A vehicle and pedestrian recognition system and method based on image recognition

Abstract

The present invention relates to the field of target recognition technology, specifically a vehicle and pedestrian recognition system and method based on image recognition, the system comprising: a video frame analysis module obtains a video stream based on a vehicle-mounted camera, compares the image content of multiple consecutive frames, and recognizes dynamic objects by analyzing the movement between pixels. In the present invention, accurate recognition of dynamic objects and calculation of their speed and direction are achieved by deeply analyzing the movement of pixels in consecutive frames in the video stream, significantly improving the accurate generation of data, real-time update of decisions ensures dynamic adaptation of the decision-making process, and continuously optimizes decision rules to adapt to the ever-changing environment, effectively improving the accuracy of target classification, and the real-time adjustment of neural network parameters enhances the system's responsiveness to dynamic targets, thereby improving the efficiency and reliability of the overall recognition system, and image enhancement technology optimizes the quality of traffic scene images, providing clearer and more accurate visual information for autonomous driving technology.

Description

Technical Field

[0001] The present invention relates to the field of target recognition technology, and in particular to a vehicle and pedestrian recognition system and method based on image recognition. Background Art

[0002] Object recognition technology, a key branch of computer vision, involves the detection, classification, and tracking of specific objects (such as vehicles, pedestrians, or other objects) in images or videos. This technology uses deep learning models to analyze visual data and automatically identify and locate target objects in images. It is widely used in security surveillance, self-driving cars, intelligent transportation systems, robotic navigation, and other fields. The development of object recognition technology relies on algorithm optimization, increased computing power, and the creation of large-scale image datasets. Convolutional neural networks (CNNs) are one of the most commonly used models.

[0003] Vehicle and pedestrian recognition systems (VPRs) are technology systems installed on vehicles that detect and identify other vehicles and pedestrians in their immediate vicinity in real time. The primary purpose of these systems is to improve road safety and reduce traffic accidents, particularly in complex traffic environments. They capture visual information of the surrounding environment through cameras and then analyze it using object recognition technology to alert the driver or trigger appropriate actions within the autonomous driving system. Such systems are crucial for achieving fully autonomous vehicles and improving the safety of existing vehicles.

[0004] When dealing with dynamic objects in complex traffic environments, existing technologies are often limited by preset models and fixed parameters, lacking sufficient adaptability and flexibility. In rapidly changing scenarios, this limitation may result in the inability to promptly identify sudden obstacles or pedestrians with abnormal behavior, increasing the risk of traffic accidents. In addition, the accuracy of image recognition is limited by the quality of the input image. For example, performance degrades under conditions such as low resolution or poor lighting, further affecting the overall performance of the system. These technical limitations are particularly significant in demanding autonomous driving scenarios, restricting the widespread application of the technology and user acceptance. Therefore, strengthening the ability to quickly adapt to dynamic environments and improving the quality of image processing are key directions. The lack of these capabilities will limit the effectiveness of the system in complex environments. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a vehicle and pedestrian recognition system and method based on image recognition.

[0006] In order to achieve the above-mentioned object, the present invention adopts the following technical solution: A vehicle and pedestrian recognition system based on image recognition includes:

[0007] The video frame analysis module obtains video streams from the vehicle camera, compares the image content of multiple consecutive frames, identifies dynamic objects by analyzing the movement between pixels, calculates the speed and direction of moving targets, and generates preliminary moving target data;

[0008] The real-time updating decision module extracts the position and speed information of the target from the preliminary moving target data, updates the node information of the decision tree in real time, re-evaluates the target category by comparing the current input data with the existing classification results, optimizes the node decision rules, and generates decision tree adjustment data;

[0009] The neural network fine-tuning module uses the decision tree to adjust the data, identify classification errors and tracking errors, adjust the target weights and parameters in the neural network, update the network to enhance its responsiveness to dynamic targets, and generate fine-tuned model parameters;

[0010] The image enhancement and generation module adjusts the input parameters of the conditional GAN, uses the fine-tuned model parameters to analyze the target traffic scene, optimizes the image quality during the generation process, and generates an enhanced traffic scene image.

