An intelligent scheduling control method and system of artificial intelligence

Abstract

The application discloses an intelligent scheduling control method of artificial intelligence, relates to the technical field of intelligent transportation, and can accurately acquire the current traffic state of a crossroad by analyzing real-time images through a first artificial intelligence model, thereby providing a solid data foundation for subsequent decision-making, and then can identify and determine the signal light execution phase of the current traffic state through a second artificial intelligence model, perform signal control according to the signal light execution phase, and construct the current traffic state, the signal light execution phase, a reward and a next traffic state into a replay experience, update the second artificial intelligence model by using the replay experience, and continuously accumulate new experience and optimize itself in actual operation, so that the scheduling strategy can adapt to long-term changes of a traffic mode and has strong self-adaptation and continuous evolution capabilities.

Description

Technical Field

[0001] This application relates to the field of intelligent transportation technology, and more specifically, to an artificial intelligence-based intelligent dispatching and control method and system. Background Technology

[0002] With the acceleration of urbanization and the surge in motor vehicle ownership, urban traffic congestion has become increasingly severe. As a key node in the urban transportation network, the efficiency of intersections directly impacts the overall operation of the road network. Traditional traffic light control mainly employs fixed-sequence schemes or simple sensor-based schemes, both of which have numerous drawbacks. Fixed-sequence schemes cannot dynamically adjust to real-time traffic flow changes, causing unnecessary waiting during off-peak hours and failing to effectively manage traffic during peak hours, leading to resource waste and exacerbated congestion. While sensor-based schemes can respond to traffic flow to some extent, their detection information is limited (usually only the presence or absence of vehicles), lacking precise perception of the overall intersection situation (such as queue length, waiting time, and vehicle type), resulting in simple decision-making logic and limited optimization effects. Summary of the Invention

[0003] This application aims to provide an intelligent scheduling and control method and system based on artificial intelligence, which is intended to solve the problem that existing technologies cannot dynamically adjust according to real-time traffic flow changes, causing unnecessary waiting during off-peak hours and failing to effectively manage traffic during peak hours, resulting in resource waste and increased congestion.

[0004] The first aspect of this application provides an intelligent scheduling and control method based on artificial intelligence, comprising:

[0005] Real-time images of the intersection are collected and processed by a first artificial intelligence model to obtain the current traffic status of the intersection.

[0006] The current traffic condition is identified using a second artificial intelligence model to determine the traffic light execution phase at the intersection.

[0007] Based on the traffic light phases at the intersection, perform intelligent traffic management at the intersection and collect the next traffic status after intelligent traffic management;

[0008] Based on the current traffic status of the intersection and the next traffic status, the reward for implementing the intelligent dispatch is determined, and a replay experience is constructed based on the current traffic status, traffic light execution phase, reward, and the next traffic status.

[0009] The second artificial intelligence model is updated based on the replay experience to obtain the updated second artificial intelligence model, and the updated second artificial intelligence model is used to enter the next round of intelligent scheduling.

[0010] The first artificial intelligence model is set as a combination of the YOLO model and the Transformer model, and the YOLO model is trained as follows:

[0011] The hyperparameters of the YOLO model are randomly initialized between the upper and lower bounds, and the randomly initialized hyperparameters are encoded into vectors to obtain the hyperparameter encoding.

[0012] Based on the hyperparameter encoding obtained through random initialization, multiple different hyperparameter encodings are obtained as follows:

[0013]

[0014] in, Let represent the d-th dimension hyperparameter encoded by the k-th hyperparameter, and when k=1, Let d represent the d-th dimension of the hyperparameter encoding obtained through random initialization, where d = 1, 2, ..., D, and D represents the total dimension of the hyperparameters in the hyperparameter encoding. Denotes the mapping factor, and Between [0,1] This represents the d-th dimension hyperparameter encoded by the (k+1)-th hyperparameter. Represents the modulo function;

[0015] Obtain the first loss function value for each hyperparameter encoding, and determine the hyperparameter encoding with the smallest first loss function value as the optimal encoding;

[0016] Based on the optimal encoding, the search position for the hyperparameter encoding is selected as follows:

[0017]

[0018] in, This represents the m-th hyperparameter encoding during the t-th training iteration, where m = 1, 2, ..., NP, and NP represents the total number of hyperparameter encodings. This represents the hyperparameter encoding after the m-th selected search position. This represents the optimal encoding. Represents the first random number between (0,1). This represents the second random number between (0,1). This represents a third random number between (0,1). This represents the fourth random number between (0,1). Let T represent an exponential function with the natural constant e as the base, and let T represent the maximum number of training iterations preset.

[0019] Learning neighborhood information for the hyperparameter encoding after selecting the search location is as follows:

[0020]

[0021]

[0022] in, This represents the d-th dimension hyperparameter encoded after the nth search position selection during the t-th training process. This represents the optimal encoding. This represents the d-th dimension hyperparameter encoded after learning the information from the nth neighborhood. This represents the neighborhood information learning coefficient, which decreases linearly from 1 to 0.0001 as the number of training iterations increases. Indicates the learning intensity of neighborhood information. This represents the maximum value of the d-th dimension hyperparameter. This represents the minimum value of the d-th dimension hyperparameter. This represents the d-th dimension of the hyperparameters of the g-th neighborhood code corresponding to the hyperparameter code after the n-th selection search position during the t-th training process, where G represents the total number of neighborhood codes. This represents the Euclidean distance between the hyperparameter encoding after the nth selected search position during the tth training process and its corresponding gth neighbor encoding; for any hyperparameter encoding, if its Euclidean distance with other hyperparameter encodings is less than the preset neighborhood radius, then the other hyperparameter encodings are neighborhood encodings. This represents the first learning intensity control parameter, and is set to 0.5; This represents the second learning intensity control parameter, and is set to 1.5; Represents the natural constant;

[0023] Information exchange is performed on the hyperparameter encoding after neighborhood information learning, including:

[0024] The hyperparameter codes learned from two neighborhood information are paired up and used as each other's paired codes to obtain the paired code corresponding to each hyperparameter code.

