Expressway accident vehicle positioning identification method based on quantum and unmanned aerial vehicle technology

Abstract

The invention discloses an expressway accident vehicle positioning identification method based on quantum and unmanned aerial vehicle technologies, and relates to the technical field of quantum technology deep learning application, and the method comprises the steps: constructing a PPP-RTK model, introducing a TTTrack algorithm, and collecting expressway vehicle image information through an unmanned aerial vehicle based on the PPP-RTK model and the TTTrack algorithm; preprocessing the image; performing feature extraction on the image through a CNN (convolutional neural network) and an Xception classification model in sequence; judging whether the vehicle contained in the image is an accident vehicle or not according to the image; if the vehicle is in the accident, quitting the picture, and if the vehicle is in the accident, preparing a quantum image through an ONEQR model, processing the quantum image through a target detection algorithm, processing the image, and extracting vehicle features including the vehicle in the accident; and establishing a database according to the extracted vehicle features. According to the invention, the efficiency and precision of image processing are obviously improved, and the accuracy of image recognition is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of quantum technology deep learning applications, and in particular to a method for positioning and identifying accident vehicles on highways based on quantum and unmanned aerial vehicle technologies. Background Art

[0002] With the rapid growth of highway traffic flow, the problems of highway traffic accident monitoring and emergency response need to be solved urgently. Especially in extreme weather, when traffic problems occur on highways, the efficiency and accuracy of vehicle information identification are relatively low. Traditional unmanned aerial vehicle technologies face many challenges in highway scenarios: 1) Insufficient positioning accuracy: The GPS-based unmanned aerial vehicle positioning technology is affected by terrain occlusion, multipath effects, etc., and is prone to centimeter-level deviations in complex environments, making it difficult to meet the high-precision positioning requirements; 2) Low image processing efficiency: For example, traditional convolutional neural network algorithms have problems of large computational resource occupancy and slow processing speed in real-time detection. Especially in bad weather such as rainy and foggy days, the image noise interference significantly reduces the recognition accuracy; 3) Data storage and transmission bottlenecks: The high-definition image data volume is huge, and traditional storage methods occupy a large amount of space. Moreover, limited by the on-board computing power of unmanned aerial vehicles, it is difficult to achieve efficient processing and real-time transmission. Summary of the Invention

[0003] In order to overcome the above problems existing in the prior art, the present invention proposes a method for positioning and identifying accident vehicles on highways based on quantum and unmanned aerial vehicle technologies.

[0004] The technical solution adopted by the present invention to solve its technical problems is: A method for positioning and identifying accident vehicles on highways based on quantum and unmanned aerial vehicle technologies, including the following steps: Step 1, construct a PPP-RTK model, introduce the TTTrack algorithm, and collect highway vehicle image information by an unmanned aerial vehicle based on the PPP-RTK model and the TTTrack algorithm; Step 2, preprocess the image obtained in Step 1; Step 3, determine whether the vehicle contained in the image obtained in Step 2 is an accident vehicle; if it is an accident vehicle, then exit the picture, if it is an accident vehicle, then enter Step 4; Step 4, sequentially perform feature extraction on the image obtained in Step 3 through a CNN convolutional neural network and an Xception classification model; Step 5, prepare a quantum image through the ONEQR model, process it through a target detection algorithm, and perform image edge detection, smoothing processing, and threshold segmentation reconstruction on the image to extract the vehicle features in the image containing the accident vehicle; Step 6, establish a database according to the vehicle features extracted in Step 5.

[0005] The above-mentioned method for positioning and identifying accident vehicles on highways based on quantum and drone technologies, where the PPP-RTK model includes a PPP-RTK function model, a PPP-RTK regional troposphere model, and a PPP-RTK regional ionosphere model; the PPP-RTK regional troposphere model adopts the HIQM4 model.

[0006] The above-mentioned method for positioning and identifying accident vehicles on highways based on quantum and drone technologies, where the PPP-RTK function model is: ; ; where are the receiver, satellite, frequency, and epoch numbers respectively; is the tropospheric wet delay projection function; is the satellite phase bias; is the ratio of the ionosphere at other frequencies to the ionosphere at the first frequency; is the receiver phase bias; is the receiver clock error solved by the new observation equation, is the satellite clock error, is the zenith tropospheric wet delay solved by the new observation equation, is the ionospheric slant delay at the first frequency, is the receiver code bias, is the satellite code bias, is the ambiguity of non-first stations, is the receiver phase bias, is the undifferenced and uncombined code after deducting the precise satellite-to-ground distance and various model errors, is the residual of the phase observation value after deducting the precise satellite-to-ground distance and various model errors.

[0007] The above-mentioned method for positioning and identifying accident vehicles on highways based on quantum and drone technologies, where the PPP-RTK regional ionosphere model is expressed as: ; where are the coefficients of the ionospheric slant delay polynomial, is the longitude of the reference station, is the latitude of the reference station, is the difference between the longitude of the reference station and the longitude of the center of the survey area, is the difference between the latitude of the reference station and the latitude of the center of the survey area.