[0011] As a further solution of the present invention, the step of obtaining the preliminary moving target data is specifically as follows:

[0012] Through the video stream captured by the vehicle camera, multiple frames f1,f2,…,f are continuously extracted. n , compare consecutive frames f t and f t+1 The pixel position in , using the formula:

[0013] Δx=x t+1 -x t

[0014] and

[0015] Δy=y t+1 -y t

[0016] Calculate pixel point movement vectors and generate a pixel point movement vector set;

[0017] Among them, f t is the tth frame, x t 、y t is the coordinate of a pixel in the tth frame, Δx and Δy are the distance from the pixel to f t to f t+1 The amount of movement, the difference in the x and y directions;

[0018] Using the pixel point motion vector set, the modulus is calculated for each vector and the speed is estimated using the formula:

[0019]

[0020] The combined effects of speed and direction changes are integrated to obtain a set of velocity vectors;

[0021] Among them, v i is the velocity estimate of the i-th pixel, Δx i , Δy i are the x and y components of the i-th motion vector;

[0022] Based on the set of velocity vectors, the direction angle is calculated for each vector using the formula:

[0023]

[0024] Generate a set of direction angles;

[0025] Among them, θ i is the direction angle of movement of the i-th pixel, Δx i , Δy i are the components of the motion vector;

[0026] Combine the direction angle set with the velocity vector set using the formula:

[0027]

[0028] generating preliminary moving target data;

[0029] in, is the average dynamic vector of the integrated reference speed and direction, v i is the velocity, θ i is the direction angle.

[0030] As a further solution of the present invention, the step of obtaining the decision tree adjustment data is specifically as follows:

[0031] Extract the position and speed information of each target from the preliminary moving target data, and calculate the position and speed information from the vector data by vector decomposition method. and speed Using the formula:

[0032]

[0033] and

[0034]

[0035] Generate position and velocity information

[0036] Among them, x i 、y i Represents the position coordinate, Δxi , Δy i represents the components of the velocity vector;

[0037] Using the position and speed information, update the node information of the decision tree, compare the current input data with the existing classification results, and use the formula:

[0038]

[0039] Calculate and generate decision comparison result R;

[0040] Among them, α z is the normalization coefficient, w p 、w s are the weight coefficients of position and velocity respectively;

[0041] According to the decision comparison result R, the new category of the target is evaluated and the decision node rules are optimized using the formula:

[0042]

[0043] Recalculate and obtain the adjusted node information D;

[0044] Among them, β z , γ z is the adjustment factor;

[0045] Combined with the adjusted node information, the structure of the entire decision tree is optimized using the formula:

[0046] T=D+∑adjust(D i )

[0047] Generate decision tree to adjust data T;

[0048] Among them, D i Indicates the adjustment information of the associated node.

[0049] As a further solution of the present invention, the step of obtaining the fine-tuned model parameters is specifically as follows:

[0050] To identify misclassification and mistracking from the decision tree training data, the formula:

[0051]

[0052] Calculate the error vector, compare the actual output with the expected output, and generate the error recognition result E;

[0053] Among them, T expected Represents the expected output;

[0054] The error recognition results are used to adjust the target weights and parameters in the neural network using the formula:

[0055]

[0056] Generate adjusted neural network parameters Θ new ;

[0057] Among them, Θ old represents the current neural network parameter set, η x represents the learning rate, Represents the gradient of the error E;

[0058] Using the adjusted neural network parameters, the network is updated to enhance its responsiveness to dynamic targets using the formula:

[0059]

[0060] Generate fine-tuned model parameters Θ final ;

[0061] Among them, δ x represents the adjustment factor, adjust(Θ new ) represents the new parameter Θ new Iterative adjustment strategy, |Θ new | represents the absolute value of the new parameter.

[0062] As a further solution of the present invention, the step of acquiring the enhanced traffic scene image is specifically as follows:

[0063] By adjusting the input parameters of the conditional generative adversarial network, the formula is used:

[0064] P new =P old +α g ·(Θ final -P old )

[0065] Calculate the input parameter update amount, update the input parameter, and generate a new input parameter P new ;

[0066] Among them, P old represents the original input parameters, α g represents the learning rate;

[0067] The target traffic scene is analyzed using the new input parameters to optimize the image quality during the generation process using the formula:

[0068]

[0069] Convert the input parameters into image quality indicators and generate the optimized image quality indicator Q;

[0070] Image generation is performed according to the image quality index using the formula:

[0071]

[0072] Calculate the image generation value and generate the enhanced traffic scene image I.