[0025] Information exchange is performed based on the paired encoding corresponding to the hyperparameter encoding:

[0026]

[0027]

[0028] in, This represents the hyperparameter encoding after learning the information of the q-th neighborhood. This represents the hyperparameter encoding after the q-th information exchange. Indicates inertia weight, express The corresponding pairing code, This represents the search constant and is set to 1 or 1.5. Let π represent the mathematical constant pi, and cos represent the cosine function. This represents the spiral direction control factor between (0,1). Represents the fifth random number between (0, 1);

[0029] The hyperparameter encoding after information exchange is subjected to a fast position transformation as follows:

[0030]

[0031] in, This represents the hyperparameter encoding after the p-th information exchange during the t-th training process. This represents the hyperparameter encoding after the p-th fast position transformation. Represents the gamma function. This represents the rapid position change control parameter, and is set to 2; Represents pi;

[0032] Determine whether the current number of training iterations is greater than or equal to the preset maximum number of training iterations. If so, redetermine the optimal encoding based on the hyperparameter encoding after the fast position transformation. Otherwise, return to the step of determining the hyperparameter encoding with the smallest first loss function value as the optimal encoding.

[0033] The hyperparameters in the optimal encoding are used as the final hyperparameters of the YOLO model to complete the training.

[0034] In one possible design approach, real-time images of the intersection are acquired, and the real-time images are processed using a first artificial intelligence model to obtain the current traffic state of the intersection, including:

[0035] Real-time images of the intersection are collected, and the first artificial intelligence model is used to identify the real-time images to determine the vehicle information in the entry lanes of each intersection.

[0036] Based on the vehicle information in the entering lanes, the current traffic status of the intersection is constructed; wherein, the current traffic status includes the current traffic light status of all entering lanes corresponding to the intersection, the queuing status of vehicles in the entering lanes, the cumulative waiting time of vehicles in the entering lanes, the presence status of emergency vehicles, the location of emergency vehicles, and the speed of emergency vehicles.

[0037] In one possible design approach, real-time images of the intersection are acquired, and a first artificial intelligence model is used to identify the real-time images to determine vehicle information in the entry lanes of each intersection, including:

[0038] Capture real-time images of the intersection;

[0039] The first artificial intelligence model identifies targets and target information in the real-time images to obtain vehicle information in the entry lanes of each intersection; wherein, the targets represent social vehicles or emergency vehicles, and the target information represents the license plate number of the targets.

[0040] In one possible design approach, the current traffic state of the intersection is constructed based on the vehicle information in the entering lane, including:

[0041] Based on the target and target information in the vehicle information of the entering lane, the number of vehicles in the entering lane is determined, and the queuing situation of the vehicles corresponding to the entering lane is obtained.

[0042] Based on the target and target information in the vehicle information of the entering lane, the waiting time of each vehicle in the entering lane is determined, and the waiting times of all vehicles are added together to obtain the cumulative waiting time of the vehicles in the entering lane.

[0043] Based on the target and target information in the vehicle information of the entering lane, the existence status of the emergency vehicle is determined; wherein, the existence status of the emergency vehicle is 1 or 0, an existence status of 1 indicates that the entering lane includes an emergency vehicle, and an existence status of 0 indicates that the entering lane does not include an emergency vehicle.

[0044] Based on the target information and target information in the vehicle information of the vehicle entering the lane, determine the number of other vehicles ahead of the emergency vehicle's location, and obtain the location of the emergency vehicle.

[0045] Based on the target and target information in the vehicle information on the lane, the distance between the center points of the detection frames of emergency vehicles at two adjacent sampling time points is determined, and the distance is converted into the speed of emergency vehicles.

[0046] Collect the current traffic light status of the lane you are entering; wherein, the traffic light status is 1 or 0, a traffic light status of 1 represents a green light, and a traffic light status of 0 represents a red light;

[0047] The current traffic state of the intersection is constructed based on the current traffic light status of all entering lanes, the queuing status of vehicles in the entering lanes, the cumulative waiting time of vehicles in the entering lanes, the presence status of emergency vehicles, the location of emergency vehicles, and the speed of emergency vehicles.

[0048] In one possible design approach, the current traffic state is identified using a second artificial intelligence model to determine the traffic light execution phase at the intersection, including:

[0049] The current traffic state is identified by a second artificial intelligence model, and an ε-greedy strategy is used to select actions to obtain the traffic light execution phase at the intersection; wherein, the second artificial intelligence model is set as the current network in the DQN algorithm.

[0050] In one possible design approach, intelligent traffic management of the intersection is performed based on the traffic light phases, and the next traffic state after intelligent management is collected, including:

[0051] The traffic lights at the intersection are adjusted according to their phases to achieve intelligent traffic management, and the traffic status after the intelligent traffic management is implemented is collected.

[0052] In one possible design approach, the reward for implementing the intelligent dispatch is determined based on the current traffic state of the intersection and the next traffic state, including:

[0053] Based on the current traffic status and the next traffic status at the intersection, the reward for implementing the intelligent dispatch is determined as follows:

[0054]

[0055]

[0056]

[0057] in, Indicates a reward. This indicates an award for emergency vehicles. This parameter indicates the priority of rewards for emergency vehicles. This indicates a reward for private vehicles. This parameter indicates the priority of rewards for social vehicles. Indicates the time weight of emergency vehicle passage. Indicates the weight of emergency vehicle waiting time. Indicates the speed weight of emergency vehicles passing through. This indicates the time loss weight for emergency vehicles, where n represents the total number of emergency vehicles passing through the intersection. This represents the time it takes for the i-th emergency vehicle to cross the intersection. This represents the total waiting time of the i-th emergency vehicle before the intersection. This represents the average speed of the i-th emergency vehicle as it passes through the intersection. This represents the time loss caused by the i-th emergency vehicle traveling below the expected speed, where the expected speed is a pre-set speed threshold. Indicates the weight of time deviation for social vehicles. Indicates the weight of waiting time for social vehicles. The weight representing the average queue length of social vehicles. This represents the speed weight of social vehicles, and m represents the total number of entering lanes at the intersection. This represents a preset constant. This represents the deviation between the average waiting time of the j-th vehicle entering the lane and the average waiting time of all vehicles. This indicates the average waiting time for all vehicles in the traffic lane corresponding to the phase in which the traffic light is in effect. This indicates the average queue length of all vehicles in the traffic lane corresponding to the phase in which the traffic light is in effect. This indicates the average speed of all vehicles in the traffic lane corresponding to the phase in which the traffic light is in effect.

