Parking space vehicle information recognition method and device based on target position correction

Abstract

The application discloses a parking space vehicle information identification method and device based on target position correction, and relates to the technical field of computer vision. The method comprises the following steps: using an image recognition algorithm, constructing a parking space vehicle detection and information identification model in the cloud, and deploying the parking space vehicle detection and information identification model to all edge servers of a parking lot; using the edge servers to collect to-be-identified image data of a target parking space, and inputting the to-be-identified image data to the parking space vehicle detection and information identification model to obtain a parking space vehicle detection result and a detection confidence; performing target position correction on the parking space vehicle detection result and the detection confidence to obtain a corrected parking space boundary box and a correction confidence; and using the parking space vehicle detection and information identification model to perform parking space vehicle information identification according to the corrected parking space boundary box and the correction confidence, and obtaining a vehicle information identification result. The method solves the problems of poor accuracy, poor adaptability and poor reliability in the prior art.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and apparatus for identifying parking space vehicle information based on target position correction. Background Technology

[0002] With the acceleration of urbanization and the surge in car ownership, parking management faces enormous challenges. Traditional parking management methods, such as manual inspections, inductive loop detectors, and ultrasonic sensors, suffer from low efficiency, high maintenance costs, and limited information dimensions (only able to determine the presence or absence of a vehicle). In recent years, intelligent parking management systems based on machine vision technology have gradually become the mainstream technology due to their advantages such as low cost, rich information, and ease of deployment.

[0003] The core technologies are parking space vehicle detection and vehicle information (mainly license plate) recognition. However, in practical applications, this technology still faces many challenges: 1) Poor accuracy: The varying lighting conditions in parking lots, such as strong light at entrances and exits, low light in dark corners, and weak light at night, can severely affect image quality, leading to a decline in model performance. Furthermore, the limited camera installation angle can easily cause perspective distortion. At the same time, occlusion between vehicles and between vehicles and pillars often results in partial obscuring of parking space lines or the vehicle itself, making it difficult for the model to accurately locate the parking space bounding box.

[0004] 2) Poor adaptability: Insufficient model generalization ability: Traditional model training methods typically involve training on a common dataset in the cloud and then directly deploying it to all parking lots. This approach ignores the differences in parking space marking patterns, lighting environments, and vehicle type distribution among different parking lots, resulting in poor model adaptability in specific scenarios.

[0005] 3) Poor reliability: In existing technologies, parking space vehicle detection and vehicle information recognition are often two independent steps. Errors in the previous detection step are directly propagated and amplified in the subsequent recognition step. For example, an off-center parking space bounding box can cause the subsequent license plate recognition module to capture an incorrect image region, resulting in recognition failure. The system lacks an effective feedback correction mechanism to correct front-end detection errors. Summary of the Invention

[0006] This invention provides a method and apparatus for identifying parking space vehicle information based on target location correction, which solves the problems of poor accuracy, poor adaptability and poor reliability in the prior art.

[0007] In a first aspect, embodiments of the present invention provide a method for identifying parking space vehicle information based on target location correction, the method comprising: Using image recognition algorithms, a parking space vehicle detection and information recognition model is built in the cloud, and the parking space vehicle detection and information recognition model is deployed to all edge servers in the parking lot; The edge server is used to collect the image data of the target parking space to be identified, and input it into the parking space vehicle detection and information recognition model to obtain the parking space vehicle detection results and detection confidence. The target position is corrected based on the vehicle detection results and detection confidence of the parking space to obtain the corrected parking space bounding box and the corrected confidence. Based on the corrected parking space bounding box and the corrected confidence level, the parking space vehicle detection and information recognition model is used to identify the parking space vehicle information and obtain the vehicle information recognition result.

[0008] The technical solution provided in this application has at least the following beneficial effects: By employing a target position correction mechanism, the system effectively utilizes the rich information from vehicle detection to infer and correct errors in parking space detection. This breaks the one-way error transmission in the traditional detection-recognition process, forming an intelligent closed-loop feedback system that significantly improves recognition accuracy in complex scenarios such as occlusion and poor lighting. The system utilizes a complex model based on image recognition algorithms, enhancing its adaptability to irregular parking spaces and occlusion. Furthermore, by integrating multi-task detection, it provides richer and more accurate vehicle information, offering a high-quality data source for subsequent position correction. Through multi-level confidence assessment and a progressive re-recognition strategy, the system achieves continuous tasks of parking space vehicle detection and vehicle information recognition, constructing a robust fault-tolerant system that ensures reliable results even under extremely difficult conditions, preventing management chaos caused by misidentification.

[0009] In one alternative implementation, an image recognition algorithm is used to build a parking space vehicle detection and information recognition model in the cloud, and this model is deployed to all edge servers in the parking lot, including: Using the cloud, we collect sample image data with preset bounding boxes and vehicle information labels, and based on the sample image data, we use image recognition algorithms to build an original parking space vehicle detection and information recognition model. The original parking space vehicle detection and information recognition model is deployed to all edge servers in the parking lot, and a local image dataset is collected in each edge server; The local image dataset with preset bounding boxes and vehicle information labels is input into the corresponding original parking space vehicle detection and information recognition model, and local iterative training is performed to obtain the trained parking space vehicle detection and information recognition model. In each local iteration of training, an optimization algorithm is used to optimize the hyperparameters of the trained parking space vehicle detection and information recognition model, and the optimized hyperparameters are applied to the next local iteration of training. If the local iterative training ends or the accuracy of the trained parking space vehicle detection and information recognition model is greater than the accuracy threshold, then the local parking space vehicle detection and information recognition model is output, the corresponding model parameter update amount is extracted, and uploaded to the cloud. Securely aggregate the model parameter updates uploaded by all edge servers in the cloud to obtain the corresponding global model parameters, and then distribute them to all edge servers. Based on the global model parameters, the local parking space and vehicle detection and information recognition model is updated on each edge server to obtain the final parking space and vehicle detection and information recognition model.