[0073] A vehicle and pedestrian recognition method based on image recognition is implemented based on the above-mentioned vehicle and pedestrian recognition system based on image recognition, and includes the following steps:

[0074] S1: The on-board camera collects video streams, compares the image content of multiple consecutive frames, identifies dynamic objects by analyzing the movement speed and direction between pixels, and calculates the object's motion parameters to establish preliminary dynamic target data;

[0075] S2: extracting position and speed information from the preliminary dynamic target data, updating nodes of the decision tree in real time, re-evaluating target categories, optimizing node decision rules, and obtaining adjusted decision data;

[0076] S3: Analyze and identify classification errors and tracking errors using the adjusted decision data, adjust target weights and parameters in the neural network, update the network's responsiveness, and obtain fine-tuned network parameters;

[0077] S4: using the fine-tuned network parameters, adjusting the input parameters of the conditional generative adversarial network, analyzing the target traffic scene, optimizing the image generation process, and generating an enhanced traffic scene image;

[0078] S5: Analyze the enhanced traffic scene image, identify vehicles and pedestrians therein, output the position and speed information of the vehicles and pedestrians, analyze environmental factors, and form a target recognition result.

[0079] Compared with the prior art, the advantages and positive effects of the present invention are:

[0080] This invention achieves precise identification of dynamic objects and calculation of their speed and direction through in-depth analysis of pixel movement across consecutive frames of a video stream, significantly improving the accuracy of data generation. Real-time decision updates ensure dynamic adaptation of the decision-making process, continuously optimizing decision rules to adapt to changing environments and effectively improving target classification accuracy. Instant adjustment of neural network parameters enhances the system's responsiveness to dynamic targets, thereby improving the overall efficiency and reliability of the recognition system. Image enhancement technology optimizes the quality of traffic scene images, providing clearer and more accurate visual information for autonomous driving technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 is a system flow chart of the present invention;

[0082] Figure 2 This is a flow chart of the steps for obtaining preliminary mobile target data of the present invention;

[0083] Figure 3 A flow chart of the steps for obtaining decision tree adjustment data of the present invention;

[0084] Figure 4 Flowchart of the steps for obtaining model parameters after fine-tuning of the present invention;

[0085] Figure 5 This is a flow chart of the steps for acquiring enhanced traffic scene images according to the present invention. DETAILED DESCRIPTION

[0086] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0087] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0088] Example 1

[0089] See also Figure 1 The present invention provides a technical solution: a vehicle and pedestrian recognition system based on image recognition includes:

[0090] The video frame analysis module obtains video streams from the vehicle camera, compares the image content of multiple consecutive frames, identifies dynamic objects by analyzing the movement between pixels, calculates the speed and direction of moving targets, and generates preliminary moving target data;

[0091] The real-time update decision module extracts the target's position and speed information from the preliminary moving target data, updates the node information of the decision tree in real time, re-evaluates the target category by comparing the current input data with the existing classification results, optimizes the node decision rules, and generates decision tree adjustment data;

[0092] The neural network fine-tuning module uses decision trees to adjust data, identify classification errors and tracking errors, adjust the target weights and parameters in the neural network, update the network to enhance its responsiveness to dynamic targets, and generate fine-tuned model parameters;

[0093] The image enhancement and generation module adjusts the input parameters of the conditional GAN, uses the fine-tuned model parameters to analyze the target traffic scene, optimizes the image quality during the generation process, and generates enhanced traffic scene images.

[0094] The preliminary moving target data is specifically dynamic object data, the decision tree adjustment data includes node information update, target category re-evaluation, and decision rule optimization, and the fine-tuned model parameters include weight adjustment and parameter update.

[0095] See also Figure 2 ,The steps for acquiring preliminary moving target data are as follows:

[0096] Through the video stream captured by the vehicle camera, multiple frames f1,f2,…,f are continuously extracted. n , compare consecutive frames f t and f t+1 The pixel position in , using the formula:

[0097] Δx=x t+1 -x t

[0098] and

[0099] Δy=y t+1 -y t

[0100] Calculate pixel point movement vectors and generate a pixel point movement vector set;

[0101] Among them, f t is the tth frame, x t 、y t is the coordinate of a pixel in the tth frame, Δx and Δy are the distance from the pixel to f t to f t+1 The amount of movement, the difference in the x and y directions;

[0102] Using the pixel point motion vector set, the modulus of each vector is calculated and the speed is estimated using the formula:

[0103]

[0104] The combined effects of speed and direction changes are integrated to obtain a set of velocity vectors;

[0105] Among them, v i is the velocity estimate of the i-th pixel, Δx i , Δyi are the x and y components of the i-th motion vector;

[0106] Based on the velocity vector set, the direction angle is calculated for each vector using the formula:

[0107]

[0108] Generate a set of direction angles;

[0109] Among them, θ i is the direction angle of movement of the i-th pixel, Δx i , Δy i are the components of the motion vector;

[0110] Combine the direction angle set with the velocity vector set using the formula:

[0111]

[0112] generating preliminary moving target data;

[0113] in, is the average dynamic vector of the integrated reference speed and direction, v i is the velocity, θ i is the direction angle.

