An artificial intelligence-based parking lot operation management system and method
By using multimodal data fusion and key binding mechanisms, the problem of high-precision vehicle identification in complex scenarios of parking lot license plate recognition systems has been solved, realizing efficient unattended vehicle management and improving the system's automation and security.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG ZHIBO INFORMATION TECH CO LTD
- Filing Date
- 2026-03-05
- Publication Date
- 2026-05-19
AI Technical Summary
Existing parking lot license plate recognition systems cannot achieve high-accuracy self-checkout when faced with vehicles displaying temporary license plates, obscured or damaged license plates, or vehicles being test-driven without license plates, resulting in a decline in automation and user experience.
A multimodal fusion vehicle identification method is adopted, which extracts and encodes features from visual images, millimeter-wave radar point cloud data and acoustic signals, generates vehicle feature sequences by combining convolutional neural networks and point cloud encoders, and uses key binding and dual-path similarity verification mechanisms to achieve high-precision identification without human intervention.
It enables high-precision vehicle identification in complex scenarios, ensures system robustness and data privacy, improves the level of automated operation, and has high scalability and deployment flexibility.
Smart Images
Figure CN121768089B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, specifically to an artificial intelligence-based parking lot operation management system and method. Background Technology
[0002] As a key component of urban traffic management, parking lot operation and management systems are increasingly being used as important indicators of modern urban governance capabilities in terms of automation, precision, and robustness. In current mainstream technical architectures, parking lots commonly employ vehicle identification mechanisms based on license plate recognition. This involves capturing images of the front and rear license plates of vehicles using cameras deployed at entrances and exits, and then using optical character recognition algorithms to parse the license plate numbers. This information is then linked to billing policies and user accounts to achieve unattended self-service payment processes.
[0003] The inherent characteristics of the aforementioned technical solutions at the principle level have gradually revealed their limitations in addressing new challenges. Specifically, when faced with special circumstances such as vehicles with temporary license plates, obscured or damaged license plates, or vehicles being test-driven without license plates, traditional license plate recognition systems fail due to a lack of effective text information. This prevents such vehicles from completing self-service settlement, necessitating manual intervention and severely weakening the overall automation level and user experience of the system. Therefore, how to achieve highly accurate, highly secure, and fully automated vehicle identification and billing linkage in complex real-world scenarios without relying heavily on physical license plates has become a key challenge and an urgent technical problem for those skilled in the art. Summary of the Invention
[0004] The purpose of this invention is to provide an artificial intelligence-based parking lot operation management system and method to solve the problems raised in the prior art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a parking lot operation and management method based on artificial intelligence.
[0006] Vehicle feature data is collected. Visual images, millimeter-wave radar point cloud data, and acoustic signals of vehicles entering or exiting the parking lot are collected simultaneously in the parking lot entrance area. The visual images are captured by a visible light camera, covering the front and side front views of the vehicle; the millimeter-wave radar point cloud data is acquired by a millimeter-wave radar array to form a three-dimensional spatial coordinate point set; and the acoustic signals are collected by a directional microphone array to record a three-second audio clip of the engine running when the vehicle is running.
[0007] The three types of raw data were standardized. For visual images, Gaussian filtering was used to remove high-frequency noise, perspective distortion was corrected by affine transformation, and the images were cropped to a uniform size. For millimeter-wave radar point cloud data, ground reflection points and static background points were first removed, and then the point cloud density was normalized to a fixed number of points by voxel grid downsampling. For acoustic signals, pre-emphasis processing was first performed to enhance high-frequency components, and then Mel-frequency cepstral coefficients were calculated after frame segmentation and windowing to generate a fixed-length spectral feature vector.
[0008] Features from three types of raw data are extracted and concatenated to form a vehicle feature sequence. Visual images are input into the backbone of a convolutional neural network, outputting a visual feature vector fv; millimeter-wave point cloud data is input into a PointNet-based point cloud encoder, outputting a radar feature vector fr; and acoustic signals are input into a one-dimensional convolutional temporal encoder, outputting an acoustic feature vector fa. These three feature vectors are concatenated along the channel dimension to form the vehicle's original feature sequence F, where F = [fv, fr, fa].
