Decision planning system and decision planning device for complex urban road conditions
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NINGBO JUNSHENG INTELLIGENT AUTOMOBILE TECH RES INST CO LTD
- Filing Date
- 2023-01-09
- Publication Date
- 2026-07-14
Smart Images
Figure CN116164748B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of automation, and more specifically, to a decision-making and planning system and device for complex road conditions in urban areas. Background Technology
[0002] Currently, the mainstream autonomous driving algorithm technology stack is mainly a layered "perception-localization-decision planning-control" model, with each module relatively independent and decoupled. From the perspective of parallel development efficiency, it is obviously impossible to leave code problems to be exposed during joint debugging of multiple modules in a real vehicle. Each module should first be responsible for its own output. The problem with the decision planning module is that, unlike the perception module, it does not have a clear true value label for the perceived target object (e.g., whether the currently seen object is a truck or a car, it is easy to determine if the perception algorithm output does not match the reality). The localization and control modules have their own localization accuracy and control accuracy as evaluation indicators. However, for the output of the decision planning module, apart from the safety requirement of avoiding collisions, there is a lack of a unique and correct true value as a reference. Therefore, the performance of the decision planning algorithm still largely depends on the joint debugging effect with the control module in a real vehicle. This is relatively detrimental to the agile development and rapid iteration of the algorithm.
[0003] Furthermore, in current mass production, due to limitations in computing power and hardware costs, the decision-planning algorithms used cannot yet effectively balance solution efficiency and the ability to handle complex scenarios. For example, in a certain urban road, assuming there are more than twenty cars traveling at different speeds around the current vehicle, traditional decision-planning algorithms struggle to simultaneously perform interactive trajectory planning for multiple traffic participants, causing their strategies to tend to be conservative and making it difficult to search for and solve for trajectories with higher driving efficiency. Summary of the Invention
[0004] Therefore, embodiments of the present invention provide a decision planning system and device for complex road conditions in urban areas, which solves the problems of low efficiency and security in existing algorithms.
[0005] To address the aforementioned issues, this invention provides a decision-making and planning system for complex urban road conditions, comprising: detecting driving parameters of a vehicle and surrounding environmental parameters to obtain multiple raw data; importing the multiple raw data into an AI model for processing to obtain multiple perception results; calculating a planned trajectory based on the multiple perception results to obtain a first planned trajectory result; the AI model obtaining a second planned trajectory result based on the multiple raw data; importing the first planned trajectory result and the second planned trajectory result into a trajectory scoring and arbitration module to verify the optimal trajectory and output the optimal trajectory.
[0006] Compared to existing technologies, the technical effects achieved by this solution are as follows: By setting and detecting the vehicle's driving parameters and surrounding environmental parameters to obtain raw data, the data source is made more closely aligned with the actual driving conditions of the vehicle. Subsequent trajectory planning is then performed based on these conditions, making the trajectory more consistent with the current driving environment. Furthermore, by importing the raw data into an AI model for processing and obtaining perception results, traditional trajectory calculations are performed to obtain the first planned trajectory result. Then, the AI model further processes the raw data to obtain the second planned trajectory result. This provides more methods for trajectory calculation, allowing for different trajectories to be obtained through different approaches, thus enhancing the practicality and safety of the trajectory. Simultaneously, the two planned trajectory results are imported into the scoring and arbitration module to verify and obtain the optimal trajectory, minimizing trajectory risk and improving algorithm efficiency. This allows for faster determination of the vehicle's optimal driving trajectory under the current driving environment, ensuring the safety of both the driver and the vehicle. Furthermore, AI models are used to process the data and generate different trajectory models, making the trajectory planning of the autonomous driving system more comprehensive, safe, and complete. The AI model can also better react and make decisions in response to the current driving environment, resulting in higher algorithm efficiency. This enables the vehicle to safely perform autonomous driving even in complex driving environments, improving the safety and practicality of autonomous driving and ensuring driver safety.