[0008] The above-mentioned method for positioning and identifying accident vehicles on highways based on quantum and drone technologies, wherein the TTTrack algorithm includes a twin backbone network, an adaptive spatio-temporal information extraction module, and a spatio-temporal context mapping module. The twin backbone network is used to extract features from the template image and the search image. The adaptive spatio-temporal information extraction module adopts a dual-template mechanism of a reference template and a dynamic template. The spatio-temporal context mapping module is used to capture the spatio-temporal correlation between consecutive frames.

[0009] The above-mentioned method for positioning and identifying accident vehicles on highways based on quantum and drone technologies, wherein the preparation of the quantum image in step 5 specifically includes: applying the operator to to obtain the intermediate state ; storing the gray level and real coordinate information corresponding to each position in ; applying operators to to complete the preparation of all pixels.

[0010] The above-mentioned method for positioning and identifying accident vehicles on highways based on quantum and drone technologies, wherein the threshold segmentation and reconstruction in step 5 specifically include: obtaining the number of gray level images, calculating the probability of the gray level image in the total number of pixels, and initializing the gray level threshold; dividing the image into two categories of background and foreground, and calculating the occurrence probability and probability average of the gray levels of different image types based on the Bayesian optimization algorithm; calculating the mean value of the gray levels of the original image, calculating the between-class variance of the image, and determining the optimal segmentation threshold by calculating the maximum between-class variance.

[0011] The above-mentioned method for positioning and identifying accident vehicles on highways based on quantum and drone technologies, wherein the image edge detection in step 5 is specifically: adopting the quantum LoG edge detection algorithm. According to the QPIE model, the image Q can be expressed as a pure quantum state , encoding the LoG operator template into a quantum system containing 5 qubits to obtain the quantum state of the LoG operator; performing a tensor product operation on the quantum states and to obtain , and finally measuring to obtain the quantum image edge through the probability amplitude permutation operator U and quantum circuit processing.

[0012] The beneficial effects of the present invention are as follows: 1) The computing efficiency is greatly improved. Quantum computing has the ability of parallel computing and can process a large amount of data simultaneously, significantly improving the speed of image feature extraction, target detection, and classification. For example, quantum machine learning algorithms (such as quantum support vector machines, quantum neural networks) can theoretically achieve exponential acceleration, greatly shortening the image processing time and meeting the real-time requirements in highway scenarios.

[0013] 2) Enhanced data collection capabilities. Drones can collect image data from multiple angles and all directions from the air, dynamically adjust their flight paths, and cover wider areas, avoiding the blind spots of fixed cameras. This provides a more comprehensive and flexible monitoring perspective, enhancing the coverage and accuracy of data collection.

[0014] 3) Improved real-time and dynamic response capabilities. Combining the efficient processing power of quantum computing with the real-time data collection capabilities of drones allows some computing tasks to be performed on edge devices or drones, reducing data transmission latency. This significantly improves the system's real-time and dynamic response capabilities, making it suitable for scenarios requiring rapid response, such as traffic accident detection and violation identification.

[0015] 4) Improved recognition accuracy. Quantum machine learning algorithms have advantages in processing high-dimensional data and nonlinear problems, enabling better extraction of image features and improving recognition accuracy. Drones can capture images from different angles, reducing occlusion issues. Combined with multi-view data fusion technology, recognition accuracy is further improved, maintaining high-precision recognition even in complex environments.

[0016] 5) Enhanced application flexibility and scalability. Drones can be deployed and flight paths adjusted at any time to adapt to different monitoring needs, providing a high degree of flexibility. The versatility and scalability of quantum computing provide more possibilities for future algorithm upgrades. The system can be flexibly adjusted according to actual needs and is applicable to a variety of scenarios (such as traffic flow monitoring, accident detection, and violation identification).

[0017] 6) Energy and cost optimization. The high efficiency of quantum computing can reduce computing resource consumption, and the flexible deployment of drones can reduce the installation and maintenance costs of fixed equipment. This can reduce energy consumption and costs in the long term, improving economic benefits.

[0018] 7) Support for large-scale data processing. Quantum computing can efficiently process large amounts of data, making it suitable for high-density vehicle monitoring on highways. It supports real-time processing of massive amounts of data, providing more powerful data support for intelligent traffic management.

[0019] 8) Future Compatibility and Technological Foresight: Quantum computing and drone technology are both cutting-edge technologies with a high degree of technological foresight, capable of providing enhanced technical support for future intelligent transportation systems. They lay the foundation for future technological upgrades and application expansion, and possess long-term development potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic flow chart of the present invention; Figure 2 It is the overall framework of the TTTrack tracker of the present invention; Figure 3 This is the CNN network model implementation architecture of the present invention; Figure 4 It is a schematic diagram of the implementation structure of the convolutional layer of the present invention; Figure 5 It is a flow chart of multi-threshold image adaptive segmentation of the present invention; Figure 6 It is the network structure of YOLOv5 of the present invention; Figure 7 It is the feature fusion network of the YOLOv5 algorithm of the present invention; Figure 8 It is the design diagram of the BLENet network of the present invention. Detailed implementation manners

[0021] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific implementation manners.