[0009] Normalize the original vehicle feature sequence F to output the vehicle feature sequence E. ;
[0010] Where W is the weight matrix and b is the bias term. This represents the L2 norm. The embedding representation E is the feature sequence of the vehicle.
[0011] Obtain the feature sequence E of the current vehicle in Iterate through the feature sequence set corresponding to all existing records in the database, and denote the i-th feature sequence in the feature sequence set as E. i ;
[0012] Calculate E in The cosine similarity s with the i-th feature sequence in the feature sequence set i : ,
[0013] If there exists any s i >τ sim , where τ sim If the preset similarity threshold is met, the vehicle entering the venue is determined to be highly similar to a parked vehicle in the database, triggering the key binding process.
[0014] A unique key k is generated using a key derivation algorithm. This key k is then used to analyze the feature sequence E of the currently entering vehicle. in Perform a symmetric encryption operation to generate the encrypted feature information sequence C. in C in =Enc(E in,k), where Enc represents the encryption function, and Cin is written to the database along with the key identifier, timestamp, and vehicle entry event metadata.
[0015] If the traversal results do not find any s i >τ sim Then directly send the unencrypted E in It is stored in the database as a plaintext feature sequence and is not associated with any key.
[0016] Record the start time of key validity t0. Key k has a maximum valid duration Tmax. It is a valid key within the time interval [t0, t0+Tmax]. If it exceeds this interval, it is marked as an expired key.
[0017] Car owners can scan a QR code at the parking lot entrance using a self-service device, or the parking management system can bind the car owner's identity and key through the cloud platform and the car owner's network terminal.
[0018] When a vehicle leaves the parking lot exit area, the feature acquisition, preprocessing, and encoding process is repeated to generate the feature sequence E of the exiting vehicle. out Retrieve all records from the database and calculate E. out The cosine similarity with each stored feature sequence is used to select those that satisfy s. j >τ sim The candidate set S, where j represents the j-th plaintext feature sequence in the database.
[0019] If the vehicle leaving the site is associated with a valid key k d If the candidate set S is not empty, then perform dual-path similarity verification: the first path calculates the field feature sequence E. out Compared with the j-th candidate plaintext feature sequence E in the database j The cosine similarity is denoted as the first similarity. The second path uses the valid key k. d For E out Encryption generation C out C out =Enc(E out k d ), and search for C in all encrypted feature information sequences in the database. out The most similar entry is used to calculate its decrypted feature sequence and E. out The similarity can be expressed as either the first similarity or the second similarity index, which can be calculated directly in the ciphertext space using the Hamming distance. The normalized similarity index is denoted as the second similarity q. j,2 The final similarity score is the maximum of the two scores, h. ;
[0020] Returns the largest h that exceeds the validation threshold τ. verify Vehicle identification information.
[0021] If the vehicle leaving the site is not bound to a key, or if it is bound to a key but the candidate set S is empty, then single-path similarity verification is performed.
[0022] In this case, only the first similarity q is calculated. j,1 Or calculate the second similarity q based solely on ciphertext matching. j,2 And determine whether it is greater than the single-path similarity threshold. If the conditions are met, the corresponding vehicle identity is returned; otherwise, the matching is deemed unsuccessful, and a manual processing procedure is triggered.
[0023] If the key associated with a vehicle leaving the parking lot has been determined by the system to be an expired key, an abnormal prompt message will be sent to the parking lot management personnel terminal regardless of the similarity calculation result. The message will include a summary of vehicle characteristics, entry and exit timestamps, and key expiration status.
[0024] The system includes:
[0025] The vehicle feature acquisition module is used to simultaneously acquire visual images, millimeter-wave radar point cloud data, and acoustic signals of vehicles entering or leaving the vehicle.
[0026] The feature preprocessing module is used to perform Gaussian filtering and affine transformation on visual images and then crop them to a uniform size. After removing ground reflection points and static background points from millimeter-wave radar point cloud data, voxel grid downsampling is performed to normalize the number of points. The acoustic signal is pre-emphasized, framed and windowed, and Mel frequency cepstral coefficients are calculated to generate spectral feature vectors.
[0027] The feature encoding module is used to concatenate the three types of preprocessed feature vectors to form a vehicle feature sequence, and then normalize and map the encoded feature sequence.