[0010] In one alternative implementation, the parking space vehicle detection and information recognition model includes a parking space detector, a vehicle detector, and an information recognizer, with the parallel parking space detector and vehicle detector both connected to the information recognizer.

[0011] In one alternative implementation, the parking space detector includes a first shared backbone network and parallel parking space bounding box detection branches, parking space key point detection branches, and a first confidence branch. The first shared backbone network is constructed based on the EfficientNet algorithm, and the first shared backbone network introduces deformable convolutions in the convolutional layers of the last at least two stages. The vehicle detector consists of a second shared backbone network, a neck network, and a multi-task unified detection head connected in sequence. The second shared backbone network is constructed based on the EfficientNet algorithm, and the neck network is an adaptive feature pyramid network constructed based on the AFPN algorithm. The adaptive feature pyramid network adds a corresponding one-channel attention module after each feature fusion layer of the FPN. The multi-task unified detection head includes parallel bounding box regression branches, orientation angle classification branches, type classification branches, and a second confidence branch.

[0012] In one alternative implementation, the COOA algorithm is used in each local iteration of training to optimize the hyperparameters of the trained parking space vehicle detection and information recognition model, and the optimized hyperparameters are applied to the next local iteration of training.

[0013] In one alternative implementation, during each local training iteration, the COOA algorithm is used to optimize the hyperparameters of the trained parking space vehicle detection and information recognition model, and the optimized hyperparameters are applied to the next local training iteration, including: The hyperparameters of the trained parking space vehicle detection and information recognition model are encoded into individual vectors of the COOA algorithm, and the fitness function, COOA population parameters and maximum number of iterations are set. Based on the COOA population parameters, the initial COOA population is obtained by initializing using the Tent chaotic mapping sequence; each COOA individual in the COOA population corresponds to a candidate hyperparameter. Based on the local image dataset, the fitness function is used to obtain the fitness value of each initial COOA individual, and the initial COOA individual with the best fitness value is taken as the optimal solution. Based on the hovering search behavior and diving capture behavior of the COOA algorithm, the initial COOA population is iteratively updated to obtain the updated COOA population. Introduce a dynamic reverse mechanism to update the probabilities. p The updated COOA individuals are dynamically reversed to obtain a reversed COOA population. Based on the local image dataset, the fitness function is used to obtain the fitness values ​​of each updated COOA individual and the reversed COOA individual, and the updated COOA individual or the reversed COOA individual with the best fitness value is updated as the optimal solution. When the number of iterations reaches the maximum number of iterations or the fitness value of the optimal solution meets the requirements, the iterative update of the COOA population is terminated, and the optimal solution of the current iteration is output. The individual vectors of the COOA individuals corresponding to the optimal solution are decoded to obtain the optimized hyperparameters of the trained parking space vehicle detection and information recognition model, and the optimized hyperparameters are applied to the next local iteration training.

[0014] In one optional implementation, an edge server is used to collect the image data of the target parking space to be identified, and the data is input into the parking space vehicle detection and information recognition model to obtain the parking space vehicle detection result and detection confidence level, including the following steps: Using cameras, video streams are captured of the target parking space and uploaded to an edge server within the communication range; Using an edge server, image data to be identified is extracted from the video stream, preprocessed to obtain preprocessed image data to be identified, and then input into the parking space vehicle detection and information recognition model. The parking space detector, which uses a parking space vehicle detection and information recognition model, performs parking space detection on the preprocessed image data to be identified, and obtains the parking space bounding box, key point coordinates and first confidence score of the candidate parking space. The vehicle detector using the parking space vehicle detection and information recognition model performs vehicle detection on the preprocessed image data to be recognized, and obtains the vehicle bounding box, vehicle type, vehicle orientation and second confidence score of the target vehicle. By integrating the parking space bounding box, key point coordinates, target vehicle bounding box, vehicle type, and vehicle orientation of the candidate parking space, the parking space vehicle detection result is obtained. The first confidence score and the second confidence score are then integrated to obtain the detection confidence score.

[0015] In one optional implementation, the vehicle detection results and detection confidence scores for parking spaces are corrected for target location to obtain the corrected parking space bounding box and corrected confidence score, including: If the first confidence level is lower than the preset first confidence threshold, the parking space position correction process is triggered; otherwise, the parking space bounding box of the candidate parking space is directly input into the information recognizer of the parking space vehicle detection and information recognition model. In the parking space location correction process, the vehicle bounding box with the highest spatial overlap with the parking space bounding box is matched, and the target location of the corresponding parking space bounding box is corrected according to the vehicle type and vehicle orientation of the target vehicle corresponding to the vehicle bounding box, as well as the key point coordinates of the candidate parking space, to obtain the corrected parking space bounding box. A revised confidence score is generated based on the second confidence score of the target vehicle, the intersection-union ratio of the vehicle bounding box and the revised parking space bounding box, the type matching score, and the angle alignment score. If the corrected confidence level is lower than the preset first confidence level threshold, the parking space location correction process is repeated; otherwise, the corrected parking space bounding box of the candidate parking space is input into the information recognizer.

