Target vehicle identification and path tracking method in video based on big data analysis

Abstract

The invention discloses a vehicle identification and path tracking method in a video based on big data analysis. The method comprises the following steps: collecting video data through a camera; performing image enhancement, noise elimination and feature extraction on the video data; identifying the vehicle by using an image analysis technology; tracking a vehicle path in real time by adopting an algorithm; predicting a vehicle path using an intelligent algorithm; a large amount of video data is stored, processed and analyzed by using a big data technology; multiple vehicles are identified and tracked at the same time; real-time processing is achieved through an optimization algorithm and hardware acceleration, and delay is reduced through edge calculation. The method combines various technologies, efficiently and accurately identifies and tracks the vehicle, and is suitable for a complex traffic environment. Through distributed computing and parallel processing technologies, a large amount of video data can be processed, and real-time requirements can be quickly responded. The method can manage a plurality of vehicles at the same time, improves the image quality, enhances the recognition and tracking accuracy, and is suitable for intelligent traffic management and automatic driving application.

Description

Technical Field

[0001] The present invention relates to video processing and analysis in intelligent transportation, and in particular to a method for target vehicle recognition and path tracking in videos based on big data analysis. Background Art

[0002] With the continuous advancement of modern urbanization, urban traffic management and autonomous driving technology are facing more and more challenges. Existing traffic management systems and autonomous driving technologies mainly rely on traditional sensors, video surveillance equipment, and simple image processing and data analysis algorithms. However, these systems have obvious limitations in processing efficiency, recognition accuracy, real-time performance, multi-target management capabilities, and path prediction.

[0003] Traditional video surveillance systems mainly rely on fixed cameras and closed-circuit television (CCTV) networks to monitor major traffic routes in real time. These systems usually rely on manual viewing of video footage or use simple image processing technology to perform basic traffic event detection. However, with the increase in traffic volume, the efficiency and accuracy of manual monitoring have greatly decreased, making it difficult to respond to complex traffic conditions in real time. In addition, these systems are inefficient in processing massive amounts of video data and cannot meet the needs of large-scale data processing and real-time analysis.

[0004] Existing vehicle recognition technologies are mostly based on image processing and machine learning algorithms, such as license plate recognition (LPR) and vehicle classification technologies. These technologies can achieve high recognition accuracy in static environments, but their multi-target recognition and tracking capabilities are insufficient in dynamic and complex traffic environments. In particular, the recognition accuracy drops significantly in situations such as lighting changes, target occlusion, and changes in vehicle appearance. In addition, existing technologies perform poorly in multi-target recognition and tracking, and it is difficult to ensure high accuracy and reliability in complex traffic environments.

[0005] Traditional path tracking and prediction methods mainly rely on GPS data and simple prediction models. These methods have great limitations when facing complex and dynamic traffic environments and multi-target tracking, and it is difficult to achieve high-precision path tracking and prediction. At the same time, the lack of comprehensive consideration of traffic flow and road information makes the accuracy and reliability of path planning low. Existing methods show insufficient path prediction capabilities in dynamically changing traffic environments and cannot meet the needs of high-precision path planning.

[0006] With the rapid growth of data volume, how to efficiently process and analyze massive traffic data has become an important challenge. Existing big data processing technologies are mainly based on distributed computing frameworks, such as Hadoop and Spark. These technologies have good performance in processing static data, but there is a lot of room for improvement in real-time data processing and analysis. Poor real-time performance and high processing delay are the main defects of existing big data processing technologies, which cannot meet the needs of real-time traffic data processing and analysis.

[0007] Multi-target tracking technology plays an important role in intelligent transportation and monitoring systems. Existing MOT technologies are mainly based on Kalman filtering, particle filtering and data association algorithms. These technologies perform well in simple scenarios, but are easily affected by occlusion, changes in lighting and changes in target appearance in complex traffic environments, resulting in reduced tracking accuracy and even target loss. Existing multi-target tracking technologies have great limitations in complex traffic environments and cannot meet the needs of high-precision, multi-target management.

[0008] Therefore, how to provide a method for target vehicle recognition and path tracking in videos based on big data analysis is an urgent problem that technicians in this field need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose a method for target vehicle recognition and path tracking in a video based on big data analysis. The present invention utilizes computer vision, deep learning, machine learning and big data processing technologies, and describes in detail the complete process from video data acquisition, preprocessing, vehicle recognition, path tracking, path prediction, data distributed processing, multi-target tracking to real-time processing and response. The method has the advantages of efficient processing of massive data, high recognition accuracy, strong real-time response capability, superior multi-target management capability, and accurate path prediction and optimization, and is suitable for intelligent traffic management and autonomous driving systems in complex traffic environments.

[0010] According to an embodiment of the present invention, a method for identifying and tracking a target vehicle in a video based on big data analysis includes the following steps:

[0011] S1. Collect video data through cameras installed on traffic arteries or vehicles;

[0012] S2, preprocessing the collected video data, including image enhancement, noise removal and key feature extraction;

[0013] S3, using visual Transformer to analyze the preprocessed video data and identify the target vehicle in the video;

[0014] S4, using recurrent neural network and long short-term memory network to track the movement path of the target vehicle in real time;

[0015] S5. Use generative adversarial networks and reinforcement learning algorithms to predict the path of the target vehicle;

[0016] S6. Use big data processing technology to perform distributed storage, processing and analysis of massive video data;

[0017] S7, identify and track multiple target vehicles simultaneously;

[0018] S8. Realize real-time processing and analysis of video data through algorithm optimization and hardware acceleration, and use edge computing to reduce latency.

[0019] Optionally, the step S1 specifically includes:

[0020] S11. Install cameras on major roads or vehicles. The camera locations and angles should be able to cover the target area.

[0021] S12, use a high-resolution camera with a resolution of R×C, where R represents the number of rows and C represents the number of columns;

[0022] S13, configure the frame rate of the camera to be f, and the frame rate range is f∈[15,60] frames per second;

[0023] S14, transmitting the collected video data to a data processing center via a network, using TCP / IP, UDP and other protocols suitable for video transmission;

[0024]

[0025] Among them, P is the data transmission efficiency, M is the number of data packets, and B j is the data volume of the jth packet, T j is the transmission time, S j is the network bandwidth, N j For noise interference;

[0026] S15, real-time monitoring and maintenance of the working status of the camera;

[0027]

[0028] Among them, R is the reliability evaluation of the camera's working status, W is the number of time periods in the monitoring cycle, and U m is the camera working status in the mth time period, V m It is an indicator of environmental complexity.