[0058] In one possible design approach, the second artificial intelligence model is updated based on the playback experience to obtain an updated second artificial intelligence model, including:

[0059] H replay experiences are randomly selected to obtain H target replay experiences;

[0060] Based on the target network in the DQN algorithm and the experience of replaying H targets, the value of the second loss function is obtained;

[0061] The second artificial intelligence model is updated based on the second loss function value to obtain the updated second artificial intelligence model.

[0062] The second aspect of this application provides an artificial intelligence-based intelligent scheduling and control system, including: a data acquisition module, a data recognition module, an intelligent scheduling and control module, a playback experience acquisition module, and a model update module;

[0063] The data acquisition module is used to acquire real-time images of the intersection and process the real-time images through a first artificial intelligence model to obtain the current traffic status of the intersection.

[0064] The data recognition module is used to identify the current traffic state through a second artificial intelligence model and determine the traffic light execution phase at the intersection.

[0065] The intelligent dispatch control module is used to perform intelligent dispatch of the intersection according to the traffic light phase of the intersection, and to collect the next traffic status after intelligent dispatch.

[0066] The replay experience acquisition module is used to determine the reward for implementing the intelligent dispatch based on the current traffic state and the next traffic state of the intersection, and to construct replay experience based on the current traffic state, traffic light execution phase, reward and the next traffic state.

[0067] The model update module is used to update the second artificial intelligence model based on the replay experience to obtain the updated second artificial intelligence model, and to use the updated second artificial intelligence model to enter the next round of intelligent scheduling.

[0068] Beneficial effects:

[0069] This application provides an intelligent dispatching and control method based on artificial intelligence. A first artificial intelligence model analyzes real-time images to accurately obtain the current traffic state of an intersection, providing a solid data foundation for subsequent decision-making. Then, a second artificial intelligence model identifies the current traffic state and determines the traffic light execution phase. Signal control is then performed based on the traffic light execution phase. The current traffic state, traffic light execution phase, reward, and the next traffic state are constructed as playback experience. This playback experience is used to update the second artificial intelligence model. In actual operation, new experience is continuously accumulated and self-optimized, enabling the dispatching strategy to adapt to long-term changes in traffic patterns and possessing strong adaptive and continuous evolution capabilities. Attached Figure Description

[0070] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is a flowchart illustrating an intelligent scheduling and control method for artificial intelligence proposed in an embodiment of this application;

[0072] Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based intelligent scheduling and control system proposed in one embodiment of this application;

[0073] Explanation of reference numerals in the attached diagram: 201-Data acquisition module, 202-Data recognition module, 203-Intelligent scheduling and control module, 204-Replay experience acquisition module, 205-Model update module. Detailed Implementation

[0074] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0075] like Figure 2 As shown in the figure, this application provides an intelligent scheduling and control method based on artificial intelligence, including:

[0076] S101. Collect real-time images of the intersection and process the real-time images using a first artificial intelligence model to obtain the current traffic status of the intersection.

[0077] Real-time images at an intersection refer to the images corresponding to each entering lane. Generally, an intersection has four entering directions, and each entering direction usually includes a straight-in lane, a right-turn lane, and a left-turn lane. By processing the real-time images using a first artificial intelligence model, vehicle information in each entering direction can be determined, thereby determining the current traffic status of the intersection.

[0078] For example, assuming the four directions of entry at an intersection are east, south, west, and north, a real-time image containing all entry lanes can be captured for the east-bound lanes. After identifying this real-time image, the vehicle information in the east-bound direction can be determined. Since this embodiment of the application requires obtaining the current traffic status of the intersection, real-time images of each direction need to be captured for identification.

[0079] Optionally, in order to accurately identify the current traffic status of the intersection, real-time images of each entering lane can be collected for identification.

[0080] S102. The current traffic state is identified by the second artificial intelligence model to determine the traffic light execution phase of the intersection.

[0081] The second artificial intelligence model can be set as a model in a deep reinforcement learning algorithm, and the action space of the deep reinforcement learning algorithm can be the execution phases of various traffic lights at the intersection. After the second artificial intelligence model identifies the current traffic state, it will select an action to obtain the execution phase of the traffic lights at the intersection.

[0082] The traffic light phases at an intersection can include four phases: north-south straight, north-south left turn, east-west straight, and east-west left turn. Considering that in practical applications, intersections are generally not directly aligned with the cardinal directions, the center of the intersection can be taken as the origin, and the road with the direction closest to the north can be taken as the north, while the roads in the other three directions can be taken as the east, south, and west, respectively.

[0083] S103. Based on the traffic light phase of the intersection, perform intelligent traffic management of the intersection and collect the next traffic status after intelligent traffic management.

[0084] After executing the traffic light phases at an intersection, a smart traffic management cycle for that intersection is completed. Then, based on a preset scheduling cycle, the traffic status for the next scheduling time is collected, providing a basis for scheduling optimization.

[0085] S104. Based on the current traffic status of the intersection and the next traffic status, determine the reward for implementing the intelligent dispatch, and construct a replay experience based on the current traffic status, traffic light execution phase, reward, and the next traffic status.

[0086] Playing back experiences can update the second artificial intelligence model, making subsequent decisions more accurate and thus gradually improving the efficiency of command and dispatch.