[0007] In one embodiment of the present invention, multiple raw data are imported into an AI model for processing to obtain multiple perception results. The method also includes: using a convolutional neural network to preprocess the multiple raw data and extract local features to obtain multiple feature vectors.
[0008] Compared with existing technologies, the technical effects achieved by this solution are as follows: By using convolutional neural networks to preprocess and extract local features from the raw data, impurities and interference factors can be quickly removed from the raw data to obtain the required data features. At the same time, local feature extraction quickly extracts the features required by the system and converts them into corresponding feature vectors for subsequent operations. This makes the algorithm more efficient, the process more accurate, and the results clearer. It also makes subsequent trajectory planning more convenient and safer, making the decision-making and planning system for complex urban road conditions more practical and safer, and ensuring the safety of drivers.
[0009] In one embodiment of the present invention, the AI model obtains a second planned trajectory result based on multiple raw data, and further includes: merging multiple feature vectors through a concatenation operation to obtain a merged feature vector; and using a fully connected neural network structure and a Transformer network structure as intermediate layers to perform feature transformation and encoding on the merged feature vector to obtain an intermediate result.
[0010] Compared with existing technologies, the technical effects achieved by this solution are as follows: by setting up the concatenation and merging of multiple feature vectors, subsequent operations and processing become more convenient and faster, and the efficiency of subsequent algorithms is also improved. Furthermore, by using a fully connected neural network structure and a Transformer network structure to perform feature transformation and encoding on the merged feature vectors, intermediate results are obtained, enabling the merged feature vectors to undergo subsequent processing operations. Moreover, the transformation process is more efficient, and the intermediate results obtained can be used directly, making it more convenient and improving the efficiency of the algorithm.
[0011] In one embodiment of the present invention, the AI model obtains a second planned trajectory result based on multiple raw data, and further includes: inputting intermediate results into an LSTM Decoder network to generate a series of trajectory sequences and pre-generating decision planning trajectory points; processing the decision planning trajectory points through a trajectory model and outputting the second planned trajectory result.
[0012] Compared with existing technologies, the technical effects achieved by this solution are as follows: By inputting intermediate results into the LSTM Decoder network to generate a series of trajectory sequences, the LSTM Decoder network can better predict the trajectories. Trajectories are a set of continuous data information with temporal continuity and spatial randomness, and the LSTM Decoder network can effectively predict and process them, making trajectory prediction more accurate. The trajectory model processes the decision planning trajectory points and outputs the second planning trajectory result. The combination of the model and the LSTM Decoder network makes the model more accurate, resulting in higher accuracy of subsequent predictions. This makes the decision planning system more comprehensive and safer, better protects the safety of drivers, and makes autonomous driving more intelligent.
[0013] In one embodiment of the present invention, the first planned trajectory result and the second planned trajectory result are imported into the trajectory scoring arbitration module to verify the best trajectory and output the best trajectory. The method also includes: performing trajectory prediction through the prediction module to obtain the predicted trajectory, and evaluating the credibility of the predicted trajectory to obtain the third planned trajectory result; and performing collision screening on the third planned trajectory result and the second planned trajectory result to obtain the fourth planned trajectory result.
[0014] Compared with existing technologies, the technical effects achieved by this solution are as follows: By setting up a prediction module to predict the trajectory, a predicted trajectory is obtained. The reliability of the predicted trajectory is then evaluated to obtain a third planned trajectory result. This allows the predicted trajectory to assist the AI model in judging and evaluating the trajectory generated. Combining multiple trajectories makes the final output trajectory more comprehensive and accurate, better reflecting the current driving environment and protecting driver safety. The third and second planned trajectory results are then compared and filtered to obtain a fourth planned trajectory result. This improves the overall solution efficiency of the algorithm, avoiding the expenditure of significant computational resources on invalid trajectories for complete feasibility checks and comfort ratings; it also serves as a safety screening. Furthermore, collisions with low-reliability trajectory points are ignored. Valid self-planned trajectories are not blindly deleted due to some low-reliability predicted trajectories, thus expanding the scope and completeness of the algorithm's solution space.