[0029] Optionally, the step S2 specifically includes:

[0030] S21, performing image enhancement on the collected video data;

[0031]

[0032] Among them, I enhanced is the enhanced image, I original is the original image, μ and σ are the mean and standard deviation of the original image, μ′ and σ′ are the target mean and standard deviation, respectively;

[0033] S22, performing noise elimination and smoothing the image using a Gaussian filter;

[0034]

[0035] Among them, I denoised (x, y) is the denoised image, I(x, y) is the original image pixel value, σ is the standard deviation of the Gaussian filter, and k is the radius of the filter;

[0036] S23, extracting key feature points in the image, and performing feature extraction using a scale-invariant feature transformation algorithm;

[0037] D(x,y,σ)=(L(x,y,kσ)-L(x,y,σ));

[0038] Where D(x, y, σ) is the difference in scale space, L(x, y, σ) is the Gaussian blurred image of the image at scale σ, and k is the scale multiple;

[0039] S24, performing inter-frame difference on the pre-processed video frames to detect motion areas;

[0040] M t (x, y) = |I t (x, y)-I t-1 (x, y)|;

[0041] Among them, M t (x, y) is the difference image between the t-th frame and the t-1-th frame, I t (x, y) is the image of the tth frame, I t-1 (x, y) is the image of the t-1th frame;

[0042] S25, performing morphological processing, dilating and corroding the motion area;

[0043]

[0044] Among them, I morph is the image after morphological processing, I is the input binary image, For expansion operation, is the corrosion operation, and B is the structural element.

[0045] Optionally, the step S3 specifically includes:

[0046] S31, converting the pre-processed video data into an image sequence {I t}, where t represents the time frame;

[0047] {I t}={I 1 , I 2 , ..., I n};

[0048] Among them, I t is the t-th frame image, n is the total number of frames;

[0049] S32, use the visual Transformer model to extract features from the image sequence, and transform each frame image I t Mapped to high-dimensional feature space, we get the feature vector F t ;

[0050] S33, all the feature vectors of the frames {F t} Input into the multi-head self-attention mechanism to calculate the similarity between each feature vector;

[0051]

[0052] Among them, A i，j is the attention weight between the i-th frame and the j-th frame, Q i and K j are the query vector and key vector of the i-th frame and j-th frame respectively;

[0053] S34, weighted summation through the attention mechanism, to obtain the weighted feature representation Z of each frame image t ;

[0054]

[0055] Among them, Z t is the weighted feature representation of the t-th frame image, V j is the value vector of the jth frame;

[0056] S35, weighted feature representation Z t Input into the classifier to identify the target vehicle in the video;

[0057] S36, performing time series analysis on the classification results of all frames to determine the position and motion trajectory of the target vehicle in the video;

[0058]

[0059] Among them, Trajectory is the trajectory analysis function, which converts the classification results Mapped to the motion trajectory of the target vehicle.

[0060] Optionally, the step S4 specifically includes:

[0061] S41, the recognition result of the target vehicle As an input sequence, it is input into the recurrent neural network;

[0062]

[0063] Among them, h t is the classification result of the t-th frame image, h t is the hidden state of the recurrent neural network at the tth time step, σ is the activation function, W h and U h is the weight matrix, b h is the bias vector;

[0064] S42. Use the long short-term memory network LSTM to further process the hidden state to capture long-term dependencies and obtain the updated hidden state

[0065]

[0066] Among them, i t 、f t , o t are the activation vectors of the input gate, forget gate, and output gate, respectively, c t is the cell state, ⊙ is the element product, W i , W f , W o , W c and U i , U f , U o , U c is the weight matrix, b i , b f , b o , b c is the bias vector;

[0067] S43, output o through long short-term memory network t Calculate the position prediction of the target vehicle at time step t;

[0068]

[0069] Among them, t is the position prediction, W o2 is the weight matrix, b o2 is the bias vector, λ is the adjustment coefficient, α jis the attention weight, m ​​is the number of past time steps;

[0070] S44. Prediction results for all time steps Perform sequence analysis to form the target vehicle's motion path;

[0071]

[0072] Among them, Path is the moving path of the target vehicle;

[0073] S45, smoothing the path according to actual conditions;

[0074]

[0075] Among them, w t is the smoothing weight, η is the adjustment coefficient, K is the size of the smoothing window, β k is the smoothing coefficient;

[0076] S46, finally obtaining the real-time motion path of the target vehicle in the video;

[0077]

[0078] Among them, γ is the application coefficient, δ is the adjustment coefficient, P is the historical path length, μ p is the weight of the historical path.

[0079] Optionally, the step S5 specifically includes:

[0080] S51, taking the historical movement path Path of the target vehicle as input data, and inputting it into the generator of the generative adversarial network GAN;

[0081] z t =Generator(Path t );

[0082] Among them, z t is the potential variable generated by the generator at time step t, Path t is the historical motion path of the target vehicle at time step t;

[0083] S52, using the discriminator of the generative adversarial network GAN to evaluate the generated path and calculate the difference between the generated path and the true path;

[0084] D(Path,z t )=σ(W d ·z t +b d );

[0085] Among them, D is the discriminator function, σ is the activation function, and W d is the weight matrix, b d is the bias vector;

[0086] S53, through adversarial training, the path generated by the generator is made close to the real path;

[0087]

[0088] Among them, L GAN is the loss function of GAN, p data is the data distribution of the real path, p z is the latent variable distribution of the generator;

[0089] S54, the potential variable z generated by the generator t Input into the reinforcement learning algorithm for path optimization and prediction;

[0090]

[0091] Among them, π(a t |s t ) is the policy function, a t For state s t The action selected next, Q(s t , a t ) is the action value function, A is the action space;

[0092] S55. Update the action value function Q using the value iteration method in reinforcement learning;

[0093]

[0094] Among them, r t is the reward obtained at time step t, γ is the discount factor, s t+1 is the state at time step t+1;

[0095] S56. Finally, an optimized target vehicle path prediction is obtained.

