Diffusion model trajectory planning method based on historical guidance and collision suppression

Abstract

The invention discloses a diffusion model trajectory planning method based on historical guidance and collision suppression. The method comprises the following steps: collecting own vehicle information, environment information and routing lane information in normal vehicle driving data; constructing negative sample scene data containing vehicle collision; during model training, encoding static obstacle states, dynamic obstacle historical tracks and surrounding lane structure information in a current environment; training a diffusion model decoder based on the future trajectory of the own vehicle, inputting the environmental condition information, the routing lane information and the own vehicle planning result of the model at the previous moment as condition information into the decoder, and enhancing the time continuity of the planning result; and training a classifier based on the collision data, and applying negative guidance to a collision category by using a classifier guidance mechanism during reasoning and sampling of the diffusion model so as to suppress a trajectory generation result with a high collision risk. According to the invention, efficient, safe and stable planning track generation can be realized in a real road scene.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving trajectory planning technology, and in particular to a diffusion model trajectory planning method based on historical guidance and collision suppression. Background Technology

[0002] In autonomous driving technology, trajectory planning, as a core module related to vehicle driving safety, efficiency, and passenger comfort, has always been an important research direction. Trajectory planning aims to generate a safe, smooth, and efficient driving trajectory based on various factors such as the vehicle's current position, surrounding environment, and road information. Existing trajectory planning methods can be broadly categorized into rule-based methods, model-based methods, and data-driven methods. With the development of deep learning technology, data-driven methods have gradually become a research hotspot due to their superior adaptability and flexibility.

[0003] In recent years, trajectory planning methods based on diffusion models have gradually emerged. Diffusion models, through a stepwise inverse process from noise to data, can generate high-quality trajectories that conform to environmental conditions, showing significant advantages, particularly in multimodal trajectory generation. Compared with traditional methods, diffusion models can better handle complex environments and dynamic constraints.

[0004] Due to the complexity and dynamism of the autonomous driving environment, the input information of the model at each moment is affected by factors such as environmental changes and sensor noise. If the diffusion model plans the trajectory based solely on the perception data at the current moment, it cannot guarantee the continuity between the current planned trajectory and the historical planned trajectory, nor the consistency of the vehicle's motion trend. Even if the model's planning result at each moment is locally optimal, a large difference between the planned trajectory at a single moment and the planned trajectories at previous moments may lead to sudden speed changes, lane changes, or steering adjustments during actual driving, affecting passenger comfort, vehicle stability, and potentially even posing safety hazards.

[0005] Furthermore, while existing diffusion model-based trajectory planning methods have demonstrated strong capabilities in trajectory generation, most works still focus on generating "positive sample" trajectories, lacking effective modeling of potential collision scenarios. Previous diffusion model training processes primarily relied on environmental conditions and probability distribution loss functions for guidance, failing to adequately incorporate negative sample data where collisions occurred.

[0006] In summary, there is an urgent need for a method to enhance the temporal consistency of autonomous driving trajectory planning and to suppress high-risk trajectory output by incorporating dangerous driving data, thereby ensuring the safety and comfort needs of passengers. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies in terms of temporal consistency and domain safety in autonomous driving trajectory planning, this invention provides a diffusion model trajectory planning method based on historical guidance and collision suppression, aiming to improve the performance of diffusion models in autonomous driving trajectory planning tasks. This method not only enhances the temporal consistency of the model's output planning results by introducing historical planned trajectory conditions, but also guides and suppresses dangerous generated trajectories through a classifier during the diffusion model's inference sampling process, thereby achieving efficient, safe, and stable planned trajectory generation in real-world road scenarios.