[0015] In one embodiment of the present invention, the first and fourth planned trajectory results are evaluated for comfort, anthropomorphism, and efficiency; the trajectory with the highest comfort, anthropomorphism, and efficiency scores is output.
[0016] Compared with existing technologies, the technical effects achieved by adopting this technical solution are as follows: by setting the first and fourth planned trajectory results to be rated for comfort, anthropomorphism, and efficiency, the output results are more representative and practical after being rated by the three criteria, and are more in line with the current driving conditions of the vehicle, so that the trajectory can ensure the safe driving of the vehicle and the safety of the driver.
[0017] In one embodiment of the present invention, the comfort score further includes: comfort = normalized (1 / trajectory curvature) + normalized (1 / rate of change of acceleration).
[0018] Compared with existing technologies, the technical effects achieved by adopting this technical solution are as follows: the trajectory is evaluated by setting a comfort level, which is mainly measured by two indicators: the trajectory curvature change rate and the planned acceleration change rate. That is, under the premise of meeting safety requirements, the lower the trajectory curvature change rate and acceleration change rate, the better, so that the trajectory with a high score is safer and more practical.
[0019] In one embodiment of the present invention, the efficiency score also includes: efficiency = normalized (total trajectory length).
[0020] Compared with existing technologies, the technical effects achieved by adopting this technical solution are as follows: by setting efficiency to score the trajectory, the longer the path, the higher the speed and the better the efficiency, so that the trajectory obtained after efficiency scoring can travel better at the current intersection.
[0021] In one embodiment of the present invention, the anthropomorphism score also includes: anthropomorphism = normalized (1 / Freche distance (the average trajectory of the human driver in the corresponding scene, the target trajectory)).
[0022] Compared with existing technologies, the technical effects achieved by adopting this technical solution are as follows: by setting anthropomorphism to score the trajectory, the trajectory can be more in line with human standards and the driving habits of existing drivers. This ensures the best route is taken under the current driving conditions, while making the trajectory more humanized and ensuring driving safety.
[0023] The present invention also provides a decision-making and planning device, comprising: an acquisition module for detecting driving parameters of a vehicle and surrounding environmental parameters to obtain multiple raw data; a processing module for importing the multiple raw data into an AI model for processing to obtain multiple perception results; calculating a planning trajectory based on the multiple perception results to obtain a first planning trajectory result; the AI model obtaining a second planning trajectory result based on the multiple raw data; and a scoring module for importing the first planning trajectory result and the second planning trajectory result into a trajectory scoring arbitration module to verify the optimal trajectory and output the optimal trajectory.
[0024] Compared with existing technologies, the technical effects achieved by this solution are as follows: By setting up an acquisition module to obtain the current driving environment of the vehicle, raw data is obtained, making the vehicle trajectory planning more consistent with the actual environment and making the planning more practical. The processing module processes the raw data to obtain the first and second planned trajectory results, making the two trajectory results more accurate and consistent with the actual situation. At the same time, the evaluation and comparison of multiple trajectory results makes the final output trajectory safer and more practical. Furthermore, the scoring module scores and verifies the two trajectories, ensuring that the output trajectory for returning home is the safest, fastest, and most user-friendly. Through the cooperation of these three modules, the algorithm becomes more efficient, can take into account different driving environments, and has higher practicality.