[0096] Optionally, the step S6 specifically includes:

[0097] S61, dividing the collected video data into a plurality of data blocks, each data block containing a certain number of frames;

[0098]

[0099] Among them, D i is the i-th data block, F i，j is the jth frame of the ith data block, T iis the time length of the i-th data block, Δt is the time interval between each frame, and n i is the number of frames in the i-th data block;

[0100] S62, storing the video data blocks on multiple storage nodes using a distributed file system;

[0101] S i =HDFS({F i，j}), for i=1, 2,...,m;

[0102] Among them, S i is the data block stored on the i-th node, HDFS is a distributed file system function that stores all frames in the i-th data block on multiple nodes;

[0103] S63, using a distributed computing framework to perform parallel processing and analysis on the video data blocks;

[0104]

[0105] Among them, SparkProcess is a distributed computing process, φ is a function for processing each frame of an image, and each frame of an image in the i-th data block is processed to obtain the processing result R i ;

[0106] S64, performing data fusion on the processing results to obtain a global analysis result;

[0107]

[0108] Among them, Fuse is the data fusion function, ω i is the weight of the ith data block, R qlobal It is the global analysis result;

[0109] S65. Apply big data processing technology to further aggregate and summarize the global analysis results;

[0110] I=Aggregate(R g local)=Ψ(R g lobal);

[0111] Among them, I is the extracted information, Aggregate is the aggregation function, Ψ is the aggregation operation, and the global analysis result R qlobal Further processing and summarization are performed to extract useful information I.

[0112] Optionally, the step S7 specifically includes:

[0113] S71. Perform target detection on each frame of video image to identify all possible target vehicles

[0114]

[0115] in, is the i-th target vehicle in the t-th frame image, M t is the total number of target vehicles detected in the t-th frame image, Detector is the target detection algorithm, I t is the t-th frame image;

[0116] S72: Extract feature vectors for each detected target vehicle

[0117]

[0118] in, is the feature vector of the i-th target vehicle in the t-th frame image, and FeatureExtractor is the feature extraction algorithm;

[0119] S73, using the Kalman filter algorithm to predict and update the state of each target vehicle to obtain an updated state estimate

[0120] Prediction steps:

[0121] Update steps:

[0122] in, For the predicted state, is the prediction covariance, is the Kalman gain, is the observation value, H is the observation matrix, F is the state transfer matrix, B is the control matrix, u t is the control input, Q is the process noise covariance, R is the observation noise covariance, and I is the identity matrix;

[0123] S74, using a multi-target tracking algorithm to perform data association, matching the target vehicle in the current frame with the target vehicle in the previous frame, and updating the trajectory of each target vehicle;

[0124]

[0125] Among them, Match is the matching function, c ij is the cost matrix, x ij is the matching result, M t-1 is the total number of target vehicles in the previous frame;

[0126] S75, for the target vehicle that is not matched, initialize a new track and add it to the target vehicle tracking list;

[0127]

[0128] Among them, NewTrajectory is the new trajectory initialization function, Init is the initialization operation, and Update is the update operation;

[0129] S76. Manage and update the trajectories of all target vehicles in real time.

[0130] Optionally, the step S8 specifically includes:

[0131] S81. Improve the efficiency of video data processing by optimizing algorithms, including parallelizing and pipelining image processing, feature extraction, and target recognition algorithms;

[0132]

[0133] Among them, Process par 、Extract par and Detect par Respectively represent parallel processing of images, feature extraction and target recognition, Time total is the total processing time;

[0134] S82, use GPU and FPGA to accelerate the processing of video data;

[0135] Speed accel =γ·Speed cpu ;

[0136] Among them, Speed accel is the processing speed after hardware acceleration, γ is the acceleration factor, Speed cpu is the CPU processing speed;

[0137] S83, using edge computing architecture to offload data processing tasks to edge devices;

[0138]

[0139] Among them, Latency edge is the total delay under the edge computing architecture, Time edge (i) is the processing time of the i-th edge device, Time cloud (j) is the jth cloud processing time, N is the number of edge devices, and M is the number of cloud servers;

[0140] S84, process the video data through edge devices and perform analysis and processing locally;

[0141]

[0142] Among them, LocalProcessing edge For local processing on edge devices, Process local (V k ) is the time for the kth video frame to be processed locally, and K is the total number of video frames processed;

[0143] S85, combining edge computing and cloud computing, sending computing tasks and big data analysis tasks to the cloud for processing;

[0144] HybridProcessing=α·EdgeProcessing+β·CloudProcessing;

[0145] Among them, HybridProcessing is the hybrid computing mode, EdgeProcessing is the processing capacity of edge computing, CloudProcessing is the processing capacity of cloud computing, α and β are the weight coefficients of edge computing and cloud computing;

[0146] S86. Optimize data transmission protocols and network architecture;

[0147]

[0148] Among them, Transmission opt For optimized data transmission delay, Data size is the amount of data transmitted, Bandwidth is the network bandwidth, Overkey is the transmission overhead;

[0149] S87. Feedback the processing results to the traffic management system or the automatic driving system to achieve monitoring and path optimization of the target vehicle.

[0150] The beneficial effects of the present invention are:

[0151] (1) The present invention proposes a method for target vehicle recognition and path tracking in videos based on big data analysis. By comprehensively applying multiple advanced technologies, the performance of traffic management and autonomous driving systems is significantly improved, and the specific beneficial effects are as follows:

[0152] (2) Efficient processing of massive data: Video data is collected by cameras on major traffic routes or vehicles, and distributed computing and parallel processing technologies are used to efficiently store, process and analyze massive video data, greatly improving data processing efficiency and meeting city-level traffic flow monitoring and management needs.

[0153] (3) High recognition accuracy: The collected video data is preprocessed, including image enhancement, noise removal, and feature extraction, and analyzed using a visual transformer to accurately identify the target vehicle in the video. In complex traffic environments, it can maintain a high recognition accuracy and effectively cope with challenges such as lighting changes, target occlusion, and changes in vehicle appearance.

[0154] (4) Strong real-time response capability: Recurrent neural networks and long short-term memory networks are used to track the movement path of the target vehicle in real time, and the real-time processing capability of the system is improved through optimization algorithms and hardware acceleration technologies (such as GPU and FPGA). At the same time, edge computing is used to reduce data transmission delays, achieve real-time response and decision-making, and meet the real-time requirements of traffic management and autonomous driving systems.

[0155] (5) Superior multi-target management capability: The system can simultaneously identify and track multiple target vehicles, and use multi-target tracking algorithms (such as Kalman filtering and Hungarian algorithm) to improve multi-target management capabilities. In complex traffic environments, the system can stably and accurately track multiple target vehicles to ensure the effectiveness of traffic management.