[0016] In one optional implementation, based on the corrected parking space bounding box and corrected confidence level, a parking space vehicle detection and information recognition model is used to identify parking space vehicle information, obtaining the vehicle information recognition result, including: The information recognizer of the parking space vehicle detection and information recognition model is used to recognize the information of the image data area in the parking space bounding box / corrected parking space bounding box to obtain the license plate information recognition result and third confidence of the target vehicle. Based on the corrected confidence level and the third confidence level, the joint confidence level is calculated. If the joint confidence level is lower than the preset second confidence level threshold, the re-identification process is initiated. Otherwise, the vehicle information identification result is obtained by combining the license plate information identification result of the target vehicle, the vehicle type, and the vehicle orientation. In the first N re-identification processes, the model parameters of the parking space vehicle detection and information recognition model are adjusted, and the image acquisition step is returned to wait for the next re-identification process. In the N+1 to N+M re-identification processes, the image data of the target parking space to be identified is uploaded to the cloud, the parking space vehicle detection and information recognition process is carried out, and the obtained vehicle information recognition result is returned, waiting for the next re-identification process; In the N+M+1th re-identification process, the image data of the target parking space to be identified is sent to the manual review terminal, and the vehicle information recognition result is returned.

[0017] Secondly, embodiments of the present invention provide a parking space vehicle information recognition device based on target location correction, used to implement a parking space vehicle information recognition method, the device comprising: The model deployment unit is used to build a parking space vehicle detection and information recognition model in the cloud using image recognition algorithms, and to deploy the parking space vehicle detection and information recognition model to all edge servers in the parking lot; The parking space vehicle detection unit is used to correct the target position of the parking space vehicle detection results and detection confidence, and obtain the corrected parking space bounding box and corrected confidence. The target position correction unit is used to trigger a cross-camera matching event when any target to be matched leaves the current camera's field of view or enters the field of view of other cameras, and to filter out a set of candidate targets from other cameras based on global trajectory segments; The information recognition unit is used to identify parking space vehicle information based on the corrected parking space bounding box and corrected confidence level, using the parking space vehicle detection and information recognition model, and obtain the vehicle information recognition result.

[0018] A third aspect of this invention provides an electronic device, which includes: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by at least one processor, such that the at least one processor can perform the method proposed in the first aspect of the present invention.

[0019] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in the first aspect of the present invention. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention; Figure 2 This is a flowchart illustrating the steps of a parking space vehicle information recognition method based on target location correction provided in an embodiment of the present invention. Figure 3This is a functional unit diagram of a parking space vehicle information recognition device based on target position correction provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0022] The present invention will be further described below with reference to the accompanying drawings.

[0023] Reference Figure 1 , Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention.

[0024] like Figure 1 As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0025] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0026] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating device, a data storage module, a network communication module, a user interface module, and electronic programs.

[0027] exist Figure 1 In the electronic device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the electronic device of the present invention can be set in the electronic device. The electronic device calls the parking space vehicle information recognition device based on target position correction stored in the memory 1005 through the processor 1001 and executes the parking space vehicle information recognition method based on target position correction provided in the embodiment of the present invention.

[0028] Reference Figure 2 The present invention provides a parking space vehicle information recognition method based on target location correction, the method comprising: S201: Using image recognition algorithms, a parking space vehicle detection and information recognition model is built in the cloud, and the parking space vehicle detection and information recognition model is deployed to all edge servers in the parking lot; S202: Use an edge server to collect the image data to be identified for the target parking space, and input it into the parking space vehicle detection and information recognition model to obtain the parking space vehicle detection results and detection confidence. S203: Correct the target position of the vehicle detection results and detection confidence of the parking space to obtain the corrected parking space bounding box and the corrected confidence; S204: Based on the corrected parking space bounding box and the corrected confidence level, use the parking space vehicle detection and information recognition model to identify the parking space vehicle information and obtain the vehicle information recognition result.

[0029] The technical solution provided in this application has at least the following beneficial effects: By employing a target position correction mechanism, the system effectively utilizes the rich information from vehicle detection to infer and correct errors in parking space detection. This breaks the one-way error transmission in the traditional detection-recognition process, forming an intelligent closed-loop feedback system that significantly improves recognition accuracy in complex scenarios such as occlusion and poor lighting. The system utilizes a complex model based on image recognition algorithms, enhancing its adaptability to irregular parking spaces and occlusion. Furthermore, by integrating multi-task detection, it provides richer and more accurate vehicle information, offering a high-quality data source for subsequent position correction. Through multi-level confidence assessment and a progressive re-recognition strategy, the system achieves continuous tasks of parking space vehicle detection and vehicle information recognition, constructing a robust fault-tolerant system that ensures reliable results even under extremely difficult conditions, preventing management chaos caused by misidentification.