[0022] As Figure 1 shown, this embodiment discloses a method for positioning and identifying accident vehicles on highways based on quantum and drone technologies. In this embodiment, research is carried out successively through theoretical research, theoretical analysis, quantum algorithm analysis, quantum algorithm design, and optimization model methods, and model simulation calculations are carried out with the aid of the Python language. Using the Python language as a framework, data images are compiled into the OpenQASM quantum programming language, which serves as an intermediate representation carrier for high-level compilers to communicate with quantum hardware, and is uploaded to the quantum cloud server for operation to complete operations such as defining quantum circuits based on ONEQR quantum image expressions, generating quantum circuits, defining quantum circuit behaviors, and printing quantum circuits, and realizing the preparation and reading process of quantum grayscale images.

[0023] First, use drone recognition technology and quantum convolutional neural network (CNN) to solve the problems of low recognition efficiency and low accuracy of vehicle information in traffic problems on highways under extreme weather conditions, and use the precise point positioning - real-time kinematic positioning technology of the Beidou satellite navigation system to provide strong signal and high-precision positioning services for detection drones.

[0024] Secondly, introduce the TTTrack algorithm to enhance the target tracking ability of drones to adapt to complex drone tracking scenarios.

[0025] Finally, taking drone technology as a carrier, use the YOLOv5 algorithm to make up for the shortcomings of deep learning technology in images to increase the operation accuracy and improve the operation efficiency. At the same time, bring the ONEQR quantum model into it to optimize the convolutional neural network, and use the LoG algorithm for image edge detection and smooth processing of the image to make the accuracy of the image higher.

[0026] The technical solution of this embodiment is implemented as follows: 1) Introduce precise point positioning - real-time kinematic (PPP-RTK) technology. Traditional UAV positioning usually uses GPS positioning, but in some areas with complex terrain or regions where the baseline of the reference station is too long, the service accuracy cannot meet the requirements of UAV inspection, and a large amount of manpower and material resources are required for manual inspection. To solve this problem, a Beidou-based PPP-RTK technology is introduced to provide high-precision positioning services for UAVs to quickly locate the disaster area.

[0027] 2) Introduce the TTTrack algorithm. In this embodiment, the SiamFC++ (AlexNet) algorithm is used as the baseline, and a time-domain Siamese network tracking algorithm called TTTrack is proposed by combining Transformer. Different from traditional single-template tracking algorithms, TTTrack introduces a dual-template mechanism. Among them, the reference template is determined by the initial frame with stable response values, while the dynamic template is updated by capturing the spatio-temporal information of historical frames through Transformer. The TTTrack algorithm adaptively uses the spatio-temporal information of historical frames to update the template instead of storing all historical information, ensuring the real-time performance of the algorithm.

[0028] 3) Prepare the quantum primary state. In this embodiment, a quantum bit array is used to prepare the quantum state of an image.

[0029] 4) Prepare the quantum image. When using a fully connected neural network to process large-size images, there are three very obvious disadvantages: unfolding the image into a vector will lose spatial information; having too many parameters leads to low efficiency and difficult training; a large number of parameters will quickly cause the network to overfit. Convolutional neural networks can well solve the above three problems. To make up for the deficiency in the accuracy of the digital image obtained by quantizing the image with the ONEQR quantum method, the YOLOv5 algorithm is used to increase the budget accuracy of the digital image. Taking advantage of the characteristics of the LOG operator, which has strong anti-noise ability, high edge positioning accuracy, and good continuity, this method is used to perform image edge detection and smoothing processing on the image.

[0030] 5) Image processing. Image processing methods such as image enhancement, smoothing filtering, and threshold segmentation are used on the characteristic images of high-speed vehicles for high-speed vehicle detection and recognition, to solve the problem of inaccurate vehicle detection in rainy and cloudy weather, make up for the shortcomings of deep learning technology, and increase the accuracy of determining the accident location.

[0031] Specifically, it includes the following steps: Step 1, construct a PPP-RTK model, introduce the TTTrack algorithm, and collect highway vehicle image information through a UAV based on the PPP-RTK model and the TTTrack algorithm.

[0032] The establishment of the PPP-RTK model is divided into several sub-parts, including the PPP-RTK function model, the regional troposphere model, and the regional ionosphere model.

[0033] (1)PPP-RTK function model PPP-RTK technology realizes high-precision positioning by integrating the advantages of precise point positioning (PPP) and real-time kinematic positioning (RTK) and using the undifferenced and uncombined GNSS observation equations. Its core observation equation is: ; In the above formula: are the receiver number (used to identify different receivers), satellite number (used to identify different satellites, such as a certain satellite in satellite systems like GPS and Beidou), frequency number (used to distinguish different frequencies of signals), and epoch number (i.e., a time point of the observation data or a certain moment in the time series), respectively; is the pseudorange observation value; is the phase observation value; is the distance from the satellite to the station; is the receiver clock error; is the satellite clock error; is the zenith tropospheric wet delay; is the tropospheric wet delay projection function; is the I-th frequency; is the ratio of the ionosphere of other frequencies to the ionosphere of the first frequency; is the wavelength of the carrier phase of the J-th frequency; is the speed of light in a vacuum; is the slant delay of the ionosphere of the first frequency; is the receiver code bias; is the satellite code bias; is the receiver phase bias; is the unmodeled error and random noise of the pseudorange; is the unmodeled error and random noise of the phase.

