias term of the unit state, σ represents the sigmoid function, and tanh represents the tanh function. The motion state information, individual feature information, and human-vehicle interaction information of pedestrians crossing the street obtained by roadside perception sensors are imported into the LSTM neural network model to train the structural weights and bias parameters of the neural network model.

[0078] Pedestrian trajectory prediction steps based on LSTM neural network model:

[0079] Step 401: 11 input features of the current vehicle and pedestrian are acquired through multiple sensors, including pedestrian speed (including X and Y directions), pedestrian position (including X and Y directions), age, and gender, as well as vehicle speed (including X and Y directions), position (including X and Y directions), and vehicle type.

[0080] Step 402: performing data normalization on the acquired input features and importing them into the trained LSTM neural network model;

[0081] Step 403: Output the predicted data and denormalize it to obtain the movement position of the pedestrian crossing the street within the first preset time period in the future, that is, the predicted trajectory of the pedestrian crossing the street within the first preset time period in the future.

[0082] The parameter calibration process for correcting the predicted trajectory of pedestrians crossing the street based on the social force model includes the following steps:

[0083] Step 501: obtaining the actual trajectory of pedestrians crossing the street under initial conditions in the pedestrian trajectory dataset;

[0084] Step 502: Set the same initial conditions, give the social force model parameters an initial value, and use the social force model to correct the predicted trajectory of the LSTM neural network model;

[0085] Step 503: Construct a maximum likelihood function based on the difference between the true trajectory and the predicted trajectory. The maximum likelihood function of the two-dimensional normal distribution is:

[0086]

[0087] in, is the parameter vector, A b Represents the zebra crossing boundary force-strength coefficient, B b Represents the zebra crossing boundary force-distance coefficient, A p represents the force-intensity coefficient between pedestrians, B p represents the force-distance coefficient between pedestrians, A v Indicates the influence of vehicles on pedestrians - intensity coefficient, B v Indicates the influence of vehicles on pedestrians - distance coefficient, represents the x-direction trajectory correction weight, Indicates the y-direction trajectory correction weight, p i (t) is the predicted trajectory of the i-th pedestrian after correction by the social force model, Δp i (t) = p i (t+1)-p i (t) is the predicted displacement of the pedestrian crossing the street from time t to time t+1, is the mean vector, ∑ p is the covariance matrix, p i (t), Δp i (t) and About function;

[0088] Step 504: Assume that there are N pedestrians crossing the street in all the real trajectories used for parameter calibration, and the time length of each pedestrian crossing trajectory is M. Then the expression of the total maximum likelihood function is:

[0089]

[0090] Among them, M is the preset time length, and N is the number of pedestrians crossing the street;

[0091] Step 505: Continuously iterate through the maximum likelihood estimation method to finally complete the calibration of the social force model parameters.