[0025] The following benefits can be obtained by adopting this technical solution:
[0026] (1) By setting the driving parameters of the detection vehicle and the surrounding environment parameters, the raw data is obtained, making the data source more in line with the actual driving situation of the vehicle. The subsequent trajectory planning is carried out according to the actual situation, so that the trajectory is more in line with the current driving environment of the vehicle. The raw data is imported into the AI model for processing and the perception results are obtained. The traditional trajectory calculation is then performed to obtain the first planned trajectory result. The AI model is then used to process the raw data to obtain the second planned trajectory result, making the trajectory calculation more diverse. Different trajectories are obtained through different methods, making the trajectory more practical and safer. At the same time, the two planned trajectory results are imported into the scoring arbitration module to verify and obtain the best trajectory, minimizing the risk of the trajectory and improving the efficiency of the algorithm. The best driving trajectory of the vehicle in the current driving environment can be obtained more quickly, ensuring the safety of the driver and the vehicle. At the same time, the AI model is used to process the data to obtain different trajectory models, making the trajectory planning of the autonomous driving system more comprehensive, safe, and complete. The AI model can better react to the current driving environment and make decisions. The algorithm is more efficient, enabling the vehicle to safely drive in complex driving environments, improving the safety and practicality of autonomous driving and ensuring the safety of the driver.
[0027] (2) By setting up and using convolutional neural networks to preprocess the original data and extract local features, the original data can be quickly cleaned of impurities and interference factors to obtain the required data features. At the same time, local feature extraction can quickly extract the features required by the system and convert them into corresponding feature vectors for subsequent operations, making the algorithm more efficient, the process more accurate, and the results clearer. This makes subsequent trajectory planning more convenient and safer, making the decision planning system for complex urban road conditions more practical and safe, and ensuring the safety of drivers.
[0028] (3) By setting up a prediction module to predict the trajectory, the predicted trajectory is obtained, and the credibility of the predicted trajectory is evaluated to obtain the third planning trajectory result. The trajectory generated by the AI model is judged and evaluated by using the predicted trajectory. Combining multiple trajectories makes the final output trajectory more comprehensive and accurate, and more in line with the current environment of the driving vehicle, thus protecting the safety of the driver. The third planning trajectory result and the second planning trajectory result are then compared and screened to obtain the fourth planning trajectory result. On the one hand, this improves the overall solution efficiency of the algorithm and avoids spending a lot of computing power on invalid trajectories to perform complete feasibility checks and comfort scores; on the other hand, it also performs a safety screening. At the same time, the collision of some low credibility trajectory points is ignored. The effective self-planned trajectory will not be blindly deleted because of some low credibility predicted trajectories, thus expanding the scope and completeness of the algorithm's solution space.
[0029] (4) By setting the first and fourth planning trajectory results to be rated for comfort, anthropomorphism and efficiency, the output results are more representative and practical after being rated by the three criteria, and are more in line with the current driving conditions of the vehicle, so that the trajectory can ensure the safe driving of the vehicle and the safety of the driver. Attached Figure Description
[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings to be used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.
[0031] Figure 1 The flowchart of a decision-making and planning system for complex urban road conditions provided by the present invention.
[0032] Figure 2 This is a schematic diagram of a decision-making and planning device provided by the present invention.
[0033] Figure 3 This invention provides a detailed flowchart of a decision-making and planning system for complex urban road conditions.
[0034] Figure 4 This is a schematic diagram of the multi-task unified AI model provided by the present invention.
[0035] Figure 5 This invention provides a trajectory scoring arbitration flowchart for a decision-making and planning system for complex urban road conditions.
[0036] Figure 6 This is a schematic diagram of the trajectory model provided by the present invention.
[0037] Explanation of reference numerals in the attached figures:
[0038] 100 is the decision-making and planning device; 110 is the acquisition module; 120 is the processing module; 130 is the scoring module. Detailed Implementation
[0039] To make the above-mentioned objectives, features, and advantages of the present invention more apparent and understandable, the technical solutions in the embodiments of the present invention are clearly and completely described. 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.
[0040] See Figure 1-6This invention provides a decision-making and planning system for complex urban road conditions, comprising:
[0041] Step S100: Detect the driving parameters of the vehicle and the surrounding environment parameters to obtain multiple raw data;
[0042] Step S200: Import multiple raw data into the AI model for processing to obtain multiple perception results; calculate the planned trajectory using the multiple perception results to obtain the first planned trajectory result; the AI model obtains the second planned trajectory result based on the multiple raw data.