[0008] A trajectory planning method for a diffusion model based on history guidance and collision suppression includes: (1) Collect vehicle information and environmental information from normal vehicle driving data, and process them by historical time, current time and future time. Vehicle information includes vehicle current time data and future trajectory data, and environmental information includes the current static obstacle status, the historical trajectory and current data of dynamic obstacles, and the surrounding lane structure information at the current time. At the same time, based on the breadth-first search algorithm, obtain the shortest lane segment structure information sequence from the current position to the end of the future trajectory, and use it as the routing lane information for subsequent steps. (2) Construct negative sample scene data containing vehicle collisions based on normal vehicle driving data, including abnormal acceleration data or non-deceleration data when there is an obstacle in front; (3) During model training, the static obstacle state, dynamic obstacle historical trajectory and surrounding lane structure information in the current environment are encoded, and the resulting environmental condition encoding is used as the environmental condition information of the diffusion model decoder. (4) When training the diffusion model decoder based on the future trajectory of the vehicle, environmental condition information and route lane information are input into the decoder as condition information. At the same time, the vehicle planning result of the model at the previous time is used as the historical planning condition information at the current time and input into the decoder in chronological order to enhance the temporal consistency and stability of trajectory planning. (5) Based on the negative sample scene data constructed in (2), a classifier is trained. During the inference sampling of the diffusion model, the classifier guidance mechanism is used to apply negative guidance to the collision category in order to suppress the trajectory generation results with high collision risk.

[0009] This invention is applicable to improving the trajectory generation quality of diffusion models in complex scenarios within the field of autonomous driving. The method first collects environmental information during vehicle operation through data preprocessing and uses a breadth-first search algorithm to obtain the shortest lane segment structure information sequence for the vehicle as routing lane information. Next, negative sample scenario data containing collision events is constructed based on normal driving data for subsequent collision classifier training. During diffusion model training, this method encodes environmental information, routing lane information, and historical planning results as conditional inputs to the diffusion model to enhance the rationality, temporal consistency, and stability of trajectory planning. During inference, this method guides the diffusion model through the trained collision classifier to suppress high-collision-risk trajectory generation results.

[0010] In step (1), the current time data and future trajectory data of the vehicle include the vehicle position coordinates, the direction of the vehicle's head, the current static obstacle status and the historical trajectory of the dynamic obstacle, and the current data information of the dynamic obstacle includes the position coordinates, direction, length, width and specific obstacle type. Among them, the static obstacle status only needs to be recorded once as the static data at the current time, while the future trajectory of the vehicle and the historical trajectory of the dynamic obstacle need to be collected and recorded multiple times over a period of time with an interval of 0.1 seconds.

[0011] In the surrounding lane structure information, the lane structure information is a sequence consisting of the position coordinates, direction, width, traffic signal status, and speed limit values ​​of equally spaced sampling points within the lane. The surrounding lanes refer to the lanes included in the environment within a certain distance range at the current moment. In the shortest lane segment structure information sequence, the shortest lane segment refers to the sequence of the fewest lanes on the map that connect the start and end points of the lane segment, given the current position of the vehicle and the lane where the vehicle's future trajectory ends, after breadth-first search. The shortest lane segment structure information sequence is a sequence consisting of the lane structure information corresponding to the lane sequence of the shortest lane segment.

[0012] In step (2), negative sample scenario data is generated by superimposing collision-related perturbations onto the data collected from the normal vehicle driving scenario obtained in step (1), including: The initial velocity components of the vehicle during normal driving at the start of its future trajectory before the collision, obtained by differentiating the vehicle's position coordinates. Add abnormal acceleration : ; Alternatively, maintain the speed from the actual collision avoidance deceleration data just before a collision. No slowdown: ; in, This represents the time elapsed from the start time to the moment the collision occurs. This indicates the speed of the vehicle at the time of the collision.

[0013] These negative sample scene data will be used for model training in subsequent steps to suppress the generation of high-collision-risk trajectories by the model.

[0014] In step (3), the static obstacle state, dynamic obstacle historical trajectory, and surrounding lane structure information in the current environment are first encoded using separate encoders, and then fused by a self-attention encoder to obtain the final environmental condition code. This environmental condition code serves as the input to the diffusion model decoder, guiding the model to generate a planned trajectory based on the environmental information.

[0015] In step (4), the diffusion model decoder adopts a diffusion Transformer structure, which iteratively denoises the trajectory noise using a multi-layer Transformer network, and generates a planned trajectory based on training on the real vehicle trajectory; the ordinary differential equation of the denoising process is: ; in, Indicates time Samples at that time It is a time-step-based coefficient used to control the overall scaling of the sample. These are time-step-based weights used to adjust the gradient distribution. It is the gradient of the log probability density function of the distribution. yes The marginal distribution; the diffusion model uses a neural network to fit the probability density gradient. .