[0092] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A pedestrian crossing trajectory prediction method based on the SFM-LSTM neural network model, characterized in that: The method comprises the following steps: Step 1: Collect data through multiple sensors to obtain pedestrian movement status information, individual characteristics information, and human-vehicle interaction scene information; Step 2: Perform data preprocessing and data enhancement to establish a pedestrian trajectory dataset; Step 3: Establish an LSTM neural network model and train it based on the pedestrian trajectory dataset; Step 4: Obtain the predicted trajectory of pedestrians crossing the street within the first preset time period in the future through the trained LSTM neural network model; Step 5: Use the maximum likelihood estimation method to calibrate the parameters of the social force model; Step 6: Correct the predicted trajectory of the LSTM neural network model using the social force model, and output the optimal predicted trajectory of pedestrians crossing the street within the first preset time period; Step 7: Broadcast the optimal predicted trajectory to nearby vehicles to assist intelligent connected vehicles in decision-making; In step 5, the process of calibrating the parameters of the social force model using the maximum likelihood estimation method specifically includes the following steps: Step 501: obtaining the actual trajectory of pedestrians crossing the street under initial conditions in the pedestrian trajectory dataset; Step 502: Set the same initial conditions and give the social force model parameters an initial value; Step 503: Construct a maximum likelihood function based on the difference between the true trajectory and the predicted trajectory. The maximum likelihood function of the two-dimensional normal distribution is: Among them, θ is the parameter vector, A b Represents the zebra crossing boundary force-strength coefficient, B b Represents the zebra crossing boundary force-distance coefficient, A p represents the force-intensity coefficient between pedestrians, B p represents the force-distance coefficient between pedestrians, A v Indicates the influence of vehicles on pedestrians - intensity coefficient, B v Indicates the influence of vehicles on pedestrians - distance coefficient, represents the x-direction trajectory correction weight, Indicates the y-direction trajectory correction weight, p i (t) is the predicted trajectory of the i-th pedestrian after correction by the social force model, Δp i (t) = p i (t+1)-p i (t) is the predicted displacement of pedestrians crossing the street from time t to time t+1, μ is the mean vector, ∑ p is the covariance matrix, p i (t), Δp i (t) and μ are both functions of θ; Step 504: Assume that there are N pedestrians crossing the street in all the real trajectories used for parameter calibration, and the time length of each pedestrian crossing trajectory is M. Then the expression of the total maximum likelihood function is: Among them, M is the preset time length, and N is the number of pedestrians crossing the street; Step 505: Continuously iterate through the maximum likelihood estimation method to finally complete the calibration of the social force model parameters.

2. A method for predicting pedestrian trajectories based on an SFM-LSTM neural network model according to claim 1, characterized in that: In step 1, the process of collecting data through multiple sensors is specifically as follows: The multi-sensor system includes roadside lidar, millimeter-wave radar and cameras. A preliminary survey is conducted on the zebra crossing area where pedestrians and vehicles are in free flow conditions. The roadside lidar and millimeter-wave radar are used to collect the trajectory coordinates of pedestrians and vehicles crossing the street. At the same time, cameras are set up at the intersection to synchronously shoot pedestrian walking videos at a head-on angle to calibrate the age and gender of pedestrians. Based on the multi-sensor information fusion algorithm, the motion status information of pedestrians crossing the street within the vehicle safety envelope, the individual characteristics of pedestrians and the human-vehicle interaction scene information are obtained.

3. A method for predicting pedestrian trajectories based on an SFM-LSTM neural network model according to claim 2, characterized in that: The movement status information of the pedestrian crossing the street includes the pedestrian's current position and speed; The human-vehicle interaction scene information includes the vehicle's current location, speed, and vehicle model; The individual characteristic information of pedestrians crossing the street includes the age, gender, and the ideal speed v of pedestrians of different ages and genders when crossing the street. d , maximum speed v max and reaction time τ; The ideal speed v d The average walking speed of pedestrians of different ages and genders when crossing the street without any external interference; The maximum speed v max The maximum speed that pedestrians of different ages and genders can reach when crossing a zebra crossing when the traffic light is flashing and about to change from green to red; The reaction time τ is the time it takes for pedestrians of different ages and genders to change from their current speed to the ideal speed v d The time required.

4. A method for predicting pedestrian trajectories based on an SFM-LSTM neural network model according to claim 3, characterized in that: In step 2, the data preprocessing process is specifically as follows: Data preprocessing is performed on the acquired pedestrian motion status information, individual characteristics, and human-vehicle interaction information. The video is segmented into images at 0.2-second intervals to achieve the purpose of discretizing the trajectory. The coordinates of the pedestrian's pixel location are obtained from the image. The actual trajectory of the pedestrian crossing the street under initial conditions is obtained through the direct linear transformation method to calibrate the parameters of the social force model. The specific process of data enhancement is as follows: The trajectory rotation method is used for data enhancement. A random point in the line connecting the starting point and the end point of the original real trajectory is used as the rotation center, and the rotation operation is performed at fixed angles to expand the data.