[0030] In one alternative implementation, an image recognition algorithm is used to build a parking space vehicle detection and information recognition model in the cloud, and this model is deployed to all edge servers in the parking lot, including: S2011: Using the cloud, collect several sample image data with preset bounding boxes and vehicle information labels, and use image recognition algorithms to construct the original parking space vehicle detection and information recognition model based on the sample image data. In this embodiment, a massive amount of sample image data from different parking lots, different environments, and different time periods is collected. All of these data are manually labeled and include tags such as parking space bounding boxes, parking space key points (such as the four corner points), vehicle bounding boxes, vehicle types (such as sedans, SUVs, and buses), vehicle orientation (such as the front of the vehicle facing left or right), and vehicle information (such as license plate numbers). Based on these labeled data, an original parking space vehicle detection and information recognition model is constructed. S2012: Deploy the original parking space vehicle detection and information recognition model to all edge servers in the parking lot, and collect local image datasets in each edge server; S2013: Input the local image dataset with preset bounding boxes and vehicle information labels into the corresponding original parking space vehicle detection and information recognition model, perform local iterative training, and obtain the trained parking space vehicle detection and information recognition model. S2014: In each local iteration of training, an optimization algorithm is used to optimize the hyperparameters of the trained parking space vehicle detection and information recognition model, and the optimized hyperparameters are applied to the next local iteration of training, thereby accelerating model convergence and improving performance. S2015: If the local iterative training ends or the accuracy of the trained parking space vehicle detection and information recognition model is greater than the accuracy threshold, then output the local parking space vehicle detection and information recognition model, extract the corresponding model parameter update amount, and upload it to the cloud. In this embodiment, when the local model training of the parking lot reaches the preset accuracy threshold or completes the set number of rounds, the training stops. At this time, the edge server does not upload the entire model (containing a large amount of data and privacy), but only calculates the model parameter update amount (i.e., weight gradient or weight difference) generated by this training, and uses a secure aggregation protocol to encrypt or anonymize the update amount before uploading it to the cloud. S2016: Securely aggregate the model parameter updates uploaded by all edge servers in the cloud to obtain the corresponding global model parameters, and then distribute them to all edge servers. In this embodiment, the cloud distributes this new global model parameter to all edge servers. Each edge server uses this parameter to update its local model, resulting in a final model with better performance and stronger generalization ability. This process is carried out periodically, realizing the silent and continuous evolution of the model. S2017: Based on the global model parameters, the local parking space vehicle detection and information recognition model is updated on each edge server to obtain the final parking space vehicle detection and information recognition model.

[0031] In one alternative implementation, the parking space vehicle detection and information recognition model includes a parking space detector, a vehicle detector, and an information recognizer, with the parallel parking space detector and vehicle detector both connected to the information recognizer.

[0032] In one alternative implementation, the parking space detector includes a first shared backbone network and parallel parking space bounding box detection branches, parking space key point detection branches, and a first confidence branch. The first shared backbone network is constructed based on the EfficientNet algorithm, and the first shared backbone network introduces deformable convolutions in the convolutional layers of the last at least two stages. The vehicle detector comprises a second shared backbone network, a neck network, and a multi-task unified detection head, which are connected in sequence. The second shared backbone network is constructed based on the EfficientNet algorithm. The neck network is an adaptive feature pyramid network constructed based on the Adaptive Feature Pyramid Network (AFPN) algorithm. The adaptive feature pyramid network adds a corresponding one-channel attention module after each feature fusion layer of the Feature Pyramid Network (FPN). The multi-task unified detection head includes parallel bounding box regression branches, orientation angle classification branches, type classification branches, and a second confidence branch.

[0033] In this embodiment, the parking space detector includes a first shared backbone network for extracting multi-scale feature maps of the image; a parking space bounding box detection branch connected after the backbone network for predicting the bounding box coordinates of the parking space; a parking space key point detection branch running parallel to the parking space bounding box detection branch and also connected after the first shared backbone network for predicting the positions of key points such as parking space corners or center lines; and a first confidence branch for outputting a first confidence score. Deformable convolutions are introduced in at least the last two stages of the convolutional layers of the first shared backbone network to enhance the model's ability to extract features from tilted, occluded, and non-standard shaped parking spaces. The loss functions of the parking space bounding box detection branch, the parking space key point detection branch, and the first confidence branch are jointly optimized so that the model learns the internal geometry of the parking space while learning to locate the parking space bounding box.

[0034] In the vehicle detector, a second shared backbone network and a neck network are used to extract multi-scale features of the image. The channel attention module of the neck network is used to dynamically adjust the weights of features at different scales, so that the network can adaptively enhance the most effective feature layer when facing large-scale vehicles (such as trucks in the foreground) and small-scale vehicles (such as cars in the distance), thereby improving the detection performance of multi-scale and occluded vehicles. In the multi-task unified detection head, the detection head contains four parallel prediction sub-tasks: a bounding box regression branch, used to predict the bounding box coordinates of the vehicle; an orientation angle classification branch, used to discretize the vehicle orientation into N categories (such as 0°, 45°, 90°, 135°) for classification prediction; a type classification branch, used to predict the type of vehicle (such as car, SUV, bus, truck); and a second confidence branch, used to output the second confidence. The total loss function of the vehicle detector is the weighted sum of the loss functions of the four sub-tasks, realizing end-to-end multi-task joint training.

[0035] Both the parking space detector and the vehicle detector employ advanced network designs. The parking space detector introduces deformable convolution and keypoint detection, enhancing its adaptability to irregular parking spaces and occlusions. The vehicle detector uses AFPN and channel attention, and integrates multi-task learning (position, type, orientation), providing richer and more accurate vehicle information, thus offering a high-quality data source for subsequent position correction.