[0028] The key management module is used to generate a key with a valid duration limit based on the vehicle owner's identity authentication during the entry stage, and associate the key with the corresponding vehicle's feature sequence within the key's validity period;
[0029] An information storage database is used to store plaintext feature sequences or feature information sequences encrypted with a key.
[0030] The similarity verification module is used to adaptively select a single-path or dual-path similarity verification strategy during the exit stage based on whether there is a valid key and whether there are candidate matching records that meet the preset similarity threshold.
[0031] The vehicle feature acquisition module is deployed at the entrance and exit areas of the parking lot to simultaneously acquire visual images, millimeter-wave radar point cloud data, and acoustic signals of vehicles entering or exiting. The visual images are captured by two visible light cameras, covering the frontal and side-view perspectives of the vehicle, respectively. The millimeter-wave radar point cloud data is acquired by a four-element millimeter-wave radar array, forming a three-dimensional spatial coordinate point set. The acoustic signals are collected by a directional microphone array, recording a three-second audio clip of the engine running while the vehicle is in operation.
[0032] The feature preprocessing module receives raw multimodal data from the vehicle feature acquisition module and performs standardization processing on each data.
[0033] The feature encoding module receives three types of standardized data output by the feature preprocessing module and inputs them into the corresponding deep neural network models to extract high-level semantic features.
[0034] The key management module determines whether to generate an encryption key with time constraints based on the feature matching results and the vehicle owner's identity authentication status during the vehicle entry stage.
[0035] The information storage database adopts a distributed key-value storage architecture, supporting high-concurrency read and write operations and fine-grained access control. Each record in the database contains the following fields: vehicle characteristic sequence, including plaintext or ciphertext sequence; key identifier, including whether the key exists; timestamp; entry / exit event type; and key status flag.
[0036] When a vehicle leaves the parking lot exit area, the similarity verification module repeatedly executes the feature acquisition, preprocessing and encoding process to generate the feature sequence of the exiting vehicle and adaptively selects a single-path or dual-path similarity verification strategy.
[0037] Compared with the prior art, the beneficial effects of the present invention are:
[0038] 1. By constructing a multimodal fusion virtual identity system, the reliance on physical license plates is completely eliminated, and high-precision identity recognition can still be achieved in scenarios with no license plates, temporary license plates, and vehicles with similar appearances; by introducing time-sensitive keys and dual-path verification mechanisms, data privacy protection is strengthened while ensuring recognition robustness.
[0039] 2. By constructing a complete technology chain of multimodal feature fusion, dynamic key binding, encrypted storage and adaptive verification path switching, and by using key lifecycle management and dual-path verification mechanisms, we have balanced recognition accuracy, system security and operational automation.
[0040] 3. By employing a unified feature embedding space and an end-to-end optimization framework, deep fusion and collaborative enhancement of heterogeneous sensing signals are achieved. This system boasts high scalability and deployment flexibility in engineering implementation, improving its applicability to various scenarios and providing core technical support for smart parking infrastructure. Attached Figure Description
[0041] Figure 1 This is a schematic diagram of the vehicle entry process in a parking lot operation and management method based on artificial intelligence according to the present invention.
[0042] Figure 2 This is a schematic diagram of the vehicle exit process in a parking lot operation and management method based on artificial intelligence according to the present invention. Detailed Implementation
[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0044] Example: Figures 1-2 As shown, the present invention provides a technical solution: a parking lot operation and management method based on artificial intelligence.
[0045] In general vehicle management scenarios, license plates play a crucial role in vehicle identification due to their uniqueness. However, when license plates are discarded as the primary vehicle identification medium, they are easily affected by similar vehicle shapes and colors, such as the same model of vehicle. Therefore, this invention integrates multiple features and encoding / decoding methods to identify vehicle identities without relying on license plate recognition.
[0046] At the physical deployment level, this method sets up multimodal sensing front-ends at the entrance and exit areas of the parking lot, which consist of a visible light camera, a millimeter-wave radar array, and a directional microphone array.
[0047] When a vehicle enters or exits the parking lot, the multimodal perception front-end synchronously triggers data acquisition. The three types of raw data acquired—visual images, millimeter-wave radar point cloud data, and acoustic signals—are transmitted in real time via gigabit Ethernet to edge computing nodes deployed near the entrance and exit.