[0114] Set up to extract two frames of images f1 and f2 from the video stream of the vehicle camera. In f1, the position of a target pixel is (x1, y1) = (100, 150), while in f2, the pixel moves to

[0115] (x2,y2)=(102,152)

[0116] Formula and calculation process:

[0117] Δx=x2-x1=102-100=2

[0118] Δy=y2-y1=152-150=2

[0119] The generated pixel motion vector (Δx, Δy) is (2, 2).

[0120] Next, use the above results to calculate the velocity without introducing unknown coefficients.

[0121] Formula and calculation process:

[0122]

[0123] Substituting Δx=2 and Δy=2,

[0124]

[0125] The speed v represents the speed at which a pixel moves from one frame to the next, and the unit can be pixel / frame.

[0126] Based on the velocity vector calculated above, the moving direction of the pixel point is calculated next.

[0127] Formula and calculation process:

[0128]

[0129] Substitute Δx i =2 and Δy i =2

[0130]

[0131] Direction angle θ i Indicates the direction of pixel movement in degrees.

[0132] Finally, the speed and direction are combined to generate preliminary moving target data.

[0133] Formula and calculation process:

[0134]

[0135] For a single pixel example, let N = 1, v1 = 2.83, θ1 = 45°,

[0136]

[0137] Indicates the average motion trend of pixels. The value means that the object moves approximately 2 pixels toward the upper right corner per frame.

[0138] See also Figure 3 , the specific steps for obtaining decision tree adjustment data are:

[0139] Extract the position and speed information of each target from the preliminary moving target data, and calculate the position and speed of each target from the vector data by vector decomposition method. and speed Using the formula:

[0140]

[0141] and

[0142]

[0143] Generate position and velocity information

[0144] Among them, x i 、y iRepresents the position coordinate, Δx i , Δy i represents the components of the velocity vector;

[0145] Using the position and speed information, update the node information of the decision tree and compare the current input data with the existing classification results using the formula:

[0146]

[0147] Calculate and generate decision comparison result R;

[0148] Among them, α z is the normalization coefficient, w p 、w s are the weight coefficients of position and velocity respectively;

[0149] According to the decision comparison result R, the new category of the target is evaluated and the decision node rules are optimized using the formula:

[0150]

[0151] Recalculate and obtain the adjusted node information D;

[0152] Among them, β z , γ z is the adjustment factor;

[0153] Combined with the adjusted node information, the structure of the entire decision tree is optimized using the formula:

[0154] T=D+∑adjust(D i )

[0155] Generate decision tree to adjust data T;

[0156] Among them, D i Indicates the adjustment information of the associated node.

[0157] formula

[0158]

[0159] and

[0160]

[0161] Represents the total position vector of the target, calculated as the sum of the x and y coordinates of all observation positions.

[0162] Represents the total velocity vector of the target, calculated as the sum of Δx and Δy of all observed velocities.

[0163] Calculation example

[0164] Set up within a certain period of time, through the sensor monitoring to obtain four position coordinate points and corresponding speed changes: Position coordinate points: (1,2), (3,4), (5,6), (7,8)

[0165] Speed ​​changes: (0.5, 0.5), (1.0, 1.5), (1.5, 2.0), (2.0, 2.5)

[0166] calculate:

[0167] Position vector: Velocity vector: formula

[0168]

[0169] α z : Normalization coefficient used to adjust the scale of the results.

[0170] w p 、w s : Weight coefficient of position and speed, adjusted according to actual application.

[0171] Calculation example

[0172] Set the weight w p =0.5, w s =0.3, normalization coefficient α z =2:

[0173] calculate:

[0174] R=2(0.5·(16,20)+0.3·(5,6.5))

[0175] R=2((8,10)+(1.5,1.95))=2×(9.5,11.95)=(19,23.9)

[0176] formula

[0177]

[0178] β z , γ z : Adjustment coefficient.