[0036] In one alternative implementation, during each local iteration of training, the hyperparameters of the trained parking space vehicle detection and information recognition model are optimized using the Convergent Osprey Optimization Algorithm (COOA), and the optimized hyperparameters are applied to the next local iteration of training.

[0037] In this embodiment, the hyperparameters are the learning rate, batch size, and weight decay of the parking space detector, vehicle detector, and information recognizer.

[0038] In one alternative implementation, during each local training iteration, the COOA algorithm is used to optimize the hyperparameters of the trained parking space vehicle detection and information recognition model, and the optimized hyperparameters are applied to the next local training iteration, including: S20141: Encode the hyperparameters of the trained parking space vehicle detection and information recognition model into individual vectors of the COOA algorithm, and set the fitness function, COOA population parameters and maximum number of iterations; The formula is: In the formula, For COOA individuals Xfitness value; The parking space positioning error measures the accuracy of the predicted parking space bounding box. To measure the accuracy of the judgment on whether a parking space is occupied, considering the classification error; This is to measure the accuracy of license plate information recognition based on information recognition error. This is the error weighting coefficient, which can be adjusted according to actual needs; X Parameters for COOA individuals; S20142: Based on the COOA population parameters, the initial COOA population is obtained by initializing using the Tent chaotic mapping sequence; each COOA individual in the COOA population corresponds to a candidate hyperparameter. The formula is: In the formula, For the initial COOA population, the first i One initial COOA individual; For the first i One chaotic variable; These are the upper and lower bounds of the search space; i For individual COOA indicators; In the formula, For the first i- One chaotic variable; compared with random initialization, chaotic initialization can ensure that the population is evenly distributed in the solution space, thus enhancing diversity. S20143: Based on the local image dataset, use the fitness function to obtain the fitness value of each initial COOA individual, and take the initial COOA individual with the best fitness value as the optimal solution; In this embodiment, several historical sample data are divided into a training sample set and a test sample set in a 7:3 ratio. The candidate hyperparameters corresponding to the initial COOA individuals are substituted into the model. The model is trained using the training sample set and evaluated on the test sample set. The corresponding fitness value is obtained according to the fitness function. S20144: Based on the hovering search behavior and diving capture behavior of the COOA algorithm, the initial COOA population is iteratively updated to obtain the updated COOA population. The formula is: In the formula, Number of iterations t+ The first result obtained from the hovering search behavior is the first i One updated COOA individual; The convergence factor; Number of iterationst The COOA individuals are the initial COOA individuals during the first iteration update; Number of iterations t The optimal solution; For randomly selected COOA individuals; Flight radius; To explore and develop weights; In the formula, These are the maximum and minimum values ​​of the convergence factor; This represents the maximum number of iterations. t This represents the current iteration number; , To adjust the parameters; It is the hyperbolic tangent function; In the formula, Number of iterations t+ The first dive-and-capture maneuver yielded the [number missing]th [missing information]. i One updated COOA individual; for Levy Distribute random numbers; b for Levy Step length, and b ∈[1,2]; S20145: Introduce a dynamic reverse mechanism to update probabilities. p The updated COOA individuals are dynamically reversed to obtain a reversed COOA population. In the formula, Number of iterations t+ 1 of i One reverse COOA individual; For the updated COOA population, the first i One updated COOA individual; For the updated COOA population, the updated first i The reverse solution for each COOA individual; The fitness function; S20146: Based on the local image dataset, use the fitness function to obtain the fitness value of each updated COOA individual and the reversed COOA individual, and update the updated COOA individual or the reversed COOA individual with the best fitness value as the optimal solution. S20147: When the number of iterations reaches the maximum number of iterations or the fitness value of the optimal solution meets the requirements, terminate the iterative update of the COOA population and output the optimal solution of the current iteration; S20148: Decode the individual vector of the COOA individual corresponding to the optimal solution to obtain the optimized hyperparameters of the trained parking space vehicle detection and information recognition model, and apply the optimized hyperparameters to the next local iteration training.

[0039] In one optional implementation, an edge server is used to collect the image data of the target parking space to be identified, and the data is input into the parking space vehicle detection and information recognition model to obtain the parking space vehicle detection result and detection confidence level, including the following steps: S2021: Use a camera to capture video streams of the target parking space and upload the video streams to an edge server within the communication range; In this embodiment, high-definition network cameras are deployed at key locations in the parking lot (such as lane intersections, above parking space areas, entrances and exits, etc.). Cameras with wide dynamic range and low-light performance are used to cope with drastic changes in light in the parking lot (such as backlight at entrances and exits) or low-light environments such as night. The installation position and angle of the cameras need to be carefully designed to ensure that the target parking space can be clearly covered and to minimize the obstruction between vehicles. S2022: Using an edge server, extract the image data to be identified from the video stream, preprocess the image data to be identified to obtain the preprocessed image data to be identified, and input the preprocessed image data to be identified into the parking space vehicle detection and information recognition model. In this embodiment, the preprocessing includes: Image enhancement: Histogram equalization or adaptive gamma correction is performed on the image data to be recognized to enhance image contrast, especially in scenes with uneven lighting. Distortion correction: If the camera lens has obvious barrel or pincushion distortion, the system will pre-calibrate the distortion parameters of the camera and perform distortion correction on the acquired image data to be identified to restore the true geometric shape, which is crucial for the accuracy of subsequent parking space detection.