[0034] To solve the problem of linear correlation between parameters, the S-basis rank deficiency elimination theory is introduced, and a full-rank equation is constructed through parameter recombination. The final PPP-RTK function model can be expressed as: ; .

[0035] (2)PPP-RTK regional troposphere model The troposphere model adopts the HIQM4 model, which simultaneously considers the different gradients of the troposphere in the N direction and the E direction. This model is expressed as: ; In the formula, is the zenith tropospheric wet delay polynomial coefficient, where q represents the polynomial coefficient index, used to describe the spatial distribution of the zenith tropospheric wet delay, q = ; and are the longitude and latitude differences between the reference station and the central longitude and latitude of the survey area respectively; ; is the geodetic height of the reference station .

[0036] (3)PPP-RTK regional ionospheric model Based on the difference between the longitude and latitude of the reference station ( , ) and the central longitude and latitude of the survey area , the two-dimensional second-order Taylor expansion is performed on the single-star ionospheric slant delay. This model is expressed as: ; ; In the formula are the ionospheric slant delay polynomial coefficients, where represents the term independent of the reference station, represents the term related to the longitude and latitude of the reference station.

[0037] To further enhance the target tracking ability of the UAV in rainy weather, the TTTrack algorithm is introduced. The TTTrack algorithm aims to improve the target tracking ability of the UAV in complex scenarios (such as rainy weather), and the overall framework is as Figure 2 shown. Its core framework includes a Siamese backbone network, an adaptive spatio-temporal information extraction module, and a spatio-temporal context mapping module. The Siamese backbone network is used to extract features from the template image and the search image; the adaptive spatio-temporal information extraction module adopts a dual-template mechanism of a reference template and a dynamic template; the spatio-temporal context mapping module is used to capture the spatio-temporal correlation between consecutive frames.

[0038] (1)Siamese backbone network 1) It is constructed based on the lightweight AlexNet and is used to extract features from the template image (127×127×3) and the search image (289×289×3).

[0039] 2) The Siamese structure with shared parameters ensures the consistency of feature extraction while taking into account the real-time requirements.

[0040] (2)Adaptive spatio-temporal information extraction module 1) The Transformer technology is introduced to dynamically update the template according to the spatio-temporal information of historical frames.

[0041] 2) Adopt a dual-template mechanism: a reference template (determined by the initial frame) and a dynamic template (adaptively generated through historical frames), enhancing the adaptability to target appearance changes.

[0042] (3) Spatiotemporal context mapping module The encoder-decoder structure fuses the response maps of the reference template and the dynamic template (generated through cross-correlation operations), capturing the spatiotemporal correlation between consecutive frames.

[0043] It is expressed by the formula: ; where is the response map of the reference template, is the response map of the dynamic template, is the real-time search feature map, is the reference template, is the dynamic template, is the sum, is the corresponding deviation.

[0044] Step 2, preprocess the image obtained in Step 1.

[0045] The image preprocessing steps include: 1) Image grayscale processing: Convert the RGB color image to a grayscale image, reducing the image complexity and highlighting the brightness features.

[0046] 2) Filtering and noise reduction: The actually acquired images are often affected by noise. Filtering techniques are used to smooth the images and reduce noise interference.

[0047] 3) Image binarization processing: Convert the grayscale image to a binary image, enhancing the contrast, separating the foreground and background, and facilitating subsequent morphological operations.

[0048] 4) Morphological noise reduction and enhancement: After binarization processing, isolated noise points or small regions that may appear in the image are removed through morphological operations such as opening and closing operations, while retaining the overall shape of the target region.

[0049] In the actual application environment of drones, the lighting conditions change gradually. Drawing on the adaptive selection mechanism of SK-Net and the modal feature-level fusion idea of UA-CMDet, the enhanced image and the original image are learned simultaneously, and selection fusion is performed at the feature level to achieve the effect of low-illumination image enhancement. The specific design of the network BLENet is as Figure 8 shown.

[0050] The input is divided into two parts. In the Normal part, preliminary feature extraction is performed through two convolutions. In the Enhance part, the image is subjected to improved Laplacian filtering for edge enhancement, and then enters the IAT to obtain the enhanced image, and then two convolutional layers are used to extract preliminary features. In the Fuse part, first, element-wise summation is performed: ; where is the output feature, and are the features of the original image and the enhanced features respectively.

[0051] Then, channel data statistics are performed on for selecting the feature maps of the original image and the enhanced image. This step uses the method of global average pooling to reduce the dimension in the spatial dimension and compress the information within the channel into a one-dimensional vector : ; where represents the information within a single channel of the output feature , u and z represent the coordinate values of individual pixels within the channel, and H and W are the length and width of the feature map respectively.

