Target detection method for unmanned driving

Abstract

The invention relates to the technical field of image target detection, in particular to a target detection method for unmanned driving, which comprises the following steps of: collecting image data of a road environment, setting the size of a sensing area of each pixel point according to the complexity of the road environment and the expected size of a target to be recognized, extracting image features in the sensing area, and extracting the image features in the sensing area; comprising spectral intensity, edge direction and texture gradient. According to the method, the size of the sensing area of each pixel point is set, and the image features of the spectral intensity, the edge direction and the texture gradient are extracted in the areas, so that the target features and the surrounding environment are matched more accurately, and the accuracy and the efficiency of target detection are improved. Weight distribution and influence evaluation are carried out based on the features, so that the target recognition model can recognize and predict the potential moving track and future behavior of the target more accurately, and more accurate environment perception capability is provided for the automatic driving vehicle.

Description

Technical Field

[0001] The present invention relates to the technical field of image target detection, and in particular to a target detection method for unmanned driving. Background Art

[0002] Image object detection focuses on automatically identifying and locating target objects in digital images or videos. By parsing visual input, the location and category of different objects in the image are calibrated. Image object detection is widely used in many scenarios, such as security monitoring, system automation, driverless cars and robot navigation. The field involves deep learning, pattern recognition and artificial intelligence technology. By using methods such as convolutional neural networks, the detection accuracy and speed can also be improved.

[0003] Among them, the target detection method used for unmanned driving specifically refers to the application of image target detection technology in unmanned driving systems. In the field of unmanned driving, target detection methods are used to identify and locate key elements such as vehicles, pedestrians, and traffic signs on the road in real time, thereby providing accurate environmental perception information for autonomous driving vehicles. This enables unmanned vehicles to effectively navigate autonomously in complex road environments. Through target detection, unmanned vehicles can make timely decision responses, optimize driving paths, and ensure driving safety.

[0004] However, in actual applications, the existing technology does not accurately recognize small or distant targets in complex road environments because the sensing area of ​​the pixels is not set finely enough. In addition, when encountering changing environmental conditions (such as lighting changes, occlusions, etc.), misjudgments or missed judgments are prone to occur, which affects the navigation and decision-making accuracy of the autonomous driving system. For example, in low-light or rainy conditions, it is difficult to accurately identify pedestrians or traffic signs, thereby increasing the risk of traffic accidents. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a target detection method for unmanned driving.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a target detection method for unmanned driving, comprising the following steps:

[0007] S1: Collect image data of the road environment, set the perception area size of each pixel according to the complexity of the road environment and the expected size of the target to be identified, extract image features within the perception area, including spectral intensity, edge direction and texture gradient, and generate a set of perception features;

[0008] S2: Based on the perceptual feature set, weights are assigned and influence is evaluated for spectral intensity, edge direction and texture gradient features, and classification criteria are established using the features, and feature parameters are introduced to initialize the convolutional neural network to obtain a target recognition model;

[0009] S3: Based on the target recognition model, analyze the road environment image data captured in real time, calculate the recognition probability of the target in the perception area, compare it with the preset threshold, judge the detection accuracy of the target, and calculate the change of the target spatial position to obtain the target detection result;

[0010] S4: Based on the target detection result, combined with the current position and speed information of the vehicle, the potential movement trajectory and future behavior changes of the target are analyzed, and the information is integrated to assist in adjusting the vehicle's driving route to obtain a dynamic target prediction result.

[0011] As a further solution of the present invention, the step of obtaining the size of the sensing area of ​​each pixel is specifically as follows:

[0012] S111: Based on the collected road environment image data, the formula is used:

[0013]

[0014] Calculate the environmental complexity C of the road image;

[0015] Where n is the total number of object categories in the image, x i is the number of objects of the i-th class, w i is the weight factor of the i-th type of object, which is set in advance according to the difficulty of detecting the object type, d i is the average distance from the i-th class of objects to the center of the image;

[0016] S112: Based on the collected road environment image data, the formula is used:

[0017]

[0018] Calculate the expected size S of the target to be identified;

[0019] Where n′ is the number of similar objects in the image, b f is the area of ​​the fth target, h f is the size coefficient of the fth target, indicating the display ratio of the target in the image;

[0020] S113: According to the environmental complexity C of the road image and the expected size S of the target to be identified, the formula is adopted:

[0021]

[0022] Calculate the perception area size G of each pixel;

[0023] Among them, k1 is the adjustment coefficient, which is obtained according to the data in the scene, p1 is the exponential attenuation factor of the environmental complexity, which is determined by experiments, ln is the natural logarithmic function, and v1 is the environmental sensitivity parameter, which is adjusted by driving data.