[0179] w p 、w s : Same weighting coefficient used above.

[0180] Calculation example

[0181] Set the adjustment factor β z =1,γ z =0.1:

[0182] calculate:

[0183] D=1·(19,23.9)+0.1(0.5·(16 2 +20 2 )+0.3·(5 2 +6.5 2 ))

[0184] D=(19,23.9)+0.1(0.5·(256+400)+0.3·(25+42.25))

[0185] D=(19,23.9)+0.1(0.5·656+0.3·67.25)

[0186] D=(19,23.9)+0.1(328+20.175)=(19,23.9)+(34.8175,34.8175)

[0187] D=(53.8175,58.7175)

[0188] formula

[0189] T=D+∑adjust(D i )

[0190] D i : Adjustment information of the associated node.

[0191] Calculation example

[0192] Setting D i The adjusted result is (2,3):

[0193] calculate:

[0194] T=(53.8175,58.7175)+(2,3)=(55.8175,61.7175)

[0195] See also Figure 4 , the steps to obtain the model parameters after fine-tuning are as follows:

[0196] Identify misclassification and mistracking from decision tree training data using the formula:

[0197]

[0198] Calculate the error vector, compare the actual output with the expected output, and generate the error recognition result E; where T expected Represents the expected output;

[0199] The error recognition results are used to adjust the target weights and parameters in the neural network using the formula:

[0200]

[0201] Generate adjusted neural network parameters Θ new ;

[0202] Among them, Θ old represents the current neural network parameter set, η x represents the learning rate, Represents the gradient of the error E;

[0203] Using the adjusted neural network parameters, update the network to enhance its responsiveness to dynamic targets using the formula:

[0204]

[0205] Generate fine-tuned model parameters Θ final ;

[0206] Among them, δ x represents the adjustment factor, adjust(Θ new ) represents the new parameter Θ new Iterative adjustment strategy, |Θ new | represents the absolute value of the new parameter.

[0207] formula:

[0208]

[0209] in:

[0210] T: The actual decision tree output, which can be the classification error rate or a target performance indicator.

[0211] E: Indicates an incorrect recognition result.

[0212] Calculation process and examples:

[0213] In a classification task, the actual error rate T of the decision tree is 15%, while the expected error rate T expected是 10%. These two values ​​can be obtained through historical data analysis or pre-set performance targets.

[0214] Calculate T and T expected The absolute value of the difference: |TT expected |=|0.15-0.10|=0.05

[0215] Calculate T expected The square root of :

[0216] These two results are used to calculate the error recognition result:

[0217] illustrate:

[0218] The error recognition result E≈0.158 represents the standardized error relative to the expected performance, which is used to evaluate the deviation between the actual performance of the decision tree and the expectation.

[0219] formula:

[0220]

[0221] in:

[0222] Θ old : Parameters of the original neural network.

[0223] η x : Learning rate, which controls the speed of parameter updates and is often set based on model tuning or previous experiments.

[0224] The gradient of the error E with respect to the parameter θ is calculated by the back-propagation algorithm.

[0225] Calculation process and examples:

[0226] Set the original parameter Θ old =2.0, learning rate η x = 0.1, and the error gradient (This means that reducing the parameter can reduce the error).

[0227] Calculate the gradient adjustment term:

[0228] Update parameter: Θ new =2.0-(-0.0455)=2.0455

[0229] illustrate:

[0230] New parameter value Θ new =2.0455 represents the parameters after one learning iteration, which is used to improve the model performance.

[0231] formula:

[0232]

[0233] in:

[0234] Θ new : Updated parameters.

[0235] δ x : Adjustment factor, used to balance the influence of the adjustment strategy.

[0236] adjust(Θ new ): Based on Θnew Further adjustments of , including additional regularization or optimization steps, are possible.

[0237] Calculation process and examples:

[0238] Setting Θ new =2.0455, adjustment factor δ x =0.02, and set adjust(Θ new )=0.1.

[0239] Calculation adjustments:

[0240] Update parameter: Θ final =2.0455+0.000978=2.046478

[0241] illustrate:

[0242] Parameter Θ final = 2.046478 represents the neural network parameters after all adjustments, aiming to optimize the model's ability to respond to dynamic targets.