[0016] In step (4), the environmental condition information from step (3), the routing lane information from step (1), and the historical planning trajectory information output by the model at the previous moment are integrated into the future planning through the cross-attention layer and the fully connected layer. If the current moment is the vehicle's starting moment and the vehicle has no historical planning trajectory from the previous moment, then a data with the same format as the planning trajectory is constructed. The position coordinates and vehicle direction at each moment are set to the coordinates and directions of the vehicle in the current starting state, representing the parking state before starting. This data is then input into the decoder. The cross-attention calculation formula in the decoder is as follows: ; Among them, the query matrix Derived from processed random noise representation, the key matrix Sum matrix The encoding representation is derived from environmental information, route lane information, and historical planned trajectory information. It is the vector dimension. It is a function that maps real numbers to attention scores.

[0017] In step (5), a classifier composed of a multi-layer perception mechanism is trained based on negative sample scene data to predict whether the vehicle's future trajectory will collide with obstacles in the future. Then, during the trajectory generation sampling process of the diffusion model, the collision condition category is added to the planned trajectory input classifier. The negative gradient guides and suppresses the generation of trajectories with high collision risk, and its formula is as follows: ; in, The diffusion model is in the first Trajectory noise at each time step In the given condition category Log probability of lower trajectory noise gradient, It is the log probability gradient of the unconditional diffusion model. The classifier determines the category to which the trajectory noise belongs. The logarithmic probability gradient, It is an adjustable suppression intensity parameter.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. Enhanced Temporal Continuity of Trajectory Planning: This invention effectively enhances the temporal consistency of trajectory planning by inputting historical planning results as conditional information for the current moment into the diffusion model decoder. Compared with traditional methods that plan trajectories based on current moment data, this invention generates planned trajectories by combining the trajectory planning results from previous moments with the guided model, ensuring good continuity between the planned trajectory at each moment and those before and after.

[0019] 2. Optimized Driving Comfort: This invention enhances the stability of trajectory planning by incorporating historical trajectory planning conditions, thereby improving vehicle driving comfort. During autonomous driving, frequent speed changes, lane changes, and steering maneuvers can cause passenger discomfort. Guided by historically planned trajectories, the model can smoothly transition between different driving behaviors, thus improving the passenger experience and making driving smoother and more comfortable.

[0020] 3. Enhanced trajectory planning safety and reduced collision risk: This invention effectively reduces the probability of generating high-collision-risk trajectories using existing diffusion model-based autonomous driving trajectory planning methods by introducing a collision negative classifier guidance mechanism. By constructing negative sample scene data containing collision behavior and introducing negative gradient guidance in the diffusion model decoder inference sampling, this invention actively suppresses the generation of high-collision-risk trajectories. This mechanism improves the safety of vehicle trajectory generation in complex environments and reduces the probability of collision accidents. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the method of the present invention.

[0023] Figure 2 This is a schematic diagram of the framework corresponding to the method of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.

[0026] Taking a typical urban road autonomous driving scenario as an example, the specific implementation process of the method of the present invention will be described below in conjunction with this scenario. In this embodiment, the vehicle type and sensor configuration are not limited, but a safe and continuous planned trajectory is generated based on the environmental perception results. Assume that on a certain urban road, an autonomous vehicle needs to travel from the starting point along the lane to the target location. During this process, the road environment includes static obstacles such as parked vehicles and construction fences, as well as dynamic obstacles such as moving vehicles and pedestrians. The road has multiple lanes and intersections, and has traffic lights and speed limits.

[0027] like Figure 1 and Figure 2 As shown, a trajectory planning method for a diffusion model based on history guidance and collision suppression includes the following steps: Step S1: Acquisition of urban road vehicle and environmental information and extraction of shortest lane segment structure information sequence.

[0028] In urban scenarios, autonomous vehicles acquire information about themselves and their environment through a perception system, and this information is segmented and stored according to historical time, current time, and future time. Specifically, this information includes: (1) current time data and future trajectory data of the vehicle, including the vehicle's position coordinates and heading; (2) the current state of static obstacles, including the position coordinates, heading, length and width of stationary vehicles, and the specific type of obstacle; (3) historical trajectory data and current data of dynamic obstacles, such as the position coordinates, heading, length and width of vehicles traveling in adjacent lanes ahead, and a sequence of obstacle type information; (4) surrounding lane structure information, including the coordinates of sampling points on multiple lanes within 100 meters of the vehicle at the current time, lane direction, width, traffic light status, and speed limit sequence. Among them, historical data and future data are separated by 0.1 seconds, with historical data at 2 seconds and future data at 8 seconds. In addition, based on the above information, a breadth-first search algorithm is used to obtain the shortest lane-level path from the vehicle's current position to the predetermined target position on the road topology map. This road topology treats each lane segment as a node, with adjacent lane segments connected to form a graph structure. A breadth-first search algorithm searches this graph for the shortest path, for example, obtaining the lane segment sequence Lane5→Lane6→Lane7, and then converts this lane segment into corresponding lane structure information. This shortest path, as routing lane information, will provide constraints in subsequent trajectory planning.