[0156] (6) Accurate path prediction and optimization: Generative adversarial networks (GANs) and reinforcement learning algorithms are used to predict and optimize the path of the target vehicle. The deep learning model captures the vehicle's movement pattern, combines traffic flow and road information, and achieves high-precision path prediction and optimization, providing reliable support for autonomous driving and emergency dispatch.

[0157] (7) Image quality improvement: In video data preprocessing, image enhancement and noise elimination techniques are used to improve video image quality and enhance the accuracy and stability of vehicle recognition and path tracking.

[0158] (8) Comprehensive technical application: The present invention combines computer vision, deep learning, machine learning and big data processing technologies to form a complete solution from data collection, preprocessing, recognition, tracking to prediction, which has strong practicality and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0159] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0160] Figure 1 This is an overall flow chart of a method for target vehicle recognition and path tracking in a video based on big data analysis proposed by the present invention;

[0161] Figure 2 Schematic diagram of big data processing and multi-target tracking; DETAILED DESCRIPTION

[0162] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0163] refer to Figure 1-2 , a method for target vehicle recognition and path tracking in a video based on big data analysis, comprising the following steps:

[0164] S1. Collect video data through cameras installed on traffic arteries or vehicles;

[0165] In this implementation, step S1 specifically includes:

[0166] S11. Install cameras on major roads or vehicles. The camera locations and angles should be able to cover the target area.

[0167] S12, use a high-resolution camera with a resolution of R×C, where R represents the number of rows and C represents the number of columns;

[0168] S13, configure the frame rate of the camera to be f, and the frame rate value range is f∈[15, 60] frames per second;

[0169]

[0170] Among them, E is the effectiveness evaluation of the frame rate configuration, K is the frame number set, and f k is the frame rate of the kth frame, f min is the lower limit of frame rate;

[0171] S14, transmitting the collected video data to a data processing center via a network, using TCP / IP, UDP and other protocols suitable for video transmission;

[0172]

[0173] Among them, P is the data transmission efficiency, M is the number of data packets, and B j is the data volume of the jth packet, T j is the transmission time, S j is the network bandwidth, N j For noise interference;

[0174] S15, real-time monitoring and maintenance of the working status of the camera;

[0175]

[0176] Among them, R is the reliability evaluation of the camera's working status, W is the number of time periods in the monitoring cycle, and U mis the camera working status in the mth time period, V m It is an indicator of environmental complexity.

[0177] S2, preprocessing the collected video data, including image enhancement, noise removal and key feature extraction;

[0178] In this implementation, S2 specifically includes:

[0179] S21, performing image enhancement on the collected video data;

[0180]

[0181] Among them, I enhanced is the enhanced image, I original is the original image, μ and σ are the mean and standard deviation of the original image, μ′ and σ′ are the target mean and standard deviation, respectively;

[0182] S22, performing noise elimination and smoothing the image using a Gaussian filter;

[0183]

[0184] Among them, I denoised (x, y) is the denoised image, I(x, y) is the original image pixel value, σ is the standard deviation of the Gaussian filter, and k is the radius of the filter;

[0185] S23, extracting key feature points in the image, and performing feature extraction using a scale-invariant feature transformation algorithm;

[0186] D(x,y,σ)=(L(x,y,kσ)-L(x,y,σ));

[0187] Where D(x, y, σ) is the difference in scale space, L(x, y, σ) is the Gaussian blurred image of the image at scale σ, and k is the scale multiple;

[0188] S24, performing inter-frame difference on the pre-processed video frames to detect motion areas;

[0189] M t (x, y) = |I t (x, y)-I t-1 (x, y)|;

[0190] Among them, M t (x, y) is the difference image between the t-th frame and the t-1-th frame, I t (x, y) is the image of the tth frame, I t-1 (x, y) is the image of the t-1th frame;

[0191] S25, performing morphological processing, dilating and corroding the motion area;

[0192]

[0193] Among them, I morph is the image after morphological processing, I is the input binary image, For expansion operation, is the corrosion operation, and B is the structural element.

[0194] S3, using visual Transformer to analyze the preprocessed video data and identify the target vehicle in the video;

[0195] In this implementation, S3 specifically includes:

[0196] S31, converting the preprocessed video data into an image sequence {I t}, where t represents the time frame;

[0197] {I t}={I 1 , I 2 , ..., I n};

[0198] Among them, I t is the t-th frame image, n is the total number of frames;

[0199] S32, use the visual Transformer model to extract features from the image sequence, and transform each frame image I t Mapped to high-dimensional feature space, we get the feature vector F t ;

[0200] S33, all the feature vectors of the frames {F t} Input into the multi-head self-attention mechanism to calculate the similarity between each feature vector;

[0201]

[0202] Among them, A i，j is the attention weight between the i-th frame and the j-th frame, Q i and K j are the query vector and key vector of the i-th frame and j-th frame respectively;

[0203] S34, weighted summation through the attention mechanism, to obtain the weighted feature representation Z of each frame image t ;

[0204]

[0205] Among them, Z tis the weighted feature representation of the t-th frame image, V j is the value vector of the jth frame;

[0206] S35, weighted feature representation Z t Input into the classifier to identify the target vehicle in the video;

[0207] S36, performing time series analysis on the classification results of all frames to determine the position and motion trajectory of the target vehicle in the video;

[0208]

[0209] Among them, Trajectory is the trajectory analysis function, which converts the classification results Mapped to the motion trajectory of the target vehicle.

[0210] S4, using recurrent neural network and long short-term memory network to track the movement path of the target vehicle in real time;

[0211] In this implementation, S4 specifically includes:

[0212] S41, the recognition result of the target vehicle As an input sequence, it is input into the recurrent neural network;

[0213]

[0214] Among them, h t is the classification result of the t-th frame image, h t is the hidden state of the recurrent neural network at the tth time step, σ is the activation function, W h and U h is the weight matrix, b h is the bias vector;

[0215] S42. Use the long short-term memory network LSTM to further process the hidden state to capture long-term dependencies and obtain the updated hidden state

[0216]

[0217]

[0218] Among them, i t 、f t , o t are the activation vectors of the input gate, forget gate, and output gate, respectively, c t is the cell state, ⊙ is the element product, W i , W f , W o , W c and Ui , U f , U o , U c is the weight matrix, b i , b f , b o , b c is the bias vector;

[0219] S43, output o through long short-term memory network t Calculate the position prediction of the target vehicle at time step t;

[0220]

[0221] Among them, t is the position prediction, W o2 is the weight matrix, b o2 is the bias vector, λ is the adjustment coefficient, α j is the attention weight, m ​​is the number of past time steps;

[0222] S44. Prediction results for all time steps Perform sequence analysis to form the target vehicle's motion path;

[0223]

[0224] Among them, Path is the moving path of the target vehicle;

[0225] S45, smoothing the path according to actual conditions;

[0226]

[0227] Among them, w t is the smoothing weight, η is the adjustment coefficient, K is the size of the smoothing window, β k is the smoothing coefficient;

[0228] S46, finally obtaining the real-time motion path of the target vehicle in the video;

[0229]

[0230] Among them, γ is the application coefficient, δ is the adjustment coefficient, P is the historical path length, μ p is the weight of the historical path.