[0040] Image cropping: If the system only cares about a specific region of interest (ROI) in the image, the ROI coordinates can be preset, and only the image of that region can be processed each time to reduce the amount of computation and improve the processing speed; S2023: A parking space detector using a parking space vehicle detection and information recognition model performs parking space detection on preprocessed image data to be recognized, and obtains the parking space bounding box, key point coordinates and first confidence score of candidate parking spaces. S2024: A vehicle detector using a parking space vehicle detection and information recognition model performs vehicle detection on the preprocessed image data to be recognized, and obtains the vehicle bounding box, vehicle type, vehicle orientation and second confidence level of the target vehicle. S2025: Integrate the parking space bounding box, key point coordinates, target vehicle bounding box, vehicle type, and vehicle orientation of the candidate parking space to obtain the parking space vehicle detection result, and integrate the first confidence and the second confidence to obtain the detection confidence.

[0041] In one optional implementation, the vehicle detection results and detection confidence scores for parking spaces are corrected for target location to obtain the corrected parking space bounding box and corrected confidence score, including: S2031: If the first confidence level is lower than the preset first confidence threshold, the parking space position correction process is triggered; otherwise, the parking space bounding box of the candidate parking space is directly input into the information recognizer of the parking space vehicle detection and information recognition model. S2032: In the parking space position correction process, match the vehicle bounding box with the highest spatial overlap with the parking space bounding box, and correct the target position of the corresponding parking space bounding box according to the vehicle type and vehicle orientation of the target vehicle corresponding to the vehicle bounding box, as well as the key point coordinates of the candidate parking space, to obtain the corrected parking space bounding box. In this embodiment, target position correction includes: Position correction: Using the center point of the high-confidence vehicle bounding box as a reference, combined with the standard parking space size and the key point coordinates of the candidate parking space, the center and length and width of the parking space box are recalculated. This is equivalent to using a definite vehicle to calibrate a fuzzy parking space. Size Correction: By integrating vehicle type information, since the vehicle is identified as an "SUV", the system will automatically adjust the corrected parking space frame size to the standard SUV parking space size, instead of the default sedan parking space size, to make it more realistic; Angle Correction: Integrating vehicle orientation, since the vehicle orientation is "90°", the system will also adjust the rotation angle θ of the corrected parking space frame to 90° to perfectly align it with the vehicle; S2033: Generate a corrected confidence score based on the second confidence score of the target vehicle, the intersection-union ratio of the vehicle bounding box and the corrected parking space bounding box, the type matching score, and the angle alignment score; The formula is: In the formula, To adjust the confidence level; The second confidence level for the target vehicle; The intersection-union ratio (IUU) is the ratio of the vehicle bounding box to the corrected parking space bounding box. The higher the IUU, the more reasonable the correction. Rate the type matching degree. For example, an SUV parked in a standard parking space has a high matching degree; a sedan parked in a large parking space also has a high matching degree. The more closely the logical relationship between the direction of the parking space frame and the vehicle's orientation is matched to the angle alignment score, the higher the score. To adjust the confidence coefficient; S2034: If the corrected confidence level is lower than the preset first confidence level threshold, repeat the parking space position correction process; otherwise, input the corrected parking space bounding box of the candidate parking space into the information recognizer. In this embodiment, the information recognizer is typically an Optical Character Recognition (OCR) model specifically designed for license plate recognition. The information recognition includes: License plate localization: Within the image area of ​​the corrected parking space bounding box, the position of the license plate is further precisely located; Character segmentation: Segmenting the located license plate image into individual characters; Character recognition: Each character image is input into a classifier to identify specific Chinese characters, letters, and numbers.

[0042] In one optional implementation, based on the corrected parking space bounding box and corrected confidence level, a parking space vehicle detection and information recognition model is used to identify parking space vehicle information, obtaining the vehicle information recognition result, including: S2041: Using the information recognizer of the parking space vehicle detection and information recognition model, information recognition is performed on the image data area within the parking space bounding box / corrected parking space bounding box to obtain the license plate information recognition result and third confidence level of the target vehicle; S2042: Calculate the joint confidence based on the corrected confidence and the third confidence. If the joint confidence is lower than the preset second confidence threshold, start the re-identification process. Otherwise, combine the license plate information recognition result of the target vehicle, the vehicle type, and the vehicle orientation to obtain the vehicle information recognition result. The formula is: In the formula, For joint confidence level; The third confidence level; This is the joint confidence coefficient; S2043: In the first N re-identification processes, adjust the model parameters of the parking space vehicle detection and information recognition model, and return to the image acquisition step to be identified, waiting for the next re-identification process; In this embodiment, the re-identification process may be an accidental error caused by the poor current state of the model. It will slightly adjust some parameters of the model (such as lowering the recognition threshold or switching to the backup model) and then return to step S2023 to re-perform the complete detection, correction and recognition process for the current frame or the next frame image. This gives the model a chance to self-correct. S2044: In the N+1 to N+M re-identification processes, the image data to be identified of the target parking space is uploaded to the cloud, the parking space vehicle detection and information recognition process is carried out, and the obtained vehicle information recognition result is returned, waiting for the next re-identification process; In this embodiment, if the local fine-tuning fails after M retries, it is assumed that the local model may no longer be able to handle this difficult case. The image data to be identified in the current frame is encrypted and uploaded to the cloud. The cloud has more powerful computing resources and a more powerful expert model to recalculate the image and return the result to the edge server. S2045: In the N+M+1th re-identification process, the image data to be identified for the target parking space is sent to the manual review terminal, and the vehicle information identification result is returned. In this embodiment, if a high-confidence result still cannot be given after recalculating M+1 times in the cloud, it indicates that this may be an extremely rare case (such as a license plate that is severely damaged or has special angles of reflection). The case is packaged and pushed to a human reviewer through the management backend for final confirmation by a human, and the confirmation result is returned. At the same time, this difficult case and its human annotation results will be carefully saved as "golden data" for the next round of federated learning to further improve the intelligence level of the model.