[0029] Step S2: Constructing the trajectory of the collision negative sample scene.

[0030] Based on the obtained normal driving data, to improve the model's ability to suppress collision risks, this step constructs negative sample scenario data for typical dangerous scenarios on urban roads. First, the normal driving trajectory is analyzed: for example, after the vehicle detects a stopped vehicle 30 meters ahead, its future normal driving trajectory will show gradual deceleration until it stops to avoid a collision. To construct the vehicle's future trajectory in the negative sample scenario, this embodiment applies an abnormal perturbation to the normal trajectory obtained in step S1: when an obstacle exists ahead and the vehicle's normal trajectory should involve deceleration, two types of abnormal trajectory data are generated. One is introducing abnormal acceleration at the point where deceleration should occur, simulating the dangerous behavior of improper acceleration. For example, if the vehicle's original speed is approximately 10 m / s and it should gradually decelerate to 0 under normal circumstances, a 2 m / s speed is introduced into the negative sample scenario. 2The abnormal trajectory data obtained from these two processes are used to simulate two scenarios: one is abnormal acceleration leading to a collision; the other is maintaining the original speed when approaching a collision, simulating a situation where the driver does not brake, for example, the vehicle moves forward at a constant speed of 10 m / s without decelerating, until it is very close to the obstacle or even collides. These abnormal trajectory data are used as the vehicle's future trajectory in the negative collision scenario data, and are used to train a multilayer perceptron classifier to predict whether the vehicle's future trajectory will collide with the obstacle in the future. Similarly, negative collision data is also constructed for dynamic obstacle scenarios. For example, for a slow-moving vehicle ahead, the normal trajectory would follow and decelerate, while the negative sample scenario's vehicle trajectory does not decelerate, leading to a dangerous situation; for a pedestrian crossing, the normal trajectory should stop and yield, while the negative sample scenario's vehicle trajectory maintains its speed, leading to a dangerous situation.

[0031] Step S3: Environmental condition coding.

[0032] The environmental information obtained in step S1 is encoded to generate an environmental feature representation for conditional guidance. In this embodiment, independent encoder networks are used for different types of information. Static obstacle information, dynamic obstacle historical trajectory sequences, and surrounding lane structure information are processed into high-dimensional vectors by three different multilayer perceptron networks and then fed into a self-attention fusion module. The attention mechanism integrates the information from each part to form a unified environmental condition encoding vector. During this fusion process, the model can focus on key obstacles and road elements. For example, if a static obstacle happens to be located close to the vehicle's path, then the feature of that obstacle will have a higher weight in the fusion result. The final environmental condition encoding will be used as a conditional input and combined with the trajectory generation process in the diffusion model decoder to guide the planned trajectory to fully consider the current environmental constraints and driving route.

[0033] Step S4: Train the diffusion model by combining environmental conditions and historical planning trajectories.