[0024] As a further solution of the present invention, the step of acquiring the perceptual feature set is specifically:

[0025] S121: According to the perception area size G of each pixel, the formula is used:

[0026]

[0027] Calculate the spectral intensity I;

[0028] Where P(x,y) is the value of pixel P at coordinate (x,y), x and y are the pixel coordinates in the perception area, and the range of x and y is from 1 to G;

[0029] S122: According to the perception area size G of each pixel, the formula is adopted:

[0030]

[0031] Calculate the edge direction D;

[0032] S123: According to the perception area size G of each pixel, the formula is adopted:

[0033]

[0034] Calculate the quantized value T of the texture gradient and integrate all the information to obtain the perceptual feature set;

[0035] Where A is the preset average brightness level.

[0036] As a further solution of the present invention, the steps of obtaining feature weight allocation and influence evaluation are specifically as follows:

[0037] S211: According to the spectral intensity I, the formula is used:

[0038]

[0039] Calculate the weight w of the spectral intensity I I ;

[0040] Among them, α1 is the adjustment coefficient, which is determined by data analysis. is the average value of the spectral intensity, is the variance of the spectral intensity, ∈ is a constant;

[0041] S212: According to the edge direction D, the formula is used:

[0042]

[0043] Calculate the weight w of edge direction D D ;

[0044] Among them, k2 is the adjustment coefficient, which is obtained through data analysis or model tuning process. It is the exponential function part, which makes the weight value between 0 and 1;

[0045] S213: According to the quantized value T of the texture gradient, the formula is adopted:

[0046]

[0047] Calculate the weight w of the quantized value T of the texture gradient T ;

[0048] Among them, w T is the weight of the texture gradient, γ is the adjustment coefficient, which is obtained through data analysis, and CV(T) is the coefficient of variation of the texture gradient.

[0049] As a further solution of the present invention, the step of acquiring the target recognition model is specifically as follows:

[0050] S221: According to the weight w of the spectral intensity I I , the weight w of edge direction D D and the weight w of the quantized value T of the texture gradient T , using the formula:

[0051] B=σ(w I I+w D ·D+w T ·T)

[0052] Define classification criteria B;

[0053] Where σ is the activation function that transforms the linear combination into an output between 0 and 1;

[0054] S222: According to the classification standard B, the formula is used:

[0055]

[0056] Establish a target recognition model M;

[0057] Among them, ReLU is the activation function used to introduce nonlinearity, conv is the convolution operation used to extract features and learn local spatial relationships, and B jIt is the feature map segmented according to B, used for convolution kernel, θ j and b j are the weights and biases of the jth convolution kernel, adjusted by the training data, and J is the number of convolutional layers.

[0058] As a further solution of the present invention, the step of obtaining the recognition probability of the target in the sensing area is specifically as follows:

[0059] S311: According to the target recognition model M, the formula is adopted:

[0060]

[0061] The extracted feature set F t ;

[0062] Among them, M j is the jth layer in the target recognition model M, which performs convolution operations on the image for feature extraction, R t is the real-time road environment image captured at time t;

[0063] S312: According to the feature set F t , using the formula:

[0064]

[0065] Calculate the recognition probability P of the target t ;

[0066] Among them, P t is the target recognition probability calculated at time t, F t,q It is from F t The qth feature in v q is the weight of the qth feature, which is obtained through training optimization, Q is the total number of features, and Z is the normalization factor.