[0052] Then, a fully connected layer is used to transform the vector to create a vector corresponding to the number of channels as the representation of the activation degree of the corresponding channel: ; where represents normalization processing, represents the ReLU activation function, represents processing the feature through the fully connected layer.

[0053] In the Select part, by performing a softmax operation on each position of and the original feature , the weight of each channel is obtained, and the process is similar to that of Fuse. By calculating the original feature and with the channel weights, its output is: ; By selecting the enhanced image feature and the original image feature respectively according to the channel activation degree, the external illumination conditions can be ignored, and the model will adaptively tend to the group of features with better performance among the two groups of feature values, thereby improving the detection performance for low-illumination images.

[0054] To make the image denoising effect more ideal and ensure that the denoised image has high quality, it is necessary to ensure the sparsity, retention, and separability of the image data.

[0055] Under normal circumstances, digital images In , two directions are completely independent, and the represented content is also different. Define the image at moment as , and obtain the time partial derivative between and through the Grünwald-Letnikov theory in the fractional-order PDE: ; In the formula, represents the order of the partial derivative, represents the time coefficient, , both represent the image level.

[0056] The image after denoising processing is expressed by the formula:

[0057] Step 3: Determine whether the vehicle included in the image obtained in Step 2 is an accident vehicle. If it is an accident vehicle, exit the picture. If it is an accident vehicle, enter Step 4.

[0058] Step 4: Perform feature extraction on the image obtained in Step 3 through the CNN convolutional neural network and the Xception classification model in sequence.

[0059] The CNN convolutional neural network includes a convolutional layer, a ReLU layer, a pooling layer, an Affine layer, and a Softmax layer. The implementation architecture is as Figure 3 shown, specifically: Convolutional layer: After the previous image acquisition and preprocessing, it is input into the convolutional layer. Subsequently, the input elements are subjected to a convolutional operation and input into the activation function to obtain the output feature map of this layer, as Figure 4 shown. The convolutional operation is as follows: ; Among them, : The elements covered by the convolutional layer feature map by convolutions; : The elements in the convolutional kernel of this layer; : Bias; : Activation function. e represents the previous layer (the The e-th input channel of the (-1)-th layer, g represents the g-th output channel of the current layer (the l-th layer) (i.e., the feature map generated by the g-th convolutional kernel), and l represents the layer index of the neural network (the l-th layer).

[0060] Extract image features through convolution operation. The number of convolutional kernels is 30, the size is 5×5, the stride is 1, and the padding is 0.

[0061] ReLU layer: An activation function layer, suitable for non-linear mapping learning, reducing the interdependence between parameters and avoiding overfitting.

[0062] Pooling layer: Reduce the image size, reduce the amount of computation, and extract features through methods such as max pooling. And output through the activation function after bias processing, as shown in the following formula: ; : The output obtained after downsampling the convolutional layer image block; : The output of this layer, which is also the input element of the next layer; : Activation function. Represents the serial number of the feature map or image block in the current layer Affine layer: A fully connected layer that performs weighted operations and bias operations. The operation in the system design is: np.dot(X, W)+B.

[0063] Softmax layer: The output layer, used for image classification, adjusting the output value to between 0 and 1.

[0064] The Xception network structure uses depthwise separable convolution for feature extraction. To improve the classification accuracy, the improved MultiXception model combines the multi-scale module in Res2Net to enhance the richness of feature information. The improvement is achieved through the following steps: 1) Feature map segmentation: The input feature map is evenly segmented into 4 parts in the channel dimension.

[0065] 2) Multi-scale feature extraction: Use different numbers of convolutional kernels to extract features from the segmented feature maps, use 3×3 Depthwise convolution instead of standard convolution, and do not perform ReLU operation after convolution.

[0066] 3) Feature fusion: Concatenate the feature maps of different scales in the channel dimension to fuse multi-scale feature information.

[0067] 4) Channel adjustment: Use 1×1 pointwise convolution to adjust the number of channels of the output feature map to complete the transmission of multi-scale features. The above process can be expressed as: ; Where \(i\) represents the serial number of the feature map segmentation in formulas (1)-(3), and \((c, d)\) in formula (4) represents the spatial position coordinates of the feature map. \(X\in R^{(N, H, W, C)}\) represents the original input feature map. \(\in R^{(N, H, W, C / 4)}\), \(k = 1, 2, 3, 4\) represents the feature maps obtained after average segmentation; in formula (2), \(DW\) is a \(3\times3\) depth-wise convolution. When \(k = 1\), the feature map directly obtains the output without passing through the depth convolution operation. , when \(k>2\), the input of each \(3\times3\) convolution is composed of the current feature map and the previous output feature map . In formula (3), \(Y\in R^{(N, H, W, C)}\) represents the feature map obtained after the multi-scale feature extraction step. \(\in R^{(N, H, W, C / 4)}\), \(k = 1, 2, 3, 4\) represents the feature map obtained after the concat operation in the channel dimension; in formula (4), represents the value at the position \((\) , , ) in the \(c\)-th channel of the output feature map. represents the \(c\)-th channel of the \(1\times1\) convolution kernel. , \((\) , , ) is the value at the position \((\) , , ) in the \(c\)-th channel of the feature map obtained after multi-scale feature extraction. \(M\) is the number of channels of the input feature map.