[0243] See also Figure 5 ,The steps for acquiring the enhanced traffic scene image are as follows:

[0244] By adjusting the input parameters of the conditional generative adversarial network, the formula is used:

[0245] P new =P old +α g ·(Θ final -P old )

[0246] Calculate the input parameter update amount, update the input parameter, and generate a new input parameter P new ;

[0247] Among them, P old represents the original input parameters, α g represents the learning rate;

[0248] The target traffic scene is analyzed using the new input parameters to optimize the image quality during the generation process using the formula:

[0249]

[0250] Convert the input parameters into image quality indicators and generate the optimized image quality indicator Q;

[0251] Image generation is performed based on image quality indicators using the formula:

[0252]

[0253] Calculate the image generation value and generate the enhanced traffic scene image I.

[0254] formula:

[0255] P new =P old +α g ·(Θ final -P old )

[0256] Parameter explanation and source:

[0257] α g : Learning rate, set to 0.1, controls the speed of parameter update and is obtained through experimental tuning. Θ final : The fine-tuned model parameters are set to 0.8, indicating the new model parameters obtained through training. Calculation process:

[0258] Calculate the parameter change:

[0259] ΔP=Θ final -P old =0.8-0.5=0.3

[0260] Calculate new parameter values:

[0261] P new =P old +α g ΔP=0.5+0.1·0.3=0.53

[0262] formula:

[0263]

[0264] Parameter explanation and source:

[0265] P new : Calculated from step 1, 0.53.

[0266] Q: Image quality index, through P new Calculated.

[0267] Calculation process:

[0268] Calculate the square root:

[0269]

[0270] Compute the negative value of an exponential:

[0271]

[0272] Calculate image quality metrics:

[0273]

[0274] formula:

[0275]

[0276] Parameter explanation and source:

[0277] Q: Calculated from the above results, 0.675.

[0278] I: Enhanced traffic scene image, calculated by Q.

[0279] Calculation process:

[0280] Calculate the square:

[0281] Q 2 =0.675 2 =0.456

[0282] Calculate the denominator:

[0283] 1-Q 2 =1-0.456=0.544

[0284] Calculate the input to the rendering function:

[0285]

[0286] Image representation, set the render function to convert 1.241 to the corresponding image enhancement level.

[0287] A vehicle and pedestrian recognition method based on image recognition is implemented based on the above-mentioned vehicle and pedestrian recognition system based on image recognition, and includes the following steps:

[0288] S1: The on-board camera collects video streams, compares the image content of multiple consecutive frames, identifies dynamic objects by analyzing the movement speed and direction between pixels, and calculates the object's motion parameters to establish preliminary dynamic target data;

[0289] S2: Extract position and speed information from preliminary dynamic target data, update decision tree nodes in real time, re-evaluate target categories, optimize node decision rules, and obtain adjusted decision data;

[0290] S3: Using the adjusted decision data, analyze and identify classification errors and tracking errors, adjust the target weights and parameters in the neural network, update the network's responsiveness, and obtain fine-tuned network parameters;

[0291] S4: Using the fine-tuned network parameters, adjust the input parameters of the conditional generative adversarial network, analyze the target traffic scene, optimize the image generation process, and produce enhanced traffic scene images;

[0292] S5: Analyze and enhance traffic scene images, identify vehicles and pedestrians, output the location and speed information of vehicles and pedestrians, analyze environmental factors, and form target recognition results.