[0043] This invention also provides a parking space vehicle information recognition device based on target location correction, referring to... Figure 3 The diagram shows a functional unit diagram of a parking space vehicle information recognition device 300 based on target location correction according to the present invention. The device may include the following units: The model deployment unit 301 is used to build a parking space vehicle detection and information recognition model in the cloud using image recognition algorithms, and to deploy the parking space vehicle detection and information recognition model to all edge servers in the parking lot; The parking space vehicle detection unit 302 is used to correct the target position of the parking space vehicle detection results and detection confidence, and obtain the corrected parking space bounding box and corrected confidence. The target position correction unit 303 is used to trigger a cross-camera matching event when any target to be matched leaves the current camera's field of view or enters the field of view of other cameras, and to filter out a set of candidate targets from other cameras based on global trajectory segments; The information recognition unit 304 is used to perform parking space vehicle information recognition based on the corrected parking space bounding box and the corrected confidence level, using the parking space vehicle detection and information recognition model, and obtain the vehicle information recognition result.

[0044] Based on the same inventive concept, another embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the parking space vehicle information recognition method based on target position correction of the present invention.

[0045] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned terminal and other devices. The memory can include Random Access Memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0046] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0047] Furthermore, to achieve the above objectives, embodiments of the present invention also propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the parking space vehicle information recognition method based on target location correction according to embodiments of the present invention.

[0048] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable vehicles (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0049] The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (apparatus), and computer program products according to embodiments of the invention. 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0050] 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.

[0051] 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.

[0052] 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. "And / or" indicates that either one or both can be chosen. 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 the element.

[0053] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying parking space vehicle information based on target location correction, characterized in that, The method includes: Using image recognition algorithms, a parking space vehicle detection and information recognition model is built in the cloud and deployed to all edge servers in the parking lot; The parking space vehicle detection and information recognition model includes a parking space detector, a vehicle detector, and an information recognizer, with the parallel parking space detector and vehicle detector both connected to the information recognizer. The image data to be identified for the target parking space is collected using an edge server and input into the parking space vehicle detection and information recognition model, which includes: Using a parking space detector, parking spaces are detected in the preprocessed image data to be identified, and the parking space bounding box, key point coordinates, and first confidence score of the candidate parking spaces are obtained. Using a vehicle detector, vehicle detection is performed on the preprocessed image data to be identified, and the vehicle bounding box, vehicle type, vehicle orientation, and second confidence score of the target vehicle are obtained. By integrating the parking space bounding box, key point coordinates, vehicle bounding box, vehicle type, and vehicle orientation, the parking space vehicle detection results are obtained. The first confidence score and the second confidence score are then integrated to obtain the detection confidence score. The target position is corrected based on the vehicle detection results and detection confidence level of the parking space, including: If the first confidence level is lower than the first confidence threshold, the parking space position correction process is triggered; otherwise, the parking space bounding box is directly input into the information recognizer. In the parking space position correction process, the target position of the corresponding parking space bounding box is corrected based on the vehicle type and vehicle orientation of the vehicle bounding box with the highest spatial overlap with the parking space bounding box, as well as the key point coordinates, to obtain the corrected parking space bounding box. A revised confidence score is generated based on the second confidence score, the intersection-union ratio of the vehicle bounding box and the revised parking space bounding box, the type matching score, and the angle alignment score. If the corrected confidence level is lower than the first confidence level threshold, the parking space position correction process is repeated; otherwise, the corrected parking space bounding box is input to the information recognizer. Based on the corrected parking space bounding box and corrected confidence level, a parking space vehicle detection and information recognition model is used to identify parking space vehicle information, including: Using an information recognizer, information recognition is performed on the image data area within the parking space boundary box or the corrected parking space boundary box to obtain the license plate information recognition result and third confidence level of the target vehicle. Based on the corrected confidence level and the third confidence level, a joint confidence level is calculated. If the joint confidence level is lower than the second confidence level threshold, the re-identification process is initiated. Otherwise, the vehicle information identification result is obtained by identifying the license plate information of the target vehicle, the vehicle type, and the vehicle orientation. In the previous N re-identification processes, the model parameters of the parking space vehicle detection and information recognition model are adjusted, and the image acquisition step is returned to wait for the next re-identification process. In the N+1 to N+M re-identification processes, the image data to be identified is uploaded to the cloud for parking space vehicle detection and information recognition processes, and the obtained vehicle information recognition results are returned, waiting for the next re-identification process; In the N+M+1th re-identification process, the image data to be identified is sent to the manual review terminal, and the obtained vehicle information recognition result is returned.