[0068] Through the multi-scale depthwise separable convolution, the model can better extract the detailed information in the image and improve the classification accuracy.

[0069] Step 5, prepare the quantum image through the ONEQR model, process it through the object detection algorithm, and perform image edge detection, smoothing, threshold segmentation and reconstruction on the image to extract the vehicle features in the image containing the accident vehicle.

[0070] The ONEQR model is used for quantum image processing, and the preparation of this model significantly reduces the time complexity of quantum image preparation.

[0071] The quantum state representation of the ONEQR model is as follows: For a grayscale image with a size of and a grayscale range of , the ONEQR quantum representation is as shown below: ; Among them, , , 𝑋 and 𝑌 respectively represent the pixel position information in the vertical and horizontal directions, and are composed of 2 qubits; At the same time, it stores the pixel grayscale and the real position information , and is composed of 2n + 1 qubits. One of the qubits is used to distinguish and ; when the single qubit is in state, represents the grayscale information; when the single qubit is in state, represents the real position information.

[0072] The preparation process of ONEQR is divided into four steps: 1) Apply the operator to to obtain the intermediate state , as shown in the following formula: ; Among them The main role of is to mark the computational ground state and is used to construct the superposition of quantum states.

[0073] 2) Store the grayscale and real coordinate information corresponding to each position in . This process requires transforming the corresponding qubits according to the grayscale and position information. As shown in the following formula (16): ; Among them, consists of a sequence of q qubit ground states, corresponding to the grayscale value at position i, represents the specific value (0 or 1) of each single qubit in this sequence of q qubit ground states, consists of a sequence of 2n qubit ground states, corresponding to the real coordinate value, then represents the specific value of each single qubit in this sequence of 2n qubit ground states.

[0074] The operators and can be obtained from the following formula, as shown below, where : ; ; 3) Apply the operator to to obtain the quantum state as shown in the following formula: ; 4) indivual Acts on , the preparation of all pixels can be completed, and the prepared quantum state is shown as follows: ; The Bayesian optimization algorithm is used to segment and reconstruct the image, so that the algorithm can make the image segmentation result tend to the optimal solution.

[0075] Set the grayscale image of the multi-threshold image to , the total number of pixels is , the grayscale histogram is , the initial threshold of the image is , the image is divided into background image by setting the initial threshold and foreground image , the optimized Bayesian algorithm is used to calculate the gray level probability of different types of images. The results are shown in the following formula: ; The probability that the number of pixels in the grayscale of the image to be segmented accounts for the total number of pixels in the image is , the gray level probability of the background image is , the gray level probability of the foreground image is .

[0076] Based on the above calculation results, the average probability of grayscale occurrence of the image is calculated. The process is shown in the following formula: ; Where, the average probability of gray level occurrence of the background image is , the average occurrence probability of the gray level of the foreground image is , the weighted coefficient is , the gray level is , the first formula Represents the grayscale value, ranging from 0 to threshold T (background area), the second formula Also grayscale, but ranging from 1 to T-1 (foreground area).

[0077] Get the overall grayscale image mean of the image to be segmented and introduce evaluation indicators Identify the pros and cons of each threshold in the image, determine the inter-class variance and overall grayscale variance of the image to be segmented, and the results are shown in the following formula: ; ; The inter-class variance of the image to be segmented is , the grayscale population variance is , the overall grayscale mean of the image is .

[0078] Since the larger the inter-class distance between the background and foreground of the image to be segmented, the better the image segmentation effect. Therefore, by calculating the inter-class variance of the image gray levels, the maximum variance value between image types is determined, and thus the optimal segmentation threshold of the image is obtained. The result is shown in the following formula: ; In the formula, the optimal segmentation threshold of the image is , and the maximum variance is . Based on the determined optimal segmentation threshold, the adaptive segmentation of the multi-threshold image is completed. The specific image segmentation process is as Figure 5 shown.

[0079] In this embodiment, the target detection algorithm uses the YOLOv5 algorithm. The target detection algorithm (YOLOv5) can be divided into an input end, a support layer, a fused feature layer, and an output layer. The input end of the target detection algorithm uses the Mosaic algorithm to implement the data augmentation design, aiming to enrich the data set and reduce the training duration to a certain extent. In the support layer, the input image is cut using a focus structure, and the image is cut into four parts. The data of each part of the image is equivalent to that obtained by double downsampling. After vertical channel splicing, convolution operations are then performed. In the fused feature layer, the Path Aggregation Network (PANet) is responsible for feature fusion. What it does is to add an information flow path on the basis of the Feature Pyramid Network (FPN), shortening the information transmission path to a certain extent.