[0293] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A vehicle and pedestrian recognition system based on image recognition, characterized in that: The system comprises: The video frame analysis module obtains the video stream based on the vehicle-mounted camera, compares the image content of multiple consecutive frames, identifies dynamic objects by analyzing the movement between pixels, calculates the speed and direction of the moving target, and generates preliminary moving target data; The real-time updating decision module extracts the position and speed information of the target from the preliminary moving target data, updates the node information of the decision tree in real time, re-evaluates the target category by comparing the current input data with the existing classification results, optimizes the node decision rules, and generates decision tree adjustment data; The steps for obtaining the decision tree adjustment data are specifically as follows: Extract the position and speed information of each target from the preliminary moving target data, and calculate the position and speed information of each target from the vector data by vector decomposition method. and speed , using the formula: and Generate position and velocity information ; in, Represents the position coordinates, represents the components of the velocity vector; Using the position and speed information, update the node information of the decision tree, compare the current input data with the existing classification results, and use the formula: Calculate and generate decision comparison results ; in, is the normalization coefficient, are the weight coefficients of position and velocity respectively; Compare the results according to the decision , evaluate the new category of the target and optimize the decision node rules, using the formula: Recalculate the adjusted node information ; in, is the adjustment factor; Combined with the adjusted node information, the structure of the entire decision tree is optimized using the formula: Generate decision tree adjustment data ; in, Indicates the adjustment information of the associated node; The neural network fine-tuning module uses the decision tree to adjust the data, identify classification errors and tracking errors, adjust the target weights and parameters in the neural network, update the network to enhance the responsiveness to dynamic targets, and generate fine-tuned model parameters; The steps for obtaining the model parameters after fine-tuning are specifically as follows: Identify misclassification and mistracking from the decision tree tuning data using the formula: Calculate the error vector, compare the actual output with the expected output, and generate the error identification result ; in, Represents the expected output; The error recognition results are used to adjust the target weights and parameters in the neural network using the formula: Generate tuned neural network parameters ; in, represents the current neural network parameter set, represents the learning rate, Representative error The gradient of Using the adjusted neural network parameters, the network is updated to enhance the responsiveness to dynamic targets using the formula: Generate fine-tuned model parameters ; in, represents the adjustment factor, Represents based on new parameters The iterative adjustment strategy of represents the absolute value of the new parameter; The image enhancement and generation module adjusts the input parameters of the conditional GAN, uses the fine-tuned model parameters to analyze the target traffic scene, optimizes the image quality during the generation process, and generates an enhanced traffic scene image.

2. The vehicle and pedestrian recognition system based on image recognition according to claim 1, characterized in that: The steps for obtaining the preliminary moving target data are specifically as follows: Continuously extract multiple frames from the video stream captured by the vehicle camera , compare consecutive frames and The pixel position in is calculated using the formula: and Calculate pixel point movement vectors and generate a pixel point movement vector set; in, is the tth frame, is the coordinate of a pixel in the tth frame, The pixel is from arrive The amount of movement, the difference in the x and y directions; Using the pixel point movement vector set, the modulus is calculated for each vector and the speed is estimated using the formula: The combined effects of speed and direction changes are considered to obtain a set of velocity vectors; in, is the velocity estimate of the ith pixel, are the x and y components of the ith motion vector; Based on the set of velocity vectors, the direction angle is calculated for each vector using the formula: Generate a set of direction angles; in, is the direction angle of movement of the i-th pixel, are the components of the motion vector; Combine the direction angle set with the velocity vector set using the formula: generating preliminary moving target data; in, is the average dynamic vector of the combined reference speed and direction, It's speed. is the direction angle.

3. The vehicle and pedestrian recognition system based on image recognition according to claim 1, characterized in that: The steps of acquiring the enhanced traffic scene image are specifically as follows: By adjusting the input parameters of the conditional generative adversarial network, the formula is used: Calculate the input parameter update amount, update the input parameter, and generate new input parameters ; in, Represents the original input parameters, represents the learning rate; The target traffic scene is analyzed using the new input parameters to optimize the image quality during the generation process, using the formula: Convert input parameters into image quality indicators and generate optimized image quality indicators ; The image is generated according to the image quality index, using the formula: Calculate the image generation value and generate the enhanced traffic scene image .

4. A vehicle and pedestrian recognition method based on image recognition, characterized in that: The vehicle and pedestrian recognition system based on image recognition according to any one of claims 1 to 3 is implemented, comprising the following steps: The on-board camera collects video streams, compares the image content of multiple consecutive frames, identifies dynamic objects by analyzing the moving speed and direction between pixels, and calculates the motion parameters of the objects to establish preliminary dynamic target data; Extracting position and speed information from the preliminary dynamic target data, updating nodes of the decision tree in real time, re-evaluating target categories, optimizing node decision rules, and obtaining adjusted decision data; Utilizing the adjusted decision data, analyzing and identifying classification errors and tracking errors, adjusting target weights and parameters in the neural network, updating the network's responsiveness, and obtaining fine-tuned network parameters; Using the fine-tuned network parameters, adjusting the input parameters of the conditional generative adversarial network, analyzing the target traffic scene, optimizing the image generation process, and generating an enhanced traffic scene image; The enhanced traffic scene image is analyzed, vehicles and pedestrians therein are identified, the position and speed information of the vehicles and pedestrians are output, environmental factors are analyzed, and a target recognition result is formed.

Images