2. The parking space vehicle information recognition method based on target position correction according to claim 1, characterized in that, Using image recognition algorithms, a parking space vehicle detection and information recognition model is built in the cloud and deployed to all edge servers in the parking lot, including: Using the cloud, we collect sample image data with preset bounding boxes and vehicle information labels, and based on the sample image data, we use image recognition algorithms to build an original parking space vehicle detection and information recognition model. The original parking space vehicle detection and information recognition model is deployed to all edge servers in the parking lot, and a local image dataset is collected in each edge server; The local image dataset with preset bounding boxes and vehicle information labels is input into the corresponding original parking space vehicle detection and information recognition model, and local iterative training is performed to obtain the trained parking space vehicle detection and information recognition model. In each local iteration of training, an optimization algorithm is used to optimize the hyperparameters of the trained parking space vehicle detection and information recognition model, and the optimized hyperparameters are applied to the next local iteration of training. If the local iterative training ends or the accuracy of the trained parking space vehicle detection and information recognition model is greater than the accuracy threshold, then the local parking space vehicle detection and information recognition model is output, the corresponding model parameter update amount is extracted, and uploaded to the cloud. Securely aggregate the model parameter updates uploaded by all edge servers in the cloud to obtain the corresponding global model parameters, and then distribute them to all edge servers. Based on the global model parameters, the local parking space and vehicle detection and information recognition model is updated on each edge server to obtain the final parking space and vehicle detection and information recognition model.

3. The parking space vehicle information recognition method based on target position correction according to claim 2, characterized in that, The parking space detector includes a first shared backbone network and parallel parking space box detection branches, parking space key point detection branches and a first confidence branch. The first shared backbone network is constructed based on the EfficientNet algorithm, and the first shared backbone network introduces deformable convolution in the convolutional layers of the last at least two stages. The vehicle detector includes a second shared backbone network, a neck network, and a multi-task unified detection head connected in sequence. The second shared backbone network is constructed based on the EfficientNet algorithm. The neck network is an adaptive feature pyramid network constructed based on the AFPN algorithm. The adaptive feature pyramid network adds a corresponding one-channel attention module after each feature fusion layer of the FPN. The multi-task unified detection head includes parallel bounding box regression branches, orientation angle classification branches, type classification branches, and a second confidence branch.

4. The parking space vehicle information recognition method based on target position correction according to claim 3, characterized in that, In each local iteration of training, the convergent Osprey-optimized COOA algorithm is used to optimize the hyperparameters of the trained parking space vehicle detection and information recognition model, and the optimized hyperparameters are applied to the next local iteration of training.

5. The parking space vehicle information recognition method based on target position correction according to claim 4, characterized in that, In each local training iteration, the COOA algorithm is used to optimize the hyperparameters of the trained parking space vehicle detection and information recognition model, and the optimized hyperparameters are applied to the next local training iteration, including: The hyperparameters of the trained parking space vehicle detection and information recognition model are encoded into individual vectors of the COOA algorithm, and the fitness function, COOA population parameters and maximum number of iterations are set. Based on the COOA population parameters, the initial COOA population is obtained by initializing using the Tent chaotic mapping sequence; each COOA individual in the COOA population corresponds to a candidate hyperparameter. Based on the local image dataset, the fitness function is used to obtain the fitness value of each initial COOA individual, and the initial COOA individual with the best fitness value is taken as the optimal solution. Based on the hovering search behavior and diving capture behavior of the COOA algorithm, the initial COOA population is iteratively updated to obtain the updated COOA population. Introduce a dynamic reverse mechanism to update the probabilities. p The updated COOA individuals are dynamically reversed to obtain a reversed COOA population. Based on the local image dataset, the fitness function is used to obtain the fitness values ​​of each updated COOA individual and the reversed COOA individual, and the updated COOA individual or the reversed COOA individual with the best fitness value is updated as the optimal solution. When the number of iterations reaches the maximum number of iterations or the fitness value of the optimal solution meets the requirements, the iterative update of the COOA population is terminated, and the optimal solution of the current iteration is output. The individual vectors of the COOA individuals corresponding to the optimal solution are decoded to obtain the optimized hyperparameters of the trained parking space vehicle detection and information recognition model, and the optimized hyperparameters are applied to the next local iteration training.

6. The parking space vehicle information recognition method based on target position correction according to claim 5, characterized in that, The image data to be identified for the target parking space is collected using an edge server and input into the parking space vehicle detection and information recognition model to obtain the parking space vehicle detection results and detection confidence. This also includes: Using cameras, video streams are captured of the target parking space and uploaded to an edge server within the communication range; Using an edge server, image data to be identified is extracted from the video stream, preprocessed to obtain preprocessed image data to be identified, and then input into the parking space vehicle detection and information recognition model.

7. A parking space vehicle information recognition device based on target location correction, used to implement the parking space vehicle information recognition method as described in any one of claims 1-6, characterized in that, The device includes: The model deployment unit is used to build a parking space vehicle detection and information recognition model in the cloud using image recognition algorithms, and to deploy the parking space vehicle detection and information recognition model to all edge servers in the parking lot; The parking space vehicle detection unit is used to correct the target position of the parking space vehicle detection results and detection confidence, and obtain the corrected parking space bounding box and corrected confidence. The target position correction unit is used to trigger a cross-camera matching event when any target to be matched leaves the current camera's field of view or enters the field of view of other cameras, and to filter out a set of candidate targets from other cameras based on global trajectory segments; The information recognition unit is used to identify parking space vehicle information based on the corrected parking space bounding box and corrected confidence level, using the parking space vehicle detection and information recognition model, and obtain the vehicle information recognition result.

Images