[0080] The data augmentation method used in the experimental part of the YOLOv5 algorithm is the Mosaic method. The Mosaic data augmentation refers to the CutMix method, which is theoretically similar, but the Mosaic uses four pictures and performs splicing in a random cropping, scaling, and arranging manner. The distribution of large, medium, and small targets in the MSCOCO data set is uneven, while the Mosaic method uses random scaling for splicing, increasing the data of small targets and enriching the data set, thus making the network more robust.

[0081] The overall network structure of the YOLOv5 algorithm is as Figure 6 shown, including CBL modules, Focus modules, SPP modules, CSP1_X, and CSP2_X modules.

[0082] The input image size of YOLOv5 is 608×608, and it has the following sections: 1) CBL module: This module consists of three parts: convolution operation (Conv), batch normalization operation (BN), and activation function (Leaky-Relu).

[0083] 2) Focus module: This structure slices the input image and then stitches together the sliced results.

[0084] 3) SPP module: This structure respectively uses max pooling with 1×1, 5×5, 9×9, and 13×13, and then stitches together the obtained results to get the fused features.

[0085] 4) CSP1_X and CSP2_X modules: Drawing on the idea of CSPNet, this module consists of three parts: a convolutional layer, a CBL module, and a ResUnit module.

[0086] The YOLOv5 object detection framework can be divided into the following parts: the input end, the backbone network, the Neck part, and the output end. The multi-scale feature fusion structure of the Neck part is improved.

[0087] In terms of the two methods of summing by pixel and stitching by channel, the improved algorithm selects the stitching method by channel dimension. When dealing with the problem of model feature fusion, the stitching method by channel dimension is adopted while using a bidirectional network.

[0088] The way of fusing features is represented by the following formula: ; Among them, the symbol represents stitching by channel dimension, , , represent three features in the bidirectional feature network, represents the feature used for subsequent object detection after fusion.

[0089] For example, Figure 7 , taking the generation process of the feature as an example. The feature is obtained by stitching together the feature obtained after sampling by , and in the channel dimension. The generation process is shown in the following formula, where the CSP2_1 operator is denoted as a function, refers to achieving two-fold downsampling through convolutional operations: ; Similarly, the generation process of the feature is shown in the following formula: .

[0090] The feature fusion network of the YOLOv5 algorithm and the improved algorithm is combined and drawn. At the same time, in order to further improve the effect of object detection accuracy, features at a lower level than are used .

[0091] Compared with traditional medium and large object detection methods, the accuracy of small object detection is relatively poor. Therefore, on the premise of ensuring the unchanged accuracy of medium and large object detection, the way of making full use of low-level features is adopted to improve the accuracy of small object detection.

[0092] Features for small object detection do not establish a connection with the features of the previous level, but utilize features and features, where the features are obtained by upsampling the features. By making full use of the features, the information of the high-resolution features is introduced into the feature fusion. Considering the wide application of the bidirectional network, the following algorithm is finally obtained: ; refers to achieving two-fold upsampling through bilinear interpolation operation. The low-level features are obtained by fusing the features and the features, as shown in the following formula: .

[0093] Considering the bidirectional feature fusion method, generating the features requires the features obtained after downsampling , and in three parts. The acquisition method of the features for small object detection is shown in the following formula: .

[0094] Similarly, the generation process of the features is as follows:

[0095] Making full use of the low-level features should improve the performance of the detection model in small object detection on the premise of ensuring the accuracy of large and medium object detection, so as to further improve the accuracy of object detection at high altitudes in rainy weather.

[0096] To improve the accuracy of image recognition, a two-dimensional Gaussian function is usually used for Gaussian filtering and combined with the Laplace operator to process the image. The output image can be expressed as , where is the two-dimensional Gaussian function, is the original image. Edge extraction is performed using the discrete LoG operator (Laplacian of Gaussian operator) by the following formula: ; The convolution kernel size of the LoG operator and has the relationship (INT represents the integer operation). The identity element of the LoG operator can be expressed as: ; ; s and t represent the row and column indices of the convolution kernel (discrete Laplacian of Gaussian operator) for locating each element position in the kernel. For example, G( s, t ) represents the element value in the s-th row and the t -th column of the convolution kernel. k is a half-width parameter used to determine the size of the convolution kernel . For example, if k = 2, the convolution kernel size is 5×5.

[0097] For a two-dimensional image , the image data can be mapped to a pure quantum state with qubits , where encodes the pixel values. The image information needs to be normalized before being written into the quantum state.

[0098] According to the QPIE model, the image Q can be expressed as a pure quantum state as follows: ; Encode the LoG operator template into a quantum system with 5 qubits to obtain the quantum state of the LoG operator: .

[0099] Perform the tensor product operation on the quantum states and to obtain . represents the joint quantum state composed of the subsystems described by and . | QLod > is a quantum state with its amplitude , corresponding to the encoding of the classical image data.