[0067] As a further solution of the present invention, the step of obtaining the target detection result is specifically:

[0068] S321: According to the recognition probability P of the target t , using the formula:

[0069] E t =|P t -τ|

[0070] Calculate the recognition probability P of the target t The deviation E from the preset threshold τ t ;

[0071] S322: According to the recognition probability P of the target t The deviation E from the preset threshold τt , using the formula:

[0072] ΔS t =(P t -P t-1 )·λ1

[0073] Calculate the change in the spatial position of the target at time t, ΔS t ;

[0074] Among them, P t is the target recognition probability at time t, P t-1 is the target recognition probability at time t-1, indicating the recognition probability at the previous moment, and λ1 is an adjustment parameter used to amplify or reduce the change in spatial position caused by the probability change, which is obtained based on experimental data or performance feedback adjustment during the optimization process.

[0075] As a further solution of the present invention, the step of obtaining the dynamic target prediction result is specifically:

[0076] S411: Based on the spatial position change ΔS of the target at time t t , using the formula:

[0077] L t+1 =L t +v t ·Δt+ΔS t

[0078] Calculate the future position L of the target t+1 ;

[0079] Among them, L t is the target position at the current time t, v t is the target speed at the current time t, Δt is the time interval, which represents the time difference from time t to t+1;

[0080] S412: Based on the future position L of the target t+1 , using the formula:

[0081] O new =O current +α2·(L t+1 -O current )

[0082] Calculate the adjusted route O new ;

[0083] Among them, O current is the current vehicle route, which represents the path of the vehicle at the current time point, and α2 is the route adjustment coefficient, which is obtained through experiments or based on feedback from vehicle dynamic control.

[0084] Compared with the prior art, the advantages and positive effects of the present invention are:

[0085] In the present invention, by setting the size of the perception area of ​​each pixel point and extracting the image features of spectral intensity, edge direction and texture gradient in these areas, the target features and the surrounding environment are matched more accurately, thereby improving the accuracy and efficiency of target detection. Based on the weight allocation and influence assessment of the features, the target recognition model can more accurately identify and predict the potential movement trajectory and future behavior of the target, thereby providing more accurate environmental perception capabilities for autonomous driving vehicles. In addition, by comparing the real-time target recognition probability with the preset threshold, the response time and accuracy of target detection are optimized, thereby improving the driving safety and efficiency of the vehicle. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 It is a schematic diagram of the main steps of the present invention;

[0087] Figure 2 A flow chart of the steps for obtaining the size of the sensing area of ​​each pixel in the present invention;

[0088] Figure 3 A flowchart of the steps for obtaining a set of perceptual features of the present invention;

[0089] Figure 4 A flow chart of the steps of obtaining feature weight allocation and influence evaluation according to the present invention;

[0090] Figure 5 A flowchart of the steps for obtaining the target recognition model of the present invention;

[0091] Figure 6 A flow chart of the steps for obtaining the recognition probability of a target within a sensing area according to the present invention;

[0092] Figure 7 A flowchart of the steps for obtaining target detection results of the present invention;

[0093] Figure 8 This is a flow chart of the steps for obtaining dynamic target prediction results of the present invention. DETAILED DESCRIPTION

[0094] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with 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.

[0095] 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 indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0096] See also Figure 1 The present invention provides a technical solution: a target detection method for unmanned driving, comprising the following steps:

[0097] S1: Collect image data of the road environment, set the perception area size of each pixel according to the complexity of the road environment and the expected size of the target to be identified, extract image features within the perception area, including spectral intensity, edge direction and texture gradient, and generate a set of perception features;

[0098] S2: Based on the perceptual feature set, the spectral intensity, edge direction and texture gradient features are weighted and evaluated for their influence. The features are used to establish classification criteria, and the feature parameters are used to initialize the convolutional neural network to obtain the target recognition model.