[0043] Step S300: Import the first planned trajectory result and the second planned trajectory result into the trajectory scoring arbitration module to verify the best trajectory and output the best trajectory.
[0044] Specifically, based on the original sensor data and input of the vehicle's driving parameters and surrounding environmental parameters, the unified multi-task AI model proposed in this invention is used for processing. The resulting perception results include: target detection results for traffic participants such as vehicles, bicycles, and pedestrians; target perception results for lane lines, traffic signs, and traffic lights; predicted behavioral trajectories of traffic participants; and the positions of other stationary obstacles. These results are then directly output to a traditional decision-making and planning algorithm for corresponding trajectory calculation, yielding the first planned trajectory result. Simultaneously, the AI model also directly outputs the second planned trajectory result, obtained through a data-driven approach. Both trajectories undergo a trajectory scoring and arbitration module before final output.
[0045] Furthermore, a unified AI model framework with multi-task capabilities is provided: at the raw input end, it supports inputs from various types of sensors (visual cameras, LiDAR, millimeter-wave radar, etc.) and multiple sensors (such as forward-facing cameras, side-facing cameras, and rear-facing cameras). For inputs from different sensors, corresponding CNNs (Convolutional Neural Networks) are used for preprocessing and local feature extraction. Then, a Concat operation is used to merge the feature vectors extracted from different sensors. Subsequently, a DNN (fully connected neural network) structure and a Transformer network structure are used as intermediate layers to further transform and encode the merged feature vectors, resulting in the intermediate layer of the entire AI model. This intermediate layer result can then be followed by typical task network blocks, such as prediction, object detection, and lane segmentation, to output corresponding multi-task perception results. Simultaneously, the intermediate layer result can be directly output to an LSTM Decoder network to generate a series of trajectory sequences for pre-generating decision-making and planning trajectory points. The pre-generated trajectory points are then processed and output using the proposed TrajectoryModel. All these modules share the same intermediate layer result, maximizing the reuse of the trained AI network model.
[0046] Specifically, this invention proposes a tree-like decision-planning trajectory representation model, namely the Trajectory Model, which ultimately requires a tree data structure for implementation. Functionally, the tree-like decision-planning trajectory model proposed in this invention can better handle multimodal decision-planning problems. That is, in the same scenario, the vehicle has multiple feasible solutions, while other vehicles also have multiple possibilities for future driving behaviors. The tree-like structure can fully capture the multiple possibilities of the vehicle's trajectory and its interaction with other vehicles. Finally, this module will transform the trajectory points generated by LSTM into a tree-like trajectory for output, representing the multiple possibilities of the vehicle's future state.
[0047] Furthermore, the prediction results from the prediction module (i.e., for multiple traffic participants around the vehicle, say 10) are evaluated for credibility. If the predicted trajectory, i.e., the third planning trajectory result, is in the form of 10 trajectory coordinate point sequences [a, b, c, d, e, f], then this module will assign a corresponding credibility discount to each point in these trajectories, such as [(a, 0.95), (b, 0.9), (c, 0.85), d (0.80), e (0.75), f (0.7)]. The predicted trajectory after credibility evaluation will be compared with the tree-like trajectory model output by the AI model for collision screening. This step mainly considers the possibility of collisions between high-credibility (e.g., above 0.8) predicted trajectory points and the vehicle's tree-like trajectory. Trajectory trees with high-credibility potential collision risks are directly pruned, i.e., the corresponding branches are deleted from the tree structure. This improves the overall solution efficiency of the algorithm and avoids collisions with unreliable data. The algorithm spends significant computational resources on performing a complete feasibility check and comfort score on the effective trajectory; this also serves as a safety screening. Collisions at trajectory points with low confidence (e.g., below 0.3) may be ignored. This step prevents the blind deletion of valid planned trajectories due to low-confidence predictions, expanding the scope and completeness of the algorithm's solution space. Finally, the trajectory tree after collision screening (the fourth planning trajectory result) and the traditional decision-planning trajectory (the first planning trajectory result), which has already undergone comprehensive collision detection, pass through three subsequent scoring stages (comfort, anthropomorphism, and efficiency). Ultimately, the single trajectory sequence with the highest total score is selected for output.