[0048] At the edge computing node, standardization preprocessing is first performed on the three types of raw data. For the visual image, a two-dimensional Gaussian kernel is first used for Gaussian filtering to suppress high-frequency noise. Then, based on the pre-calibrated camera intrinsic parameter matrix and distortion coefficient vector, perspective distortion is corrected through an affine transformation model. Finally, the corrected image is cropped to a fixed area centered on the vehicle's front bumper and uniformly scaled to a size of H×W pixels, where H and W are constants, forming the standardized visual input I. v .
[0049] For millimeter-wave radar point cloud data, ground point removal is performed first; then, a voxel grid downsampling method is used to set the voxel side length and downsample the remaining point cloud, retaining a representative point in each non-empty voxel, preferably the geometric center of the point in the voxel, and finally outputting a point cloud set with a fixed number of points.
[0050] For acoustic signals, pre-emphasis processing is first performed: the original audio sequence is obtained as s[n], and the pre-emphasized signal s'[n] = s[n] - α s[n-1], where α∈(0,1), and s[-1] is initialized to 0; then, s'[n] is divided into overlapping frames; a Hamming window is added to each frame; after performing a short-time Fourier transform on each frame, the logarithmic energy is calculated through a Mel filter bank, and then the Mel frequency cepstral coefficients are obtained through discrete cosine transform. The main dimension components are taken, such as the first n dimensions, and a first-order difference is added to form a feature vector; after normalizing the feature vectors of all frames, a fixed-length spectral feature vector f is generated through max pooling operation. a .
[0051] After preprocessing, the three types of data are input into their respective feature encoders. Visual Image I v The input convolutional neural network backbone is preferably a ResNet-50 architecture. Its original ImageNet pre-trained weights are loaded, and end-to-end training is performed using a parking lot dataset during the fine-tuning phase. The last fully connected layer is replaced with an output linear projection layer. After forward propagation, the output visual feature vector f is obtained. v .
[0052] Millimeter wave point cloud P r The input is a PointNet-based point cloud encoder, coderR. A preferred approach is to use a three-layer MLP, with each layer followed by batch normalization and ReLU activation; then a global max pooling layer; and finally two fully connected layers to output the radar feature vector f. r .
[0053] acoustic eigenvector f aThe input is a one-dimensional convolutional temporal encoder coderA. Preferably, coderA contains four one-dimensional convolutional layers, each followed by batch normalization and ReLU activation function, with a convolution stride of 1. Finally, the acoustic feature vector f is output through global average pooling and a linear projection layer. a .
[0054] The three feature vectors are concatenated along the channel dimension to form the original feature sequence F of the vehicle, where F = [f v f r f a The comma indicates a vector concatenation operation.
[0055] A learnable normalized embedding mapping is performed on the original vehicle feature sequence F to generate a normalized vehicle feature sequence E. A weight matrix W and a bias term b are defined; W and b are trainable parameters obtained through end-to-end training on a large-scale vehicle dataset before system deployment. The optimization objective is to minimize the angular distance between the feature embeddings of the same vehicle entering and exiting multiple times on a unit hypersphere. The intermediate embedding vector z of the original feature sequence is calculated; further, L2 norm normalization is performed on z to obtain the final vehicle feature sequence E. ;
[0056] This embedding represents a compact feature representation on the unit hypersphere, which facilitates subsequent cosine similarity calculation.
[0057] After the vehicle completes feature extraction in the entrance area, the feature sequence E of the currently entering vehicle is obtained. in ; Traverse the set of feature sequences {E1, E2, ..., E} corresponding to all existing records in the central feature database. M}, where M is the total number of currently registered vehicles; for i ∈ {1, 2, ..., M}, calculate the cosine similarity s. i : ;
[0058] Set a similarity threshold τ sim = 0.85; if there exists any i such that s i >τ sim If the current entering vehicle is highly similar to the parked vehicle corresponding to the i-th record in the database, the key binding process is triggered; otherwise, it is determined to be a newly entering vehicle, and the plaintext storage process is executed, i.e., E... in Write directly to the database and mark it as "no key".