[0231] S5. Use generative adversarial networks and reinforcement learning algorithms to predict the path of the target vehicle;

[0232] In this implementation, S5 specifically includes:

[0233] S51, taking the historical movement path Path of the target vehicle as input data, and inputting it into the generator of the generative adversarial network GAN;

[0234] z t =Generator(Path t );

[0235] Among them, z t is the potential variable generated by the generator at time step t, Path t is the historical motion path of the target vehicle at time step t;

[0236] S52, using the discriminator of the generative adversarial network GAN to evaluate the generated path and calculate the difference between the generated path and the true path;

[0237] D(Path,z t )=σ(W d ·z t +b d );

[0238] Among them, D is the discriminator function, σ is the activation function, and W d is the weight matrix, b d is the bias vector;

[0239] S53, through adversarial training, the path generated by the generator is made close to the real path;

[0240]

[0241] Among them, L GAN is the loss function of GAN, p data is the data distribution of the real path, p z is the latent variable distribution of the generator;

[0242] S54, the potential variable z generated by the generator t Input into the reinforcement learning algorithm for path optimization and prediction;

[0243]

[0244] Among them, π(a t |s t ) is the policy function, a t For state s t The action selected next, Q(s t ,a t ) is the action value function, A is the action space;

[0245] S55. Update the action value function Q using the value iteration method in reinforcement learning;

[0246]

[0247] Among them, r t is the reward obtained at time step t, γ is the discount factor, s t+1 is the state at time step t+1;

[0248] S56. Finally, an optimized target vehicle path prediction is obtained.

[0249] S6. Use big data processing technology to perform distributed storage, processing and analysis of massive video data;

[0250] In this implementation, S6 specifically includes:

[0251] S61, dividing the collected video data into a plurality of data blocks, each data block containing a certain number of frames;

[0252]

[0253] Among them, D i is the i-th data block, F i，j is the jth frame of the ith data block, T i is the time length of the i-th data block, Δt is the time interval between each frame, and n i is the number of frames in the i-th data block;

[0254] S62, storing the video data blocks on multiple storage nodes using a distributed file system;

[0255] S i =HDFS({F i，j}), for i=1, 2,..., m;

[0256] Among them, S i is the data block stored on the i-th node, HDFS is a distributed file system function that stores all frames in the i-th data block on multiple nodes;

[0257] S63, using a distributed computing framework to perform parallel processing and analysis on the video data blocks;

[0258]

[0259] Among them, SparkProcess is a distributed computing process, φ is a function for processing each frame of an image, and each frame of an image in the i-th data block is processed to obtain the processing result R i ;

[0260] S64, performing data fusion on the processing results to obtain a global analysis result;

[0261]

[0262] Among them, Fuse is the data fusion function, ω i is the weight of the ith data block, R global It is the global analysis result;

[0263] S65. Apply big data processing technology to further aggregate and summarize the global analysis results;

[0264] I=Aggregate(R global )=Ψ(R global );

[0265] Among them, I is the extracted information, Aggregate is the aggregation function, Ψ is the aggregation operation, and the global analysis result R global Further processing and summarization are performed to extract useful information I.

[0266] S7, identify and track multiple target vehicles simultaneously;

[0267] In this implementation, S7 specifically includes:

[0268] S71. Perform target detection on each frame of video image to identify all possible target vehicles

[0269]

[0270] in, is the i-th target vehicle in the t-th frame image, M t is the total number of target vehicles detected in the t-th frame image, Detector is the target detection algorithm, I t is the t-th frame image;

[0271] S72: Extract feature vectors for each detected target vehicle

[0272]

[0273] in, is the feature vector of the i-th target vehicle in the t-th frame image, and FeatureExtractor is the feature extraction algorithm;

[0274] S73, using the Kalman filter algorithm to predict and update the state of each target vehicle to obtain an updated state estimate

[0275] Prediction steps:

[0276] Update steps:

[0277] in, For the predicted state, is the prediction covariance, is the Kalman gain, is the observation value, H is the observation matrix, F is the state transfer matrix, B is the control matrix, u t is the control input, Q is the process noise covariance, R is the observation noise covariance, and I is the identity matrix;

[0278] S74, using a multi-target tracking algorithm to perform data association, matching the target vehicle in the current frame with the target vehicle in the previous frame, and updating the trajectory of each target vehicle;

[0279]

[0280] Among them, Match is the matching function, c ij is the cost matrix, x ij is the matching result, M t-1 is the total number of target vehicles in the previous frame;

[0281] S75, for the target vehicle that is not matched, initialize a new track and add it to the target vehicle tracking list;

[0282]

[0283] Among them, NewTrajectory is the new trajectory initialization function, Init is the initialization operation, and Update is the update operation.

[0284] S76, managing and updating the trajectories of all target vehicles in real time;

[0285] S8. Realize real-time processing and analysis of video data through algorithm optimization and hardware acceleration, and use edge computing to reduce latency.