[0087] S105. The second artificial intelligence model is updated based on the replay experience to obtain the updated second artificial intelligence model, and the updated second artificial intelligence model is used to enter the next round of intelligent scheduling.

[0088] This application provides an intelligent dispatching and control method based on artificial intelligence. A first artificial intelligence model analyzes real-time images to accurately obtain the current traffic state of an intersection, providing a solid data foundation for subsequent decision-making. Then, a second artificial intelligence model identifies the current traffic state and determines the traffic light execution phase. Signal control is then performed based on the traffic light execution phase. The current traffic state, traffic light execution phase, reward, and the next traffic state are constructed as playback experience. This playback experience is used to update the second artificial intelligence model. In actual operation, new experience is continuously accumulated and self-optimized, enabling the dispatching strategy to adapt to long-term changes in traffic patterns and possessing strong adaptive and continuous evolution capabilities.

[0089] In one possible design approach, real-time images of the intersection are acquired, and the real-time images are processed using a first artificial intelligence model to obtain the current traffic state of the intersection, including:

[0090] Real-time images of the intersection are collected, and a first artificial intelligence model is used to identify the real-time images to determine the vehicle information in the entry lanes of each intersection.

[0091] Based on the vehicle information in the entering lanes, the current traffic state of the intersection is constructed. This current traffic state includes the current traffic light status for all entering lanes, the queuing status of vehicles in the entering lanes, the cumulative waiting time of vehicles in the entering lanes, the presence status of emergency vehicles, the location of emergency vehicles, and the speed of emergency vehicles.

[0092] In this embodiment, the first artificial intelligence model can be set as a combination of the YOLO model and the Transformer model (a neural network model based on an encoder-decoder structure). The YOLO model can be used to identify real-time images, determine each vehicle and its type label, and then crop and segment the license plate area before feeding it into the Transformer model for recognition. This will identify the vehicle's license plate number and thus obtain the vehicle information.

[0093] Both the YOLO and Transformer models should be trained before use to ensure that they can accurately perform image recognition tasks.

[0094] For example, the YOLO model can be trained as follows:

[0095] The hyperparameters of the YOLO model are randomly initialized between the upper and lower bounds, and the randomly initialized hyperparameters are encoded into vectors to obtain the hyperparameter encoding.

[0096] The hyperparameters of the YOLO model can be the connection weights between its network layers.

[0097] Based on the hyperparameter encoding obtained through random initialization, multiple different hyperparameter encodings are obtained as follows:

[0098]

[0099] in, Let represent the d-th dimension hyperparameter encoded by the k-th hyperparameter, and when k=1, Let d represent the d-th dimension of the hyperparameter encoding obtained through random initialization, where d = 1, 2, ..., D, and D represents the total dimension of the hyperparameters in the hyperparameter encoding. Denotes the mapping factor, and Between [0,1] This represents the d-th dimension hyperparameter encoded by the (k+1)-th hyperparameter. This represents the remainder function.

[0100] After initialization through the above process, the initial position of the population composed of hyperparameter encoding has certain advantages compared to random allocation. It is more evenly distributed in the solution space, reducing the adverse effects of random initialization on the population's search for the optimal solution and improving the algorithm's solution performance to a certain extent.

[0101] Obtain the first loss function value for each hyperparameter encoding, and determine the hyperparameter encoding with the smallest first loss function value as the optimal encoding.

[0102] Based on the optimal encoding, the search position for the hyperparameter encoding is selected as follows:

[0103]

[0104] in, This represents the m-th hyperparameter encoding during the t-th training iteration, where m = 1, 2, ..., NP, and NP represents the total number of hyperparameter encodings. This represents the hyperparameter encoding after the m-th selected search position. This represents the optimal encoding. Indicates the first random code. This represents the mean encoding of one-third of the hyperparameter codes randomly, meaning that the hyperparameters in any dimension of the mean encoding are the mean of the hyperparameter codes in the same dimension. Represents the first random number between (0,1). This represents the second random number between (0,1). This represents a third random number between (0,1). This represents the fourth random number between (0,1). Let T represent an exponential function with the natural constant e as the base, and let T represent the maximum number of training iterations preset.

[0105] Choosing the search position for hyperparameter encoding allows each hyperparameter encoding to combine information from other hyperparameter encodings to select its own search position. This reduces collisions in search paths between different hyperparameter encodings, allows searching more regions, and improves the ability to find the global optimum.

[0106] Learning neighborhood information for the hyperparameter encoding after selecting the search location is as follows:

[0107]

[0108]

[0109] in, This represents the d-th dimension hyperparameter encoded after the nth search position selection during the t-th training process. This represents the optimal encoding. This represents the d-th dimension hyperparameter encoded after learning the information from the nth neighborhood. This represents the neighborhood information learning coefficient, which decreases linearly from 1 to 0.0001 as the number of training iterations increases. Indicates the learning intensity of neighborhood information. This represents the maximum value of the d-th dimension hyperparameter. This represents the minimum value of the d-th dimension hyperparameter. This represents the d-th dimension of the hyperparameters of the g-th neighborhood code corresponding to the hyperparameter code after the n-th selection search position during the t-th training process, where G represents the total number of neighborhood codes. This represents the Euclidean distance between the hyperparameter encoding after the nth selected search position during the tth training process and its corresponding g-th neighbor encoding. For any hyperparameter encoding, if its Euclidean distance to other hyperparameter encodings is less than a preset neighborhood radius, then the other hyperparameter encodings are considered neighbor encodings. This represents the first learning intensity control parameter, and it is set to 0.5. This represents the second learning intensity control parameter, and it is set to 1.5. Represents the natural constant.

[0110] This neighborhood information learning can effectively integrate information from other hyperparameter codes around the hyperparameter code, improve the solution space search rate, thereby increasing the algorithm's ability to find the global optimum. Furthermore, as the algorithm progresses, the search accuracy will become higher and higher, ensuring the convergence of the algorithm.

[0111] Information exchange is performed on the hyperparameter encoding after neighborhood information learning, including:

[0112] The hyperparameter codes learned from two neighborhood information are paired up and used as each other's paired codes to obtain the paired code corresponding to each hyperparameter code.