[0034] This embodiment uses a diffusion Transformer structure as the trajectory generation model, treating the future planned trajectory as a sequence obtained by progressively denoising random noise. In this step, the diffusion model is trained based on the real future trajectory. The environmental condition encoding obtained in step S3, the shortest lane segment routing lane information obtained in step S1, and the historical planned trajectory of the model at the previous moment are used as conditions to train the diffusion model and model the distribution of the planned trajectory. If the current moment is the vehicle's starting moment and the vehicle has no historical planned trajectory from the previous moment, then data with the same format as the planned trajectory is constructed. The position coordinates and vehicle direction at each moment are set to the coordinates and direction of the vehicle in the current starting state, representing the parking plan before starting, as the historical planned trajectory information. Specifically, the diffusion model decoder integrates environmental conditions, routing lane information, and historical planned trajectory information into the trajectory generation process through a multi-head cross-attention module. Let the random noise of the future trajectory processed by the fully connected layer and normalization layer be represented as a query vector, and the condition encoding be a key-value matrix, and calculate the cross-attention. Through the above process, the model can constrain the planned trajectory based on environmental and route information. For example, when the planned trajectory point corresponding to the query vector is near the lane line, the lane structure information contained in the key-value matrix will guide the model to extend the planned trajectory along the lane direction. Furthermore, inputting the planned trajectory generated by the model in the previous time step into the decoder as a reference for the current time step trajectory generation ensures that the planned trajectory not only conforms to environmental constraints but also connects with historical planned trajectories. For instance, if the model planned the vehicle to travel straight along the current lane and decelerate in the previous time step, then in the current time step planning, the model will tend to continue traveling straight and decelerate smoothly under the guidance of historical trajectories, rather than suddenly accelerating or changing lanes. Through this training strategy, the model's output trajectories from adjacent time steps will be more continuous in speed and direction. This enhanced temporal consistency significantly improves the smoothness of autonomous driving and ride comfort.

[0035] Step S5: Train a collision classifier based on negative sample scene data, and use the classifier to guide the suppression of high-risk collision trajectory generation during actual inference of the diffusion model.

[0036] In this embodiment, to ensure the safety of the planned trajectory during actual inference, a classifier-guided collision risk suppression mechanism is introduced during the diffusion model sampling process. Specifically, a trajectory collision category classifier is pre-trained to predict whether a given trajectory has a collision risk. At each denoising time step of the trajectory generated by the diffusion model sampling, it is predicted whether the currently generated trajectory belongs to the collision category. The gradient information of the collision category classifier is used to control and correct the evolution direction of the diffusion process, penalizing the model's tendency to generate collision trajectories.

[0037] Through the above steps, this embodiment demonstrates in detail the implementation process of the end-to-end trajectory planning method of the present invention in the context of an urban road scenario. This method utilizes a diffusion model to generate trajectories. While integrating environmental perception information and routing conditions, it introduces historical planned trajectories to ensure the continuity and consistency of planning results at each time step, and reduces the generation of dangerous trajectories through collision negative classification. Ultimately, the obtained planned trajectory ensures smooth trajectory changes at continuous time steps and meets the safety and comfort requirements of autonomous driving on actual urban roads.

[0038] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A trajectory planning method for a diffusion model based on historical guidance and collision suppression, characterized in that, include: (1) Collect vehicle information and environmental information from normal vehicle driving data, and process them by historical time, current time and future time. Vehicle information includes current time data and future trajectory data of the vehicle. Environmental information includes the current static obstacle status, the historical trajectory and current data of dynamic obstacles, and the surrounding lane structure information at the current time. At the same time, based on the breadth-first search algorithm, obtain the shortest lane segment structure information sequence from the current position to the end of the future trajectory of the vehicle, as the routing lane information. (2) Construct negative sample scenario data containing vehicle collisions based on normal vehicle driving data, including data on abnormal acceleration or no deceleration of the vehicle when there is an obstacle in front; (3) During model training, the static obstacle state, dynamic obstacle historical trajectory and surrounding lane structure information in the current environment are encoded, and the resulting environmental condition encoding is used as the environmental condition information of the diffusion model decoder. (4) When training the diffusion model decoder based on the future trajectory of the vehicle, environmental condition information and route lane information are input into the decoder as condition information. At the same time, the vehicle planning result of the model at the previous time is used as the historical planning condition information at the current time and input into the decoder in chronological order to enhance the temporal consistency and stability of trajectory planning. (5) Based on the negative sample scene data constructed in (2), a classifier is trained. During the inference sampling of the diffusion model, the classifier guidance mechanism is used to apply negative guidance to the collision category in order to suppress the trajectory generation results with high collision risk.

2. The diffusion model trajectory planning method based on history guidance and collision suppression according to claim 1, characterized in that, In step (1), the current time data and future trajectory data of the vehicle include the vehicle position coordinates, the direction of the vehicle's head, the current static obstacle status and the historical trajectory of the dynamic obstacle, and the current data information of the dynamic obstacle includes the position coordinates, direction, length, width and specific obstacle type. Among them, the static obstacle status only needs to be recorded once as the static data at the current time, while the future trajectory of the vehicle and the historical trajectory of the dynamic obstacle need to be collected and recorded multiple times over a period of time with an interval of 0.1 seconds.