[0059] In the key binding process, the system calls the key derivation function KDF to generate a unique key k. k = KDF(E in,t0), where t0 is the current system timestamp, KDF uses the HKDF-SHA256 algorithm, the salt value is empty, the info field of the information context contains the parking lot ID and the entrance device serial number; the output key k is a 256-bit symmetric key, output in 32-byte binary form.
[0060] Use key k to analyze the feature sequence E of the currently entering vehicle. in Perform AES-256-GCM symmetric encryption to generate the encrypted feature information sequence C. in Preferred key generation methods include: C in = Enc(Ei n, k) = AES-256-GCM.Encrypt(k, nonce, E in The nonce is a 12-byte random number generated by a hardware security module integrated into the edge computing node, which conforms to relevant security standards; the encrypted output bytes contain ciphertext and an authentication tag.
[0061] C in Key identifier kid, timestamp t0, vehicle entry event metadata, such as entrance device ID, collection time, and raw data hash value SHA3-256 (I v P r f a This information is written into the database to form an encrypted record, which is then marked as "with key".
[0062] Meanwhile, the system records the start time t0 of the key's validity and sets the maximum validity period Tmax = 72 hours; within the time interval [t0, t0 + Tmax], key k is a valid key; after exceeding this interval, the system background task automatically scans the database, marks records where t0 + Tmax is less than the current time as "key expired", and prohibits them from participating in subsequent decryption or verification operations.
[0063] Vehicle owners can complete the binding of their vehicle identity and key by scanning a dynamically generated QR code at the self-service terminal at the parking lot entrance, or by accessing the web interface provided by the parking lot cloud platform via a mobile device. For example, vehicle owners can authenticate via SMS verification code; after successful authentication, the system establishes a relationship between the vehicle owner's identity ID and the key identifier kid, and stores it in the user identity database.
[0064] When a vehicle leaves the parking lot exit area, the above multimodal data acquisition, preprocessing, and feature encoding process is repeated to generate the feature sequence E of the exiting vehicle. out .
[0065] The system retrieves all records from the central feature database, dividing them into plaintext and encrypted record sets; first, it calculates E.out Cosine similarity with each plaintext feature sequence Ej: ;
[0066] Filter out those that satisfy q j,1 A candidate index set S with τsim = 0.85; if S is non-empty and the valid key k associated with the exiting vehicle is... d If the key exists, meaning a valid key is found through the vehicle owner's identity and the corresponding key has not expired, then dual-path similarity verification will be performed.
[0067] First approach: Directly use plaintext similarity q j,1 As the primary similarity metric.
[0068] Second path: Using a valid key k d For E out Encryption is performed using the same encryption method as upon entry, resulting in ciphertext C. out ;
[0069] Then, an approximate nearest neighbor search algorithm based on Hamming distance is used to match all encrypted feature information sequences in the database. Specifically, the ciphertext features are treated as binary hash codes, and their Hamming distances with each ciphertext entry in the database are calculated. The distance values are then normalized to the [0,1] interval through a linear mapping, and the resulting value is denoted as the second similarity q. j,2 ;
[0070] The final similarity score h is the maximum of the two scores: ;
[0071] Set the verification threshold τ verify = 0.90; Returns the maximum h and h > τ. verify The vehicle's identity information, including entry time, parking duration, and fees.
[0072] If the departing vehicle is not bound to a key, or although it is bound to a key but the candidate set S is empty (i.e., there is no plaintext record matching), then single-path similarity verification is performed:
[0073] If a valid key k exists d If so, only the second path will be executed, and q will be calculated. j,2 and the single-path similarity threshold τ single =0.88 comparison;
[0074] If no valid key is available, only the first path will be executed, and q will be calculated. j,1 and τ single Compare.
[0075] If the similarity of any path is greater than τ singleIf the match is successful, the corresponding vehicle identity will be returned; otherwise, the matching will fail, triggering a manual processing procedure and pushing an alarm message to the exit gate terminal, including E. out The first few dimensions of the data, the original data hash, and suggested verification instructions.
[0076] If the key associated with a vehicle leaving the parking lot has been determined by the system to be an expired key (i.e., the current time t>t0 + Tmax), then regardless of the similarity calculation result, the system will send an abnormal prompt message to the parking lot management personnel terminal. The message includes a vehicle feature summary, entry and exit timestamps, key expiration status, and original data hash value for manual verification.