[0048] Preferably, raw data is obtained by setting and detecting the vehicle's driving parameters and surrounding environmental parameters. This ensures that the data source is more closely aligned with the actual driving conditions of the vehicle. Based on these conditions, subsequent trajectory planning is performed, making the trajectory more consistent with the current driving environment. The raw data is then imported into an AI model for processing to obtain perception results. This is used for traditional trajectory calculation to obtain a first planned trajectory result. The AI model then processes the raw data again to obtain a second planned trajectory result, providing more ways to calculate the trajectory and obtain different trajectories. This enhances the practicality and safety of the trajectory. The two planned trajectory results are then imported into a scoring and arbitration module to verify and obtain the optimal trajectory, minimizing trajectory risk and improving algorithm efficiency. This allows for faster determination of the optimal driving trajectory for the vehicle in the current driving environment, ensuring the safety of both the driver and the vehicle. Furthermore, processing data using an AI model to obtain different trajectory models makes the trajectory planning of the autonomous driving system more comprehensive, safe, and complete. The AI model can better react and make decisions based on the current driving environment, resulting in higher algorithm efficiency. This allows the vehicle to safely perform autonomous driving even in complex driving environments, improving the safety and practicality of autonomous driving and ensuring driver safety.
[0049] Specifically, multiple raw data are imported into an AI model for processing to obtain multiple perception results. This also includes using a convolutional neural network to preprocess and extract local features from multiple raw data to obtain multiple feature vectors.
[0050] Preferably, by using a convolutional neural network to preprocess the raw data and extract local features, the raw data can be quickly cleaned of impurities and interference factors to obtain the required data features. At the same time, local feature extraction can quickly extract the features required by the system and convert them into corresponding feature vectors for subsequent operations. This makes the algorithm more efficient, the process more accurate, and the results clearer. It also makes subsequent trajectory planning more convenient and safer, making the decision-making and planning system for complex urban road conditions more practical and safer, and ensuring the safety of drivers.
[0051] Specifically, the AI model obtains the second planned trajectory result based on multiple raw data, and also includes: merging multiple feature vectors through a concatenation operation to obtain a merged feature vector; and using a fully connected neural network structure and a Transformer network structure as intermediate layers to perform feature transformation and encoding on the merged feature vector to obtain an intermediate result.
[0052] Preferably, by setting up the concatenation and merging of multiple feature vectors, subsequent operations and processing become more convenient and faster, while the efficiency of subsequent algorithms is also improved. Then, a fully connected neural network structure and a Transformer network structure are used to perform feature transformation and encoding on the merged feature vectors to obtain intermediate results. This allows the merged feature vectors to be used for subsequent processing operations, and the transformation process is more efficient. At the same time, the intermediate results can be used directly, which is more convenient and makes the algorithm more efficient.
[0053] Specifically, the AI model obtains the second planning trajectory result based on multiple raw data, and also includes: inputting the intermediate results into the LSTM Decoder network to generate a series of trajectory sequences and pre-generating decision planning trajectory points; processing the decision planning trajectory points through the trajectory model and outputting the second planning trajectory result.
[0054] Preferably, by inputting intermediate results into an LSTM Decoder network to generate a series of trajectory sequences, the LSTM Decoder network can better predict trajectories. Trajectories are a set of continuous data information with temporal continuity and spatial randomness. The LSTM Decoder network can effectively predict and process these trajectories, making trajectory prediction more accurate. The trajectory model processes the decision-planning trajectory points and outputs the second planning trajectory result. The combination of the model and the LSTM Decoder network can improve the accuracy of the model, making subsequent predictions more accurate, making the decision-planning system more comprehensive and safer, better ensuring the safety of drivers, and making autonomous driving more intelligent.