[0113] Optionally, to ensure the smooth progress of the training process, NP can be set to an integer multiple of 6.

[0114] Information exchange is performed based on the paired encoding corresponding to the hyperparameter encoding:

[0115]

[0116]

[0117] in, This represents the hyperparameter encoding after learning the information of the q-th neighborhood. This represents the hyperparameter encoding after the q-th information exchange. Indicates inertia weight, express The corresponding pairing code, This represents the search constant and can be set to 1 or 1.5. Let π represent the mathematical constant pi, and cos represent the cosine function. This represents the spiral direction control factor between (0,1). This represents the fifth random number between (0,1), and can also be... Replace with a fixed constant, such as 0.6.

[0118] This information exchange can effectively enable two hyperparameter codes to perform collaborative search, improving the search capability for the region between the two hyperparameter codes. In the early stages, the search step size is larger, which is more conducive to finding the global optimum. In the later stages of the algorithm, local optimization is gradually strengthened to improve the training accuracy of the algorithm.

[0119] The hyperparameter encoding after information exchange is subjected to a fast position transformation as follows:

[0120]

[0121] in, This represents the hyperparameter encoding after the p-th information exchange during the t-th training process. This represents the hyperparameter encoding after the p-th fast position transformation. Represents the gamma function. This represents the rapid position change control parameter, and it is set to 2. It represents pi (π).

[0122] Optionally, after the hyperparameter encoding is changed, out-of-bounds processing can be performed to ensure that the hyperparameter encoding always stays within its upper and lower limits. For example, out-of-bounds hyperparameters can be randomly regenerated within their corresponding upper and lower limits.

[0123] This fast position transformation provides strong position transformation capability in the early stages of the algorithm, helping it escape local optima. Furthermore, it gradually transforms the position within its neighborhood in the later stages, ensuring the algorithm's convergence accuracy. To improve training speed, simulated annealing can be used to probabilistically accept inferior solutions after the fast position transformation.

[0124] Determine if the current number of training iterations is greater than or equal to the preset maximum number of training iterations. If so, redetermine the optimal encoding based on the hyperparameter encoding after the fast position transformation. Otherwise, return to the step of determining the hyperparameter encoding with the smallest first loss function value as the optimal encoding.

[0125] The hyperparameters in the optimal encoding are used as the final hyperparameters of the YOLO model to complete the training.

[0126] This training algorithm not only possesses powerful global search capabilities, enabling it to escape local optima and search for the global optimum in the high-dimensional solution space, but also improves the training speed and ensures convergence in later stages. This allows the trained YOLO model to accurately identify targets, improve vehicle recognition accuracy, and thus guarantee the accuracy of scheduling and control.

[0127] In one possible design approach, real-time images of the intersection are acquired, and a first artificial intelligence model is used to identify the real-time images to determine vehicle information in the entry lanes of each intersection, including:

[0128] Capture real-time images of the intersection.

[0129] The first artificial intelligence model identifies targets and target information in the real-time images to obtain vehicle information in the entry lanes of each intersection. The targets represent civilian vehicles or emergency vehicles, and the target information represents the license plate numbers of the targets.

[0130] This application embodiment, by identifying social vehicles or emergency vehicles, can prioritize the passage of emergency vehicles during command and dispatch, thus ensuring the driving efficiency of emergency vehicles. Emergency vehicles can refer to vehicles such as ambulances, fire trucks, and police cars.

[0131] In one possible design approach, the current traffic state of the intersection is constructed based on the vehicle information in the entering lane, including:

[0132] Based on the target and target information in the vehicle information of the entering lane, the number of vehicles in the entering lane is determined, and the queuing situation of the vehicles corresponding to the entering lane is obtained.

[0133] Based on the target and target information in the vehicle information of the entering lane, the waiting time of each vehicle in the entering lane is determined, and the waiting times of all vehicles are added together to obtain the cumulative waiting time of the vehicles in the entering lane.

[0134] The presence status of emergency vehicles is determined based on the target information and target details in the vehicle information of the vehicle entering the lane. The presence status of emergency vehicles is either 1 or 0. A presence status of 1 indicates that the lane includes an emergency vehicle, while a presence status of 0 indicates that the lane does not include an emergency vehicle.

[0135] Based on the target information and target information in the vehicle information of the vehicle entering the lane, the number of other vehicles ahead of the emergency vehicle's location is determined, and the location of the emergency vehicle is obtained.

[0136] Based on the target and target information in the vehicle information on the lane, the distance between the center points of the detection frames of emergency vehicles at two adjacent sampling time points is determined, and the distance is converted into the speed of emergency vehicles.

[0137] For example, the relationship between the distance between two pixels and the actual distance can be established. Then, before the next scheduling time, a preset data sampling frequency can be used for data collection and analysis. This allows the determination of the distance between the center points of the detection frames of the emergency vehicle at two adjacent sampling time points. After converting the distance between the center points of the detection frames of the emergency vehicle at two adjacent sampling time points into the actual distance, and then dividing it by the time difference between the sampling time points, the speed of the emergency vehicle between the two sampling time points can be determined. It is worth noting that the above method is only an example. The speed can also be obtained by identifying the actual location of the emergency vehicle and using the time difference between different actual locations and the distance between different actual locations. Speed ​​radar can also be used for measurement.

[0138] Optionally, a scheduling cycle can be divided into an integer number of sampling time points by the data sampling frequency. That is, any scheduling moment coincides with a sampling time point. This not only facilitates the acquisition of instantaneous data (e.g., the emergency vehicle speed between the sampling time point corresponding to the scheduling moment and the previous sampling time point can be used as the emergency vehicle speed corresponding to the scheduling moment), but also allows the acquisition of data between two scheduling moments (e.g., the average speed can be obtained by averaging all emergency vehicle speeds between two scheduling moments).