3. The diffusion model trajectory planning method based on history guidance and collision suppression according to claim 1, characterized in that, In step (1), the surrounding lane structure information is a sequence of position coordinates, direction, width, traffic signal status, and speed limit values ​​of equally spaced sampling points within the lane. The surrounding lanes refer to the lanes included in the environment within a certain distance range at the current moment. In the shortest lane segment structure information sequence, the shortest lane segment refers to the minimum number of lanes on the map that connect the starting point and the ending point after a breadth-first search is performed, given the current position of the vehicle and the lane where the vehicle's future trajectory ends. The shortest lane segment structure information sequence is a sequence of lane structure information corresponding to the lane sequence of the shortest lane segment.

4. The diffusion model trajectory planning method based on history guidance and collision suppression according to claim 1, characterized in that, In step (2), negative sample scenario data is generated by superimposing collision-related perturbations onto the data collected from the normal vehicle driving scenario obtained in step (1), including: The initial velocity components of the vehicle during normal driving at the start of its future trajectory before the collision, obtained by differentiating the vehicle's position coordinates. Add abnormal acceleration : ； Alternatively, maintain the speed from the actual collision avoidance deceleration data just before a collision. No slowdown: ； in, This represents the time elapsed from the start time to the moment the collision occurs. This indicates the speed of the vehicle at the time of the collision.

5. The diffusion model trajectory planning method based on history guidance and collision suppression according to claim 1, characterized in that, In step (3), the static obstacle state, dynamic obstacle historical trajectory, and surrounding lane structure information in the current environment are first encoded based on separate encoders, and then fused by a self-attention encoder to obtain the final environmental condition code.

6. The diffusion model trajectory planning method based on history guidance and collision suppression according to claim 1, characterized in that, In step (4), the diffusion model decoder adopts a diffusion Transformer structure, which iteratively denoises the trajectory noise using a multi-layer Transformer network, and generates a planned trajectory based on training on the real vehicle trajectory; the ordinary differential equation of the denoising process is: ； in, Indicates time Samples at that time It is a time-step-based coefficient used to control the overall scaling of the sample. These are time-step-based weights used to adjust the gradient distribution. It is the gradient of the log probability density function of the distribution. yes The marginal distribution; the diffusion model uses a neural network to fit the probability density gradient. .

7. The diffusion model trajectory planning method based on history guidance and collision suppression according to claim 1, characterized in that, In step (4), environmental condition information, route lane information, and historical planning trajectory information output by the model at the previous moment are incorporated into the future planning through independent cross-attention layers and fully connected layers, respectively. The cross-attention calculation formula in the decoder is as follows: ； Among them, the query matrix Derived from processed random noise representation, the key matrix Sum matrix The encoding representation is derived from environmental condition information, route lane information, and historical planned trajectory information. It is the vector dimension. It is a function that maps real numbers to attention scores.

8. The diffusion model trajectory planning method based on history guidance and collision suppression according to claim 1, characterized in that, In step (4), the historical planning results of the previous moment are input into the diffusion model decoder in the form of a trajectory point sequence as historical condition information to guide the model generation process. The decoder injects the historical condition information into the planning trajectory generation process through a cross-attention mechanism. If the current moment is the vehicle's starting moment and the vehicle has no historical planned trajectory from the previous moment, then a data set with the same format as the planned trajectory is constructed. The position coordinates and vehicle direction at each moment are set to the coordinates and direction of the vehicle in the current starting state, representing the parking plan before starting, and this is input into the decoder as historical condition information.

9. The trajectory planning method for a diffusion model based on history guidance and collision suppression according to claim 1, characterized in that, In step (5), a classifier composed of a multi-layer perception mechanism is trained based on negative sample scene data to predict whether the vehicle's future trajectory will collide with obstacles in the future. Then, during the trajectory generation sampling process of the diffusion model, the collision condition category is added to the planned trajectory input classifier. The negative gradient guides and suppresses the generation of trajectories with high collision risk, and its formula is as follows: ； in, The diffusion model is in the first Trajectory noise at each time step In the given condition category Log probability of lower trajectory noise gradient, It is the log probability gradient of the unconditional diffusion model. The classifier determines the category to which the trajectory noise belongs. The logarithmic probability gradient, It is an adjustable suppression intensity parameter.

Images