[0286] In this implementation, S8 specifically includes:

[0287] S81. Improve the efficiency of video data processing by optimizing algorithms, including parallelizing and pipelining image processing, feature extraction, and target recognition algorithms;

[0288]

[0289] Among them, Process par 、Extract par and Detect par Respectively represent parallel processing of images, feature extraction and target recognition, Time total is the total processing time;

[0290] S82, use GPU and FPGA to accelerate the processing of video data;

[0291] Speed accel =γ·Speed cpu ;

[0292] Among them, Speed accel is the processing speed after hardware acceleration, γ is the acceleration factor, Speed cpu is the CPU processing speed;

[0293] S83, using edge computing architecture to offload data processing tasks to edge devices;

[0294]

[0295] Among them, Latency edge is the total delay under the edge computing architecture, Time edge (i) is the processing time of the i-th edge device, Time cloud (j) is the jth cloud processing time, N is the number of edge devices, and M is the number of cloud servers;

[0296] S84, process the video data through edge devices and perform analysis and processing locally;

[0297]

[0298] Among them, LocalProcessing edge For local processing on edge devices, Process local (V k ) is the time for the kth video frame to be processed locally, and K is the total number of video frames processed;

[0299] S85, combining edge computing and cloud computing, sending computing tasks and big data analysis tasks to the cloud for processing;

[0300] HybridProcessing=α·EdgeProcessing+β·CloudProcessing;

[0301] Among them, HybridProcessing is the hybrid computing mode, EdgeProcessing is the processing capacity of edge computing, CloudProcessing is the processing capacity of cloud computing, α and β are the weight coefficients of edge computing and cloud computing;

[0302] S86. Optimize data transmission protocols and network architecture;

[0303]

[0304] Among them, Transmission opt For optimized data transmission delay, Data size is the amount of data transmitted, Bandwidth is the network bandwidth, Overhead is the transmission overhead;

[0305] S87. Feedback the processing results to the traffic management system or the automatic driving system to achieve monitoring and path optimization of the target vehicle.

[0306] Embodiment 1:

[0307] In order to alleviate the increasingly serious traffic congestion problem, the traffic management department of a large city decided to introduce a method of target vehicle recognition and path tracking in videos based on big data analysis. Traffic accidents and congestion frequently occur on the main roads and important intersections in the city. Traditional manual monitoring and simple image processing technology can no longer meet the needs of efficient and accurate traffic management. Therefore, the use of advanced technical means to comprehensively monitor and analyze urban traffic has become the key to improving the efficiency and accuracy of traffic management.

[0308] High-resolution cameras are installed on the city's main traffic arteries and important intersections, which can collect video data around the clock. The camera installation height and angle are carefully designed to ensure maximum coverage of areas with dense traffic flow. Video data is transmitted to the data processing center via a high-speed network for real-time analysis.

[0309] First, the collected video data is preprocessed, including image enhancement, noise removal, and feature extraction. This step improves the quality of the video image through advanced image processing algorithms and ensures the accuracy of subsequent processing. Next, the preprocessed video data is analyzed using the Visual Transformer to identify the target vehicle in the video. The Visual Transformer uses a deep learning algorithm that can accurately identify the location and category of each vehicle in complex backgrounds and various vehicle types.

[0310] After identifying the target vehicle, the system uses recurrent neural networks and long short-term memory networks to track the target vehicle's movement path in real time. These network models can capture the vehicle's movement pattern, accurately predict the vehicle's next move, and update the vehicle's movement trajectory in real time. In order to improve the accuracy of path prediction, the system also uses generative adversarial networks and reinforcement learning algorithms, combined with historical traffic flow data and current traffic conditions, to make high-precision predictions of the vehicle's future path.

[0311] In addition, the system uses big data processing technology to perform distributed storage, processing and analysis of massive video data from multiple cameras. Through the distributed computing framework, the system can process and analyze a large amount of video data in a short period of time, ensuring that traffic management departments can grasp the overall situation of urban traffic in real time.

[0312] In order to cope with complex traffic environments, the system has designed a multi-target tracking algorithm that can identify and track multiple target vehicles at the same time. The multi-target tracking algorithm uses Kalman filtering and Hungarian algorithm to stably and accurately track each vehicle even when the vehicle is blocked or the vehicle appearance changes, ensuring the reliability of the system.

[0313] The real-time processing capability of the system has been significantly improved through algorithm optimization and hardware acceleration technologies (such as GPU and FPGA). The introduction of edge computing architecture enables some data processing tasks to be completed locally, reducing data transmission delays and further improving the real-time response capability of the system.

[0314] During the implementation process, three major traffic intersections in the city center were selected as pilot sites, and a total of 50 high-resolution cameras were installed. The data processing center is equipped with a high-performance server cluster and multiple GPU acceleration modules to ensure efficient data processing.

[0315] Experimental data show that during peak traffic hours (7am to 9am and 5pm to 7pm), the system can process video data from 50 cameras in real time and complete preprocessing, vehicle identification, path tracking and path prediction for each frame of video in less than 1 second. Compared with traditional manual monitoring, the system's processing efficiency has increased by nearly 10 times.

[0316] In actual application, the system successfully identified and tracked more than 5 million vehicle movement trajectories within a month. The recognition accuracy rate reached 98.7%, and the average error of path prediction was less than 5 meters. The system can detect traffic congestion and traffic accidents in real time, and issue an alarm within 5 seconds after a traffic incident occurs, notifying traffic management personnel to intervene.

[0317] Through optimized management of traffic flow, the pilot operation has significantly improved traffic flow in the city center. Traffic congestion time has been reduced by 15%, and the traffic accident rate has been reduced by 20%. The punctuality rate of public transportation has increased by 12%, greatly improving the overall efficiency of urban transportation.

[0318] In order to verify the system's multi-target management capabilities, multiple tests were conducted in complex traffic environments. In the tests, the system was able to identify and track up to 200 vehicles at the same time, and the Kalman filter and Hungarian algorithm maintained high stability and accuracy under occlusion and lighting changes.

[0319] Table 1: System performance and pilot operation results

[0320] index Traditional systems Method of the present invention Improvement rate Video data processing efficiency 50 fps 500 fps +900％ Recognition accuracy 75％ 98.7％ +23.7％ Average processing time 10 seconds / frame 1 second / frame -90％ Real-time responsiveness 30 seconds 5 seconds -83.3％ Average path prediction error 15 meters Less than 5 meters -66.7％ Traffic congestion time 100 hours / month 85 hours / month -15％ Traffic accident rate 50 times / month 40 times / month -20％ Public transport punctuality 70％ 82％ +12％ Number of vehicle identification 3,000,000 times 5,000,000 times +66.7％

[0321] Table 2: System resource usage and event response time

[0322]

[0323]

[0324] From the table data, we can see that the method of the present invention shows significant advantages in many aspects. The video data processing efficiency is increased to 500 frames per second, which is 900% higher than the traditional system, the recognition accuracy is increased to 98.7%, the average processing time is reduced to 1 second per frame, the real-time response capability is enhanced to 5 seconds, and the average error of path prediction is reduced to less than 5 meters. In terms of traffic management effect, after the pilot operation, traffic congestion time was reduced by 15%, the traffic accident rate was reduced by 20%, the public transportation punctuality rate was increased by 12%, and the number of vehicles identified and tracked increased to 5,000,000 times, an increase of 66.7%. In addition, the method of the present invention has also significantly improved the efficiency of hardware resource utilization, with GPU utilization rate reaching 85% and edge computing device utilization rate reaching 75%. In terms of event response time, the response time of traffic congestion detection, traffic accident detection and vehicle path deviation detection was shortened to 5 seconds, 10 seconds and 8 seconds respectively. These data comprehensively demonstrate the significant advantages of the method of the present invention in terms of processing efficiency, recognition accuracy, real-time response capability, path prediction accuracy, system resource utilization efficiency, and event response time, providing a solid foundation for the development of intelligent traffic management and autonomous driving technology.