[0139] The system collects the current traffic light status corresponding to the lane the vehicle is entering. The traffic light status is either 1 or 0, with 1 representing a green light and 0 representing a red light.

[0140] The current traffic state of the intersection is constructed based on the current traffic light status of all entering lanes, the queuing status of vehicles in the entering lanes, the cumulative waiting time of vehicles in the entering lanes, the presence status of emergency vehicles, the location of emergency vehicles, and the speed of emergency vehicles.

[0141] The data corresponding to each entering lane can be constructed into a row of data, and then the data of all entering lanes can be constructed into a matrix to obtain the current traffic status of the intersection.

[0142] In one possible design approach, the current traffic state is identified using a second artificial intelligence model to determine the traffic light execution phase at the intersection, including:

[0143] The current traffic state is identified using a second artificial intelligence model, and an ε-greedy strategy is employed to select actions, thereby obtaining the traffic light execution phase at the intersection. The second artificial intelligence model is configured as the current network within the DQN (Deep Q-Network) algorithm.

[0144] A second artificial intelligence model based on a deep Q-network is adopted as the decision-making core. Through reinforcement learning, the system can autonomously learn the optimal traffic light control strategy through continuous interaction with the environment. This strategy is no longer a simple rule, but can make globally optimal or near-optimal decisions based on complex and ever-changing real-time traffic flow.

[0145] In one possible design approach, intelligent traffic management of the intersection is performed based on the traffic light phases, and the next traffic state after intelligent management is collected, including:

[0146] The traffic lights at the intersection are adjusted according to their phases to achieve intelligent traffic management, and the traffic status after the intelligent traffic management is implemented is collected.

[0147] In one possible design approach, the reward for implementing the intelligent dispatch is determined based on the current traffic state of the intersection and the next traffic state, including:

[0148] Based on the current traffic status and the next traffic status at the intersection, the reward for implementing the intelligent dispatch is determined as follows:

[0149]

[0150]

[0151]

[0152] in, Indicates a reward. This indicates an award for emergency vehicles. This parameter indicates the priority of rewards for emergency vehicles. This indicates a reward for private vehicles. This parameter indicates the priority of rewards for social vehicles. Indicates the time weight of emergency vehicle passage. Indicates the weight of emergency vehicle waiting time. Indicates the speed weight of emergency vehicles passing through. This indicates the time loss weight for emergency vehicles, where n represents the total number of emergency vehicles passing through the intersection. This represents the time it takes for the i-th emergency vehicle to cross the intersection. This represents the total waiting time of the i-th emergency vehicle before the intersection. This represents the average speed of the i-th emergency vehicle as it passes through the intersection. This represents the time loss caused by the i-th emergency vehicle traveling below the expected speed, where the expected speed is a pre-set speed threshold. Indicates the weight of time deviation for social vehicles. Indicates the weight of waiting time for social vehicles. The weight representing the average queue length of social vehicles. This represents the speed weight of social vehicles, and m represents the total number of entering lanes at the intersection. This represents a preset constant. This represents the deviation between the average waiting time of the j-th vehicle entering the lane and the average waiting time of all vehicles. This indicates the average waiting time for all vehicles in the traffic lane corresponding to the phase in which the traffic light is in effect. This indicates the average queue length of all vehicles in the traffic lane corresponding to the phase in which the traffic light is in effect. This indicates the average speed of all vehicles in the traffic lane corresponding to the phase in which the traffic light is in effect.

[0153] This application designs a composite reward function. By setting priority parameters and different reward sub-items for emergency vehicles and social vehicles, it achieves the goal of giving emergency vehicles extremely high passage priority while ensuring the overall traffic efficiency of social vehicles.

[0154] In one possible design approach, the second artificial intelligence model is updated based on the playback experience to obtain an updated second artificial intelligence model, including:

[0155] H replay experiences are randomly selected to obtain H target replay experiences.

[0156] Based on the target network in the DQN algorithm and the experience of replaying H targets, the value of the second loss function is obtained.

[0157] The second artificial intelligence model is updated based on the second loss function value to obtain the updated second artificial intelligence model.

[0158] By employing a closed-loop mechanism of execution-feedback-learning-update and utilizing key DQN technologies such as experience replay and target networks, the stability and convergence of model training are ensured. In actual operation, it will continuously accumulate new experience and self-optimize, enabling the scheduling strategy to adapt to long-term changes in traffic patterns and possessing strong adaptive and continuous evolution capabilities.

[0159] like Figure 2 As shown, based on the same inventive concept, this application provides an artificial intelligence-based intelligent scheduling and control system, including: a data acquisition module 201, a data recognition module 202, an intelligent scheduling and control module 203, a playback experience acquisition module 204, and a model update module 205.

[0160] The data acquisition module 201 is used to acquire real-time images of the intersection and process the real-time images through a first artificial intelligence model to obtain the current traffic status of the intersection.

[0161] The data recognition module 202 is used to identify the current traffic state through a second artificial intelligence model and determine the traffic light execution phase at the intersection.

[0162] The intelligent dispatch control module 203 is used to perform intelligent dispatch of the intersection according to the traffic light phase of the intersection, and to collect the next traffic status after intelligent dispatch.

[0163] The replay experience acquisition module 204 is used to determine the reward for implementing the intelligent dispatch based on the current traffic state of the intersection and the next traffic state, and to construct replay experience based on the current traffic state, traffic light execution phase, reward and the next traffic state.

[0164] The model update module 205 is used to update the second artificial intelligence model based on the playback experience to obtain the updated second artificial intelligence model, and to use the updated second artificial intelligence model to enter the next round of intelligent scheduling.

[0165] The intelligent scheduling and control system based on artificial intelligence provided in this application embodiment can execute the above-described method and technical solution. Its principle and beneficial effects are similar, and will not be repeated here.