[0077] The technical effects of the present invention are illustrated below through specific embodiments and comparative examples.
[0078] The following concrete example further illustrates the key binding process:
[0079] A vehicle entered the parking lot at a certain time on a certain day. After the multimodal sensing unit collected the data, it underwent preprocessing and encoding to obtain:
[0080] f v = [0.32, -0.15, ..., 0.41];
[0081] f r = [0.28, 0.09, ..., -0.22];
[0082] f a = [-0.11, 0.33, ..., 0.17];
[0083] By concatenation, we get F = [f v f r f a ].
[0084] Let the weight matrix W be the parameters obtained during training, and the bias b be the zero vector;
[0085] After normalizing the original feature sequence F, E in = [0.0641, -0.0427, ..., 0.0507].
[0086] There are already two records in the database:
[0087] E1 = [0.0635, -0.0431, ..., 0.0512], showing a high degree of similarity, suggesting they may be the same car model;
[0088] E2 = [0.1200, 0.0100, ..., -0.0300], low similarity does not affect vehicle model identification;
[0089] Similarity calculation: After calculation, s1≈ 0.872, s2≈ 0.315;
[0090] Since s1 = 0.872 > τ sim = 0.85, triggering key binding.
[0091] System call KDF, with E in Using t0 as input, a 256-bit key k is generated.
[0092] Encrypt Ein using k to obtain Cin.
[0093] Write the encrypted ciphertext, containing the key k, start time, and maximum validity period, into a database record and mark it as "with key":
[0094] Car owners can complete the identity binding process by scanning a QR code using a mobile app.
[0095] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A parking lot operation and management method based on artificial intelligence, characterized in that: The methods include: When a vehicle enters the parking lot entrance area, vehicle feature data is collected. After preprocessing, the vehicle feature data is input into an encoding model for encoding to obtain the feature sequence of the currently entering vehicle. The vehicle characteristic data includes: The vehicle's visual image, millimeter-wave radar point cloud data, and acoustic signal are preprocessed and feature extracted respectively to generate corresponding visual feature vectors, radar point cloud feature vectors, and acoustic spectrum feature vectors, which are then spliced together to form the vehicle's entry feature sequence. The feature sequences corresponding to the vehicles already parked in the parking lot are formed into a feature sequence set. If there is a feature sequence in the feature sequence set that has a similarity greater than the similarity threshold with the feature sequence of the currently entering vehicle, the key binding process is triggered to generate a key with a valid duration. The feature information sequence generated by encrypting the feature sequence of the currently entering vehicle is then stored in the database. When a vehicle leaves the parking lot exit area, the vehicle feature data collection and encoding operations are repeated to generate a feature sequence of the exiting vehicle. Obtain the feature sequence of the vehicle leaving the site, and search the database for feature sequences. If there is a feature sequence in the database that has a similarity greater than the similarity threshold with the feature sequence of the vehicle leaving the site, and the owner of the vehicle leaving the site has a key bound to it, then use dual-path similarity verification; otherwise, use single-path similarity verification. The dual-path similarity verification includes: when obtaining the valid key associated with the exiting vehicle, encrypting the feature sequence of the exiting vehicle using the valid key to generate an exiting feature information sequence; The feature sequences in the database that have a similarity greater than a similarity threshold with the feature sequences of the vehicles leaving the market are used as reference sequences, and the similarity between the feature sequences of the vehicles leaving the market and the reference sequences is calculated as the first similarity. Obtain the feature information sequence with the highest similarity to the exit feature information sequence from the database, and record the similarity between the feature information sequence and the exit feature information sequence as the second similarity. The maximum value between the first similarity score and the second similarity score is obtained as the final similarity score, and the vehicle identity information corresponding to the final similarity score is returned. The single-path similarity verification includes: When the owner of the vehicle leaving the vehicle has not bound a key, only the first similarity of the feature sequence of the vehicle leaving the vehicle is calculated; When there is no feature sequence in the database that has a similarity greater than the similarity threshold with the feature sequence of the vehicle leaving the site, and the owner of the vehicle leaving the site has a binding key, only the second similarity score is returned. Set a single-path similarity threshold. During single-path similarity verification, if the first or second similarity is greater than the single-path similarity threshold, return the corresponding vehicle identity information; otherwise, trigger the manual processing procedure. The vehicle identity information obtained through single-path or dual-path similarity verification is fed back to the parking lot's automatic billing system; When the key associated with a vehicle leaving the parking lot expires, a notification message is sent to the parking lot management personnel.