[0325] The method for target vehicle recognition and path tracking in videos based on big data analysis proposed in the present invention successfully solves the shortcomings of the existing technology in processing efficiency, recognition accuracy, real-time and multi-target management capabilities, etc. by efficiently processing massive video data, accurately identifying and tracking target vehicles, predicting vehicle paths in real time and optimizing traffic flow management, and provides a solid foundation for the development of intelligent traffic management and autonomous driving technology.

[0326] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for target vehicle identification and path tracking in video based on big data analysis, characterized in that: The following steps are involved: S1. Collect video data through cameras installed on traffic arteries or vehicles; S2, preprocessing the collected video data, including image enhancement, noise removal and key feature extraction; S3, using visual Transformer to analyze the preprocessed video data and identify the target vehicle in the video; S4, using recurrent neural network and long short-term memory network to track the movement path of the target vehicle in real time; S5. Use generative adversarial networks and reinforcement learning algorithms to predict the path of the target vehicle; S6. Use big data processing technology to perform distributed storage, processing and analysis of massive video data; S7, identify and track multiple target vehicles simultaneously; S8. Realize real-time processing and analysis of video data through algorithm optimization and hardware acceleration, and use edge computing to reduce latency.

2. The method for target vehicle identification and path tracking in video based on big data analysis according to claim 1 is characterized in that: The S1 step includes: S11. Install cameras on major roads or vehicles. The camera locations and angles should be able to cover the target area. S12, use a high-resolution camera with a resolution of R×C, where R represents the number of rows and C represents the number of columns; S13, configure the frame rate of the camera to be f, and the frame rate range is f∈[15,60] frames per second; S14, transmitting the collected video data to a data processing center via a network, using TCP / IP, UDP and other protocols suitable for video transmission; Among them, P is the data transmission efficiency, M is the number of data packets, and B j is the data volume of the jth packet, T j is the transmission time, S j is the network bandwidth, N j For noise interference; S15, real-time monitoring and maintenance of the working status of the camera; Among them, R is the reliability evaluation of the camera's working status, W is the number of time periods in the monitoring cycle, and U m is the camera working status in the mth time period, V m It is an indicator of environmental complexity.

3. The method for target vehicle identification and path tracking in video based on big data analysis according to claim 1 is characterized in that: The S2 step includes: S21, performing image enhancement on the collected video data; Among them, I enhanced is the enhanced image, I original is the original image, μ and σ are the mean and standard deviation of the original image, μ′ and σ′ are the target mean and standard deviation, respectively; S22, performing noise elimination and smoothing the image using a Gaussian filter; Among them, I denoised (x, y) is the denoised image, I(x, y) is the original image pixel value, σ is the standard deviation of the Gaussian filter, and k is the radius of the filter; S23, extracting key feature points in the image, and performing feature extraction using a scale-invariant feature transformation algorithm; D(x,y,σ)=(L(x,y,kσ)-L(x,y,σ)); Where D(x, y, σ) is the difference in scale space, L(x, y, σ) is the Gaussian blurred image of the image at scale σ, and k is the scale multiple; S24, performing inter-frame difference on the pre-processed video frames to detect motion areas; M t (x,y)=|I t (x,y)-I t-1 (x,y)|; Among them, M t (x, y) is the difference image between the t-th frame and the t-1-th frame, I t (x, y) is the image of the tth frame, I t-1 (x,y) is the image of the t-1th frame; S25, performing morphological processing, dilating and corroding the motion area; I morph =(I⊕B)°B; Among them, I morph is the image after morphological processing, I is the input binary image, ⊕ is the dilation operation, ° is the erosion operation, and B is the structural element.

4. The method for target vehicle identification and path tracking in video based on big data analysis according to claim 1 is characterized in that: The S3 step includes: S31, converting the preprocessed video data into an image sequence {I t }, where t represents the time frame; {I t }={I1,I2,…,I n }; Among them, I t is the t-th frame image, n is the total number of frames; S32, use the visual Transformer model to extract features from the image sequence, and transform each frame image I t Mapped to high-dimensional feature space, we get the feature vector F t ; S33, all the feature vectors of the frames {F t } Input into the multi-head self-attention mechanism to calculate the similarity between each feature vector; Among them, A i,j is the attention weight between the i-th frame and the j-th frame, Q i and K j are the query vector and key vector of the i-th frame and j-th frame respectively; S34, weighted summation through the attention mechanism, to obtain the weighted feature representation Z of each frame image t ; Among them, Z t is the weighted feature representation of the t-th frame image, V j is the value vector of the jth frame; S35, weighted feature representation Z t Input into the classifier to identify the target vehicle in the video; S36, performing time series analysis on the classification results of all frames to determine the position and motion trajectory of the target vehicle in the video; Among them, Trajectory is the trajectory analysis function, which converts the classification results Mapped to the motion trajectory of the target vehicle.

5. The method for target vehicle identification and path tracking in video based on big data analysis according to claim 1 is characterized in that: The S4 step includes: S41, the recognition result of the target vehicle As an input sequence, it is input into the recurrent neural network; Among them, h t is the classification result of the t-th frame image, h t is the hidden state of the recurrent neural network at the tth time step, σ is the activation function, W h and U h is the weight matrix, b h is the bias vector; S42. Use the long short-term memory network LSTM to further process the hidden state to capture long-term dependencies and obtain the updated hidden state Among them, i t 、f t , o t are the activation vectors of the input gate, forget gate, and output gate, respectively, c t is the cell state, ⊙ is the element product, W i ,W f ,W o ,W c and U i ,U f ,U o ,U c is the weight matrix, b i ,b f ,b o ,b c is the bias vector; S43, output o through long short-term memory network t Calculate the position prediction of the target vehicle at time step t; Among them, t is the position prediction, W o2 is the weight matrix, b o2 is the bias vector, λ is the adjustment coefficient, α j is the attention weight, m ​​is the number of past time steps; S44. Prediction results for all time steps Perform sequence analysis to form the target vehicle's motion path; Among them, Path is the moving path of the target vehicle; S45, smoothing the path according to actual conditions; Among them, w t is the smoothing weight, η is the adjustment coefficient, K is the size of the smoothing window, β k is the smoothing coefficient; S46, finally obtaining the real-time motion path of the target vehicle in the video; Among them, γ is the application coefficient, δ is the adjustment coefficient, P is the historical path length, μ p is the weight of the historical path.