[0166] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0167] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices, and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0168] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0169] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0170] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0171] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0172] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A smart scheduling and control method based on artificial intelligence, characterized in that, include: Real-time images of the intersection are collected and processed by a first artificial intelligence model to obtain the current traffic status of the intersection. The current traffic condition is identified using a second artificial intelligence model to determine the traffic light execution phase at the intersection. Based on the traffic light phases at the intersection, perform intelligent traffic management at the intersection and collect the next traffic status after intelligent traffic management; Based on the current traffic status of the intersection and the next traffic status, the reward for implementing the intelligent dispatch is determined, and a replay experience is constructed based on the current traffic status, traffic light execution phase, reward, and the next traffic status. The second artificial intelligence model is updated based on the replay experience to obtain the updated second artificial intelligence model, and the updated second artificial intelligence model is used to enter the next round of intelligent scheduling. The first artificial intelligence model is set as a combination of the YOLO model and the Transformer model, and the YOLO model is trained as follows: The hyperparameters of the YOLO model are randomly initialized between the upper and lower bounds, and the randomly initialized hyperparameters are encoded into vectors to obtain the hyperparameter encoding. Based on the hyperparameter encoding obtained through random initialization, multiple different hyperparameter encodings are obtained as follows: in, Indicates the first k The first hyperparameter encoding d dimensional hyperparameters, and k= At 1 o'clock, The first hyperparameter encoding obtained from random initialization represents the... d dimensional hyperparameters, d= 1,2,...,D, where D represents the total dimension of hyperparameters in hyperparameter encoding. Denotes the mapping factor, and Between [0,1] Indicates the first k+ The first hyperparameter encoding d dimensional hyperparameters, Represents the modulo function; Obtain the first loss function value for each hyperparameter encoding, and determine the hyperparameter encoding with the smallest first loss function value as the optimal encoding; Based on the optimal encoding, the search position for the hyperparameter encoding is selected as follows: in, Indicates the first t During the training process, the first m Hyperparameter encoding, m =1,2,...,NP, where NP represents the total number of hyperparameter codes. Indicates the first m Hyperparameter encoding after selecting a search location This represents the optimal encoding. Indicates the first random code. This represents the mean encoding of one-third of the hyperparameter codes randomly, meaning that the hyperparameters in any dimension of the mean encoding are the mean of the hyperparameter codes in the same dimension. Represents the first random number between (0,1). This represents the second random number between (0,1). This represents a third random number between (0,1). This represents the fourth random number between (0,1). Let T represent an exponential function with the natural constant e as the base, and let T represent the maximum number of training iterations preset. Learning neighborhood information for the hyperparameter encoding after selecting the search location is as follows: in, Indicates the first t During the training process, the first n The first hyperparameter encoding after selecting the search position. d dimensional hyperparameters, This represents the optimal encoding. Indicates the first n The hyperparameter encoding after learning the neighborhood information is the first... d dimensional hyperparameters, This represents the neighborhood information learning coefficient, which decreases linearly from 1 to 0.0001 as the number of training iterations increases. Indicates the learning intensity of neighborhood information. Indicates the first d The maximum value of the dimensional hyperparameter, Indicates the first d The minimum value of the hyperparameter. Indicates the first t During the training process, the first n The hyperparameter encoding corresponding to the selected search position is the first one. g The th neighborhood encoding of the th d The hyperparameter G represents the total number of neighborhood codes. Indicates the first t During the training process, the first n The hyperparameter encoding after selecting the search position and its corresponding first... g The Euclidean distance between neighboring codes; for any hyperparameter code, if its Euclidean distance with other hyperparameter codes is less than a preset neighborhood radius, then the other hyperparameter codes are neighboring codes; This represents the first learning intensity control parameter, and is set to 0.5; This represents the second learning intensity control parameter, and is set to 1.5; Represents the natural constant; Information exchange is performed on the hyperparameter encoding after neighborhood information learning, including: The hyperparameter codes learned from two neighborhood information are paired up and used as each other's paired codes to obtain the paired code corresponding to each hyperparameter code. Information exchange is performed based on the paired encoding corresponding to the hyperparameter encoding: in, Indicates the first q Hyperparameter encoding after learning neighborhood information Indicates the first q Hyperparameter encoding after information exchange Indicates inertia weight, express The corresponding pairing code, This represents the search constant and is set to 1 or 1.

5. Let π represent the mathematical constant pi, and cos represent the cosine function. This represents the spiral direction control factor between (0,1). Represents the fifth random number between (0, 1); The hyperparameter encoding after information exchange is subjected to a fast position transformation as follows: in, Indicates the first t During the training process, the first p Hyperparameter encoding after information exchange Indicates the first p Hyperparameter encoding after a fast position transformation Represents the gamma function. This represents the rapid position change control parameter, and is set to 2; Represents pi; Determine whether the current number of training iterations is greater than or equal to the preset maximum number of training iterations. If so, redetermine the optimal encoding based on the hyperparameter encoding after the fast position transformation. Otherwise, return to the step of determining the hyperparameter encoding with the smallest first loss function value as the optimal encoding. The hyperparameters in the optimal encoding are used as the final hyperparameters of the YOLO model to complete the training.

2. The intelligent scheduling and control method for artificial intelligence according to claim 1, characterized in that, Real-time images of the intersection are acquired, and the images are processed using a first artificial intelligence model to obtain the current traffic status of the intersection, including: Real-time images of the intersection are collected, and the first artificial intelligence model is used to identify the real-time images to determine the vehicle information in the entry lanes of each intersection. Based on the vehicle information in the entering lanes, the current traffic status of the intersection is constructed; wherein, the current traffic status includes the current traffic light status of all entering lanes corresponding to the intersection, the queuing status of vehicles in the entering lanes, the cumulative waiting time of vehicles in the entering lanes, the presence status of emergency vehicles, the location of emergency vehicles, and the speed of emergency vehicles.

3. The intelligent scheduling and control method for artificial intelligence according to claim 2, characterized in that, Real-time images of the intersection are acquired, and a first artificial intelligence model is used to identify the real-time images to determine vehicle information in the entry lanes of each intersection, including: Capture real-time images of the intersection; The first artificial intelligence model identifies targets and target information in the real-time images to obtain vehicle information in the entry lanes of each intersection; wherein, the targets represent social vehicles or emergency vehicles, and the target information represents the license plate number of the targets.