[0055] Specifically, the process of importing the first and second planned trajectory results into the trajectory scoring arbitration module to verify the best trajectory and output the best trajectory also includes: predicting the trajectory through the prediction module to obtain the predicted trajectory, and evaluating the credibility of the predicted trajectory to obtain the third planned trajectory result; and performing collision screening on the third planned trajectory result and the second planned trajectory result to obtain the fourth planned trajectory result.
[0056] Preferably, a prediction module is set up to predict the trajectory, and the predicted trajectory is evaluated for reliability to obtain a third planned trajectory result. This allows the trajectory generated by the AI model to be evaluated using the predicted trajectory. Combining multiple trajectories makes the final output trajectory more comprehensive and accurate, better reflecting the current driving environment and protecting driver safety. The third and second planned trajectory results are then compared and filtered to obtain a fourth planned trajectory result. This improves the overall solution efficiency of the algorithm, avoiding the expenditure of significant computational resources on invalid trajectories for complete feasibility checks and comfort ratings; it also serves as a safety screening. Furthermore, collisions at some low-reliability trajectory points are ignored. Valid self-planned trajectories are not blindly deleted due to some low-reliability predicted trajectories, thus expanding the scope and completeness of the algorithm's solution space.
[0057] Specifically, the results of the first and fourth planning trajectories are scored for comfort, anthropomorphism, and efficiency; the trajectories with higher scores are output.
[0058] Preferably, by setting the first and fourth planned trajectory results to be rated for comfort, anthropomorphism, and efficiency, the output results are more representative and practical after being rated by the three criteria, and are more in line with the current driving conditions of the vehicle, so that the trajectory can ensure the safe driving of the vehicle and the safety of the driver.
[0059] Specifically, the comfort score also includes: Comfort = Normalized (1 / Track curvature) + Normalized (1 / Rate of change of acceleration).
[0060] Preferably, the trajectory is evaluated by setting a comfort level. The comfort level is mainly measured by two indicators: the trajectory curvature change rate and the planning acceleration change rate. That is, under the premise of meeting safety, the lower the trajectory curvature change rate and acceleration change rate, the better, so that the trajectory with a high score is safer and more practical.
[0061] Specifically, the efficiency score also includes: Efficiency = Normalized (Total Trajectory Length).
[0062] Specifically, the efficiency of a trajectory can be measured by the total length of the planned trajectory. That is, when each trajectory planning is based on the vehicle's future (for example) state 10 seconds in advance, the longer the path, the higher the speed and the better the efficiency.
[0063] Preferably, the trajectory is scored by setting efficiency, so that the longer the path, the higher the speed and the better the efficiency, so that the trajectory obtained by efficiency scoring can travel better at the current intersection.
[0064] Specifically, the anthropomorphism score also includes: Anthropomorphism = Normalized (1 / Freche distance (the average trajectory of the human driver in the corresponding scene, the target trajectory)).
[0065] Specifically, the anthropomorphism metric measures the similarity between the current trajectory and the trajectory of a human driver. The closer it is to human standards, the more easily it will be accepted by users and the higher the score will be. However, this step first requires assuming that the average trajectory of a human driver in the corresponding scenario / condition is obtained based on big data.
[0066] Furthermore, the smaller the Frescher distance between two curves, the higher the similarity between the two curves.
[0067] Specifically, the aforementioned indicator scores, namely comfort score, anthropomorphism score, and efficiency score, have all been normalized to ensure that the final total score is not negatively affected by the difference in the order of magnitude between different parameters.
[0068] Preferably, the trajectory is scored by setting anthropomorphism, so that the trajectory can be more in line with human standards and the driving habits of existing drivers. This ensures the best route is taken under the current driving conditions, while making the trajectory more humanized and ensuring driving safety.