5. The method for predicting pedestrian trajectories based on the SFM-LSTM neural network model according to claim 4 is characterized in that: In step 2, the pedestrian trajectory data set includes the actual trajectory of pedestrians crossing the street and pedestrian attribute data, and the pedestrian attribute data includes age and gender.

6. The method for predicting pedestrian trajectories based on the SFM-LSTM neural network model according to claim 1, characterized in that: In step 3, the LSTM neural network model uses the pedestrian observation sequence as input features, adopts an encoder-decoder framework, and introduces an attention mechanism into the LSTM neural network model to mine hidden factors in the observation sequence that affect the future trajectory. The structure of the LSTM neural network model includes: Hidden layer: The dimension is set to 256 and consists of LSTM units. The LSTM unit includes three control gates: input gate, forget gate, and output gate, which are used to control the relationship between input, output, and internal state. Input layer: includes multiple input units, each of which uses the tanh activation function, and each input unit corresponds to the input feature; Output layer: includes multiple output units, each of which uses a tanh activation function, and the corresponding outputs are the X-direction movement position and the Y-direction movement position of the pedestrian crossing the street within the first preset time length.

7. The method for predicting pedestrian trajectories based on the SFM-LSTM neural network model according to claim 6, characterized in that: The pedestrian observation sequences are pedestrian speed in the X direction, pedestrian speed in the Y direction, pedestrian position in the X direction, pedestrian position in the Y direction, age, gender, vehicle speed in the X direction, vehicle speed in the Y direction, vehicle position in the X direction, vehicle position in the Y direction, and vehicle type.

8. The method for predicting pedestrian trajectories based on the SFM-LSTM neural network model according to claim 7, characterized in that: In step 3, the process of training the LSTM neural network model specifically includes the following steps: Step 301: Based on the relationship between input, output, and internal state, the expression of the LSTM model is obtained as follows: i t =σ(W i ·[h t-1 ,X t ]+b i ) f t =σ(W f ·[h t-1 ,X t ]+b f ) The t =σ(W o ·[h t-1 ,X t ]+b o ) C t =f t *C t-1 +i t *tanh(W c ·[h t-1 ,X t ]+b c ) h t =O t *tanh(C t ) Among them, i t represents the input gate, f t represents the forget gate, O t represents the output gate, C t represents the unit state at the current time t, h t represents the hidden state at the current time t, C t-1 Indicates the unit state at the previous moment, h t-1 Indicates the hidden state of the previous moment, X t Represents the input vector at the current time t, W i Represents the weight matrix of the input gate, W f Represents the weight matrix of the forget gate, W o Represents the weight matrix of the output gate, W c The weight matrix representing the unit state, b i represents the bias term of the input gate, b f represents the bias term of the forget gate, b o represents the bias term of the output gate, b c Represents the bias term of the unit state, σ represents the sigmoid function, and tanh represents the tanh function; Step 302: Obtain the movement status information, individual characteristic information, and human-vehicle interaction information of pedestrians crossing the street through roadside perception sensors, perform data preprocessing and data enhancement on them, and then import them into the LSTM neural network model to train the structural weights and bias parameters of the LSTM neural network model.

9. The method for predicting pedestrian trajectories based on the SFM-LSTM neural network model according to claim 8, characterized in that: In step 4, the process of obtaining the predicted trajectory of pedestrians crossing the street within the first preset time period in the future based on the LSTM neural network model specifically includes the following steps: Step 401: Acquire 11 current input features of vehicles and pedestrians through multiple sensors; Step 402: performing data normalization on the acquired input features and importing them into the trained LSTM neural network model; Step 403: Output the predicted data and denormalize it to obtain the movement position of the pedestrian crossing the street within the first preset time period in the future, that is, the predicted trajectory of the pedestrian crossing the street within the first preset time period in the future.

Images