6. The method for target vehicle identification and path tracking in video based on big data analysis according to claim 1, characterized in that: The S5 step includes: S51, taking the historical movement path Path of the target vehicle as input data, and inputting it into the generator of the generative adversarial network GAN; S52, using the discriminator of the generative adversarial network GAN to evaluate the generated path and calculate the difference between the generated path and the true path; D(Path,z t )=σ(W d ·z t +b d ); Among them, D is the discriminator function, σ is the activation function, and W d is the weight matrix, b d is the bias vector; S53, through adversarial training, the path generated by the generator is made close to the real path; Among them, L GAN is the loss function of GAN, p data is the data distribution of the real path, p z is the latent variable distribution of the generator; S54, the potential variable z generated by the generator t Input into the reinforcement learning algorithm for path optimization and prediction; Among them, π(a t |s t ) is the policy function, a t For state s t The action selected next, Q(s t ,a t ) is the action value function, A is the action space; S55. Update the action value function Q using the value iteration method in reinforcement learning; Among them, r t is the reward obtained at time step t, γ is the discount factor, s t+1 is the state at time step t+1. S56. Finally, an optimized target vehicle path prediction is obtained.

7. The method for target vehicle identification and path tracking in video based on big data analysis according to claim 1 is characterized in that: The S6 step includes: S61, dividing the collected video data into a plurality of data blocks, each data block containing a certain number of frames; Among them, D i is the i-th data block, F i,j is the jth frame of the ith data block, T i is the time length of the i-th data block, Δt is the time interval between each frame, and n i is the number of frames in the i-th data block; S62, storing the video data blocks on multiple storage nodes using a distributed file system; S i =HDFS({F i,j }),for i=1,2,…,m; Among them, S i is the data block stored on the i-th node, HDFS is a distributed file system function that stores all frames in the i-th data block on multiple nodes; S63, using a distributed computing framework to perform parallel processing and analysis on the video data blocks; Among them, SparkProcess is a distributed computing process, φ is a function for processing each frame of an image, and each frame of an image in the i-th data block is processed to obtain the processing result R i ; S64, performing data fusion on the processing results to obtain a global analysis result; Among them, Fuse is the data fusion function, ω i is the weight of the ith data block, R global It is the global analysis result; S65. Apply big data processing technology to further aggregate and summarize the global analysis results; I=Aggregate(R global )=Ψ(R global ); Among them, I is the extracted information, Aggregate is the aggregation function, Ψ is the aggregation operation, and the global analysis result R global Further processing and summarization are performed to extract useful information I.

8. The method for target vehicle identification and path tracking in video based on big data analysis according to claim 1 is characterized in that: The S7 step includes: S71. Perform target detection on each frame of video image to identify all possible target vehicles in, is the i-th target vehicle in the t-th frame image, M t is the total number of target vehicles detected in the t-th frame image, Detector is the target detection algorithm, I t is the t-th frame image; S72: Extract feature vectors for each detected target vehicle in, is the feature vector of the i-th target vehicle in the t-th frame image, and FeatureExtractor is the feature extraction algorithm; S73: Use the Kalman filter algorithm to predict and update the state of each target vehicle to obtain an updated state estimate. Prediction Steps: Update steps: in, For the predicted state, is the prediction covariance, is the Kalman gain, is the observation value, H is the observation matrix, F is the state transfer matrix, B is the control matrix, u t is the control input, Q is the process noise covariance, R is the observation noise covariance, and I is the identity matrix; S74, using a multi-target tracking algorithm to perform data association, matching the target vehicle in the current frame with the target vehicle in the previous frame, and updating the trajectory of each target vehicle; Among them, Match is the matching function, c ij is the cost matrix, x ij is the matching result, M t-1 is the total number of target vehicles in the previous frame; S75, for the target vehicle that is not matched, initialize a new track and add it to the target vehicle tracking list; Among them, NewTrajectory is the new trajectory initialization function, Init is the initialization operation, and Update is the update operation; S76. Manage and update the trajectories of all target vehicles in real time.

9. The method for target vehicle identification and path tracking in video based on big data analysis according to claim 1, characterized in that: The S8 step includes: S81. Improve the efficiency of video data processing by optimizing algorithms, including parallelizing and pipelining image processing, feature extraction, and target recognition algorithms; Among them, Process par 、Extract par and Detect par Respectively represent parallel processing of images, feature extraction and target recognition, Time total is the total processing time; S82, use GPU and FPGA to accelerate the processing of video data; Speed accel =γ·Speed cpu ; Among them, Speed accel is the processing speed after hardware acceleration, γ is the acceleration factor, Speed cpu is the CPU processing speed; S83, using edge computing architecture to offload data processing tasks to edge devices; Among them, Latency edge is the total delay under the edge computing architecture, Time edge (i) is the processing time of the i-th edge device, Time cloud (j) is the jth cloud processing time, N is the number of edge devices, and M is the number of cloud servers; S84, process the video data through edge devices and perform analysis and processing locally; Among them, LocalProcessing edge For local processing on edge devices, Process local (V k ) is the time for the kth video frame to be processed locally, and K is the total number of video frames processed; S85, combining edge computing and cloud computing, sending computing tasks and big data analysis tasks to the cloud for processing; HybridProcessing=α·EdgeProcessing+β·CloudProcessing; Among them, HybridProcessing is the hybrid computing mode, EdgeProcessing is the processing capacity of edge computing, CloudProcessing is the processing capacity of cloud computing, α and β are the weight coefficients of edge computing and cloud computing; S86. Optimize data transmission protocols and network architecture; Among them, Transmission opt For optimized data transmission delay, Data size is the amount of data transmitted, Bandwidth is the network bandwidth, Overhead is the transmission overhead; S87. Feedback the processing results to the traffic management system or the automatic driving system to achieve monitoring and path optimization of the target vehicle.

Images