4. The intelligent scheduling and control method for artificial intelligence according to claim 3, characterized in that, Based on the vehicle information in the entering lane, the current traffic state of the intersection is constructed, including: Based on the target and target information in the vehicle information of the entering lane, the number of vehicles in the entering lane is determined, and the queuing situation of the vehicles corresponding to the entering lane is obtained. Based on the target and target information in the vehicle information of the entering lane, the waiting time of each vehicle in the entering lane is determined, and the waiting times of all vehicles are added together to obtain the cumulative waiting time of the vehicles in the entering lane. Based on the target and target information in the vehicle information of the entering lane, the existence status of the emergency vehicle is determined; wherein, the existence status of the emergency vehicle is 1 or 0, an existence status of 1 indicates that the entering lane includes an emergency vehicle, and an existence status of 0 indicates that the entering lane does not include an emergency vehicle. Based on the target information and target information in the vehicle information of the vehicle entering the lane, determine the number of other vehicles ahead of the emergency vehicle's location, and obtain the location of the emergency vehicle. Based on the target and target information in the vehicle information on the lane, the distance between the center points of the detection frames of emergency vehicles at two adjacent sampling time points is determined, and the distance is converted into the speed of emergency vehicles. Collect the current traffic light status corresponding to the lane you are entering; wherein, the traffic light status is 1 or 0, a traffic light status of 1 represents a green light, and a traffic light status of 0 represents a red light; The current traffic state of the intersection is constructed based on the current traffic light status of all entering lanes, the queuing status of vehicles in the entering lanes, the cumulative waiting time of vehicles in the entering lanes, the presence status of emergency vehicles, the location of emergency vehicles, and the speed of emergency vehicles.

5. The intelligent scheduling and control method for artificial intelligence according to claim 1, characterized in that, The current traffic state is identified using a second artificial intelligence model to determine the traffic light execution phase at the intersection, including: The current traffic state is identified by a second artificial intelligence model, and an ε-greedy strategy is used to select actions to obtain the traffic light execution phase at the intersection; wherein, the second artificial intelligence model is set as the current network in the DQN algorithm.

6. The intelligent scheduling and control method for artificial intelligence according to claim 5, characterized in that, Based on the traffic light phases at the intersection, the system performs intelligent traffic management at the intersection and collects the next traffic status after intelligent traffic management, including: The traffic lights at the intersection are adjusted according to their phases to achieve intelligent traffic management, and the traffic status after the intelligent traffic management is implemented is collected.

7. The intelligent scheduling and control method for artificial intelligence according to claim 6, characterized in that, Based on the current traffic state and the next traffic state at the intersection, the reward for implementing the intelligent dispatch is determined, including: Based on the current traffic status and the next traffic status at the intersection, the reward for implementing the intelligent dispatch is determined as follows: in, Indicates a reward. This indicates an award for emergency vehicles. This parameter indicates the priority of rewards for emergency vehicles. This indicates a reward for private vehicles. This parameter indicates the priority of rewards for social vehicles. Indicates the time weight of emergency vehicle passage. Indicates the weight of emergency vehicle waiting time. Indicates the speed weight of emergency vehicles passing through. This indicates the time loss weight for emergency vehicles, where n represents the total number of emergency vehicles passing through the intersection. Indicates the first i The time it takes for an emergency vehicle to pass through an intersection. Indicates the first i The total waiting time of emergency vehicles before the intersection Indicates the first i The average speed of an emergency vehicle passing through an intersection Indicates the first i The time loss caused by an emergency vehicle traveling at a speed lower than the expected speed, where the expected speed is a pre-set speed threshold. Indicates the weight of time deviation for social vehicles. Indicates the weight of waiting time for social vehicles. The weight representing the average queue length of social vehicles. Indicates the speed weight of social vehicles. m This indicates the total number of lanes for entering at the intersection. This represents a preset constant. Indicates the first j The average waiting time for each vehicle entering the lane deviates from the average waiting time for all vehicles. This indicates the average waiting time for all vehicles in the traffic lane corresponding to the phase in which the traffic light is in effect. This indicates the average queue length of all vehicles in the traffic lane corresponding to the phase in which the traffic light is in effect. This indicates the average speed of all vehicles in the traffic lane corresponding to the phase in which the traffic light is in effect.

8. The intelligent scheduling and control method for artificial intelligence according to claim 7, characterized in that, The second artificial intelligence model is updated based on the replay experience to obtain the updated second artificial intelligence model, including: H replay experiences are randomly selected to obtain H target replay experiences; Based on the target network in the DQN algorithm and the experience of replaying H targets, the value of the second loss function is obtained; The second artificial intelligence model is updated based on the second loss function value to obtain the updated second artificial intelligence model.

9. An artificial intelligence-based intelligent scheduling and control system, which is capable of executing the artificial intelligence-based intelligent scheduling and control method according to any one of claims 1 to 8, characterized in that, include: The system includes a data acquisition module, a data recognition module, an intelligent scheduling and control module, a playback experience acquisition module, and a model update module. The data acquisition module is used to acquire real-time images of the intersection and process the real-time images through a first artificial intelligence model to obtain the current traffic status of the intersection. The data recognition module is used to identify the current traffic state through a second artificial intelligence model and determine the traffic light execution phase at the intersection. The intelligent dispatch control module is used to perform intelligent dispatch of the intersection according to the traffic light phase of the intersection, and to collect the next traffic status after intelligent dispatch. The replay experience acquisition module is used to determine the reward for implementing the intelligent dispatch based on the current traffic state and the next traffic state of the intersection, and to construct replay experience based on the current traffic state, traffic light execution phase, reward and the next traffic state. The model update module is used to update the second artificial intelligence model based on the playback experience to obtain the updated second artificial intelligence model, and to use the updated second artificial intelligence model to enter the next round of intelligent scheduling.

Images