[0100] Define a probability amplitude permutation operator U. Through the probability amplitude permutation operator U and quantum circuit processing, the edge of the quantum image is finally measured. After obtaining all the edge pixel points of the original image, the edge image can also be binarized, that is ; where the threshold is calculated by using the maximum inter-class variance method, and thus a binary image of the edge image can be obtained.

[0101] Step 6: Establish a database according to the vehicle features extracted in Step 5.

[0102] The above embodiments are only exemplary embodiments of the present invention and are not used to limit the present invention. Those skilled in the art can make various modifications or equivalent replacements to the present invention within the essence and protection scope of the present invention, and such modifications or equivalent replacements should also be regarded as falling within the protection scope of the present invention.

Claims

1. A method for positioning and identifying accident vehicles on highways based on quantum and drone technologies, characterized in that, It includes the following steps: Step 1: Construct a PPP-RTK model, introduce the TTTrack algorithm, and collect highway vehicle image information by drones based on the PPP-RTK model and the TTTrack algorithm; Step 2: Preprocess the images obtained in Step 1; Step 3: Determine whether the vehicles contained in the images obtained in Step 2 are accident vehicles; If they are accident vehicles, exit the pictures. If they are accident vehicles, enter Step 4; Step 4: Extract features from the images obtained in Step 3 through a CNN convolutional neural network and an Xception classification model in sequence; Step 5: Prepare quantum images through the ONEQR model, process them through an object detection algorithm, and perform image edge detection, smoothing processing, and threshold segmentation reconstruction on the images to extract vehicle features in the images containing accident vehicles; Step 6: Establish a database based on the vehicle features extracted in Step 5.

2. The method for positioning and identifying accident vehicles on expressways based on quantum and drone technologies according to claim 1, wherein The PPP-RTK model includes a PPP-RTK function model, a PPP-RTK regional tropospheric model, and a PPP-RTK regional ionospheric model; the PPP-RTK regional tropospheric model adopts the HIQM4 model.

3. The method for positioning and identifying accident vehicles on highways based on quantum and drone technologies according to claim 2, characterized in that, The PPP-RTK function model is: ; ; wherein, are the receiver, satellite, frequency, and epoch number respectively; is the tropospheric wet delay projection function; is the satellite phase deviation; is the ratio of the ionosphere at other frequencies to the ionosphere at the first frequency; is the receiver phase deviation; is the receiver clock error solved by the new observation equation, is the satellite clock error, is the zenith tropospheric wet delay solved by the new observation equation, is the slant delay of the ionosphere at the first frequency, is the receiver code deviation, is the satellite code deviation, is the ambiguity of non-first stations, is the receiver phase deviation, is the undifferenced and uncombined code after deducting the precise satellite-to-ground distance and various model errors, is the residual of the phase observation value after deducting the precise satellite-to-ground distance and various model errors.

4. The method for positioning and identifying accident vehicles on highways based on quantum and drone technologies according to claim 2, characterized in that, The PPP-RTK regional ionospheric model is expressed as: ; Among them, are the coefficients of the ionospheric slant delay polynomial, is the longitude of the reference station, is the latitude of the reference station, is the difference between the longitude of the reference station and the longitude of the center of the survey area, is the difference between the latitude of the reference station and the latitude of the center of the survey area.

5. The method for positioning and identifying accident vehicles on highways based on quantum and drone technologies according to claim 1, wherein, The TTTrack algorithm includes a Siamese backbone network, an adaptive spatio-temporal information extraction module, and a spatio-temporal context mapping module. The Siamese backbone network is used to extract features from the template image and the search image; the adaptive spatio-temporal information extraction module adopts a dual-template mechanism of a reference template and a dynamic template; the spatio-temporal context mapping module is used to capture the spatio-temporal correlation between consecutive frames.

6. The method for positioning and identifying accident vehicles on a highway based on quantum and drone technologies according to claim 1, wherein The preparation of the quantum image in step 5 specifically includes: applying the operator to to obtain the intermediate state ; storing the gray level and real coordinate information corresponding to each position in ; applying operators to to complete the preparation of all pixels.

7. The method for positioning and identifying accident vehicles on highways based on quantum and drone technologies according to claim 1, characterized in that The threshold segmentation reconstruction in Step 5 specifically includes: obtaining the number of gray image levels, calculating the probability of the gray image in the total pixels, and initializing the gray threshold; dividing the image into two categories: background and foreground, and calculating the gray-level occurrence probability and the probability average of different image types based on the Bayesian optimization algorithm; calculating the mean value of the gray levels of the original image, calculating the between-class variance of the image, and determining the maximum between-class variance to calculate the optimal segmentation threshold.

8. The method for positioning and identifying accident vehicles on expressways based on quantum and drone technologies according to claim 1, wherein The specific image edge detection in step 5 is as follows: The quantum LoG edge detection algorithm is adopted. According to the QPIE model, the image Q can be expressed as a pure quantum state , and the LoG operator template is encoded into a quantum system containing 5 qubits to obtain the quantum state of the LoG operator ; The tensor product operation is performed on the quantum states and to obtain . Through the probability amplitude permutation operator U and quantum circuit processing, the quantum image edge is finally measured.

Images