[0069] The present invention also provides a decision planning device 100, comprising: an acquisition module 110 for detecting driving parameters of a vehicle and surrounding environmental parameters to obtain multiple raw data; a processing module 120 for importing the multiple raw data into an AI model for processing to obtain multiple perception results; calculating a planning trajectory based on the multiple perception results to obtain a first planning trajectory result; the AI model obtaining a second planning trajectory result based on the multiple raw data; and a scoring module 130 for importing the first planning trajectory result and the second planning trajectory result into a trajectory scoring arbitration module to verify the optimal trajectory and output the optimal trajectory.
[0070] Preferably, the current driving environment of the vehicle is obtained by setting the acquisition module 110 to obtain raw data, so that the vehicle trajectory planning is more in line with the actual situation of the current environment and the planning is more practical. The processing module 120 processes the raw data to obtain the first planned trajectory result and the second planned trajectory result, so that the two trajectory results are more accurate and in line with the actual situation. At the same time, the evaluation and comparison of multiple trajectory results can make the final output trajectory safer and more practical. The scoring module 130 scores and verifies the two trajectories to make the output trajectory that is the safest, fastest and most user-friendly for returning home. Through the cooperation of the three modules, the algorithm is more efficient and can take into account different driving environments, making it more practical.
[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A decision-making and planning system for complex urban road conditions, characterized in that, include: The system detects the driving parameters of the vehicle and the surrounding environmental parameters to obtain multiple raw data. The multiple raw data are imported into an AI model for processing to obtain multiple perception results; A convolutional neural network is used to preprocess and extract local features from the multiple raw data to obtain multiple feature vectors; The first planned trajectory result is obtained by calculating the planned trajectory using the multiple perception results; The AI model obtains the second planned trajectory result based on the multiple raw data; The multiple feature vectors are merged through a concatenation operation to obtain a merged feature vector; An intermediate result is obtained by using a fully connected neural network structure and a Transformer network structure as intermediate layers to perform feature transformation and encoding on the merged feature vector. The intermediate results are input into the LSTM Decoder network to generate a series of trajectory sequences, and pre-generated decision planning trajectory points are generated. The decision planning trajectory points are processed by the trajectory model, and the output is the second planning trajectory result; The first and second planned trajectory results are imported into the trajectory scoring arbitration module to verify the optimal trajectory and output the optimal trajectory.
2. The decision-making and planning system for complex urban road conditions according to claim 1, characterized in that, The step of importing the first planned trajectory result and the second planned trajectory result into the trajectory scoring arbitration module to verify the optimal trajectory and output the optimal trajectory also includes: The trajectory is predicted by the prediction module to obtain the predicted trajectory, and the credibility of the predicted trajectory is evaluated to obtain the third planned trajectory result. The third planning trajectory result and the second planning trajectory result are subjected to collision filtering to obtain the fourth planning trajectory result.
3. The decision-making and planning system for complex urban road conditions according to claim 2, characterized in that, The first and fourth planning trajectory results are scored for comfort, anthropomorphism, and efficiency. The results of the comfort score, the anthropomorphism score, and the efficiency score will be used to output the trajectory with the highest score.
4. A decision-making and planning device, characterized in that, include: Acquisition module: detects the driving parameters of the vehicle and the surrounding environment parameters to obtain multiple raw data; Processing Module: The module imports the multiple raw data sets into an AI model for processing, obtaining multiple perception results; it preprocesses and extracts local features from the raw data sets using a convolutional neural network, obtaining multiple feature vectors; it calculates a planned trajectory using the multiple perception results, obtaining a first planned trajectory result; the AI model obtains a second planned trajectory result based on the multiple raw data sets; it merges the multiple feature vectors through a concatenation operation, obtaining a merged feature vector; it uses a fully connected neural network structure and a Transformer network structure as intermediate layers to perform feature transformation and encoding on the merged feature vector, obtaining an intermediate result; and it inputs the intermediate result into an LSTM Decoder network to generate a series of trajectory sequences, pre-generating decision planning trajectory points. The decision planning trajectory points are processed by the trajectory model, and the output is the second planning trajectory result; Scoring module: Import the first planned trajectory result and the second planned trajectory result into the trajectory scoring arbitration module to verify the best trajectory and output the best trajectory.