2. The parking lot operation and management method based on artificial intelligence according to claim 1, characterized in that: Methods for matching feature sequences in a set of feature sequences include: When the feature sequence of the currently entering vehicle is obtained, the feature sequences stored in the database are traversed. When there is a feature sequence in the database with a similarity greater than a threshold to the feature sequence of the currently entering vehicle, the key binding process is triggered. The feature sequence of the currently entering vehicle is encrypted using a key to generate a feature information sequence, and the feature information sequence is saved to the database. If there is no feature sequence in the database that has a similarity greater than the threshold with the feature sequence of the currently entering vehicle, then the feature sequence of the currently entering vehicle will be saved to the database.
3. The parking lot operation and management method based on artificial intelligence according to claim 2, characterized in that: The key binding process includes: The key is bound to the vehicle owner's identity through a terminal device connected to the parking management server. The terminal device includes a gate at the parking lot entrance, the vehicle owner's handheld mobile device, and the vehicle's onboard mobile device. The key is set with a maximum validity period. The moment when the key is bound to the vehicle owner's identity is recorded as the valid start time of the key. The key is valid within the maximum validity period, starting from the valid start time. Otherwise, the key is an expired key.
4. An artificial intelligence-based parking lot operation management system, used to execute the artificial intelligence-based parking lot operation management method described in claim 1, characterized in that: The system includes: The vehicle feature acquisition module is used to simultaneously acquire visual images, millimeter-wave radar point cloud data, and acoustic signals of vehicles entering or leaving the vehicle. The feature preprocessing module is used to perform Gaussian filtering and affine transformation on the visual image and then crop it to a uniform size; to remove ground reflection points and static background points from the millimeter-wave radar point cloud data and then perform voxel grid downsampling to normalize the number of points; and to pre-emphasize, frame and window the acoustic signal and calculate the Mel frequency cepstral coefficients to generate a spectral feature vector. The feature encoding module is used to encode the three types of preprocessed data to obtain three types of feature vectors, then concatenate the three types of feature vectors to form the original feature sequence of the vehicle, and then normalize and map the original feature sequence to obtain the vehicle feature sequence. The key management module is used to generate a key with a valid duration limit based on the vehicle owner's identity authentication during the entry stage, and associate the key with the corresponding vehicle's feature sequence within the key's validity period; An information storage database is used to store plaintext feature sequences or feature information sequences encrypted by the key; The similarity verification module is used to adaptively select a single-path or dual-path similarity verification strategy during the exit stage based on whether there is a valid key and whether there are candidate matching records that meet the preset similarity threshold.
5. The parking lot operation management system based on artificial intelligence according to claim 4, characterized in that: The similarity verification module includes a single-path similarity verification unit and a dual-path similarity verification unit; The dual-path similarity verification unit includes: a first path that calculates the similarity between the exit vehicle feature sequence and the candidate plaintext feature sequence in the database; a second path that encrypts the exit vehicle feature sequence using a valid key, calculates the Hamming distance between the encrypted feature sequence and the stored encrypted feature information sequence in the ciphertext space, and normalizes it to a second similarity; and finally, the maximum value of the two similarities is taken. The single-path similarity verification unit includes: identity determination based on cosine similarity of plaintext feature sequences or normalized similarity based solely on ciphertext matching.
6. The parking lot operation management system based on artificial intelligence according to claim 4, characterized in that: The similarity verification module also includes: If the key associated with a vehicle leaving the parking lot has been determined by the system to be an expired key, an abnormal prompt message will be sent to the parking lot management personnel terminal regardless of the similarity calculation result. The message will include a summary of vehicle characteristics, entry and exit timestamps, and key expiration status.
7. The parking lot operation management system based on artificial intelligence according to claim 4, characterized in that: The similarity verification module interfaces with the parking lot automatic billing system through a standard API or network interface. After receiving the valid vehicle identity information returned by the similarity verification module, it automatically calls the billing instruction to complete the parking fee calculation.