[0099] S3: Based on the target recognition model, the real-time captured road environment image data is analyzed, the recognition probability of the target in the perception area is calculated, and compared with the preset threshold to determine the detection accuracy of the target, and the change of the target spatial position is calculated to obtain the target detection result;

[0100] S4: Based on the target detection results, combined with the vehicle's current position and speed information, the target's potential movement trajectory and future behavior changes are analyzed, and the integrated information is used to assist in adjusting the vehicle's driving route to obtain dynamic target prediction results;

[0101] The perceptual feature set includes spectral peak data of local areas, directional statistical information of dominant edges, and spatial distribution characteristics of key textures. The target recognition model includes a classification layer for vehicle and pedestrian recognition, an input preprocessing layer based on ambient brightness adjustment, and a specialized convolutional layer for long-distance target detection. The target detection results include real-time detected obstacle existence probability, pedestrian crossing possibility assessment, and traffic sign recognition accuracy. The dynamic target prediction results include impending vehicle lateral movement warnings, predicted pedestrian movement paths, and safe distances from the vehicle in front.

[0102] See also Figure 2, the specific steps for setting the perception area size of each pixel are:

[0103] S111: Based on the collected road environment image data, the formula is used:

[0104]

[0105] Calculate the environmental complexity C of the road image;

[0106] Where n is the total number of object categories in the image, x i is the number of objects of the i-th class, from image analysis, w i is the weight factor of the i-th type of object, which is set in advance according to the difficulty of detecting the object type, d i is the average distance of the i-th class object to the center of the image, in pixels, where i is the index of the object class;

[0107] S112: Based on the collected road environment image data, the formula is used:

[0108]

[0109] Calculate the expected size S of the target to be identified;

[0110] Where n′ is the number of similar objects in the image, which indicates the total number of objects analyzed and taken into account, and b f is the area of ​​the fth target, extracted from the image using image analysis techniques. This parameter is directly related to the actual area occupied by the target. f is the size coefficient of the fth target, which indicates the display ratio of the target in the image, taking into account the relative distance between the target and the camera and the image resolution;

[0111] S113: According to the environmental complexity C of the road image and the expected size S of the target to be identified, the formula is adopted:

[0112]

[0113] Calculate the perception area size G of each pixel;

[0114] Among them, k1 is the adjustment coefficient, which is obtained through data analysis in experiments or actual scenarios and is used to balance the size of the perception area. p1 is the exponential decay factor of environmental complexity, which is determined through experiments and is used to adjust the influence of complexity on the size of the perception area. ln is the natural logarithm function, which is used to deal with the nonlinear effect of C. v1 is the environmental sensitivity parameter, which is adjusted through actual driving data and affects the reaction speed of the perception area to changes in environmental complexity.

[0115] The calculation process is as follows:

[0116] For the formula:

[0117]

[0118] Setting parameters From the actual road traffic image analysis data, there are three types of objects: vehicles, pedestrians, and bicycles, so n = 3, x i The image processing software is used to automatically detect and count the road images. In one image, there are 10 vehicles, 15 pedestrians, 5 bicycles, and 5 vehicles. i According to the previous experimental data settings, vehicles are 1.0, pedestrians are 1.2, bicycles are 1.1, and d i The average distances are vehicles = 50px, pedestrians = 30px, and bicycles = 40px.

[0119]

[0120] C≈0.2+0.564+0.135≈0.899

[0121] For the formula:

[0122]

[0123] The parameter n′ is obtained by image analysis technology. The number of vehicles identified in the image is 10, and b f It is calculated from the image that the average area of ​​each vehicle is 2000 square pixels, h f Calculated by distance, the average is 0.5.

[0124]

[0125] For the formula:

[0126]

[0127] The parameter k1 is set based on experimental data analysis and is set to 1.5, p1 is determined through experiments and is set to 0.5, and v1 is adjusted based on driving data and is set to 2.

[0128]

[0129] G≈1.5·(31.622 / 0.948+0.641 / 2)≈1.5·(33.364+0.319)≈1.5·33.683

[0130] ≈50.533

[0131] The calculated G value is 50.533, which means that the perception area size of each pixel is 50.533 square pixels. This indicates the local sensitivity under the conditions of this environment complexity and target size. Compared with the baseline perception area size (e.g. 50 square pixels), this result shows a slight increase.

[0132] See also Figure 3 , the specific steps for obtaining the perceptual feature set are:

[0133] S121: According to the perception area size G of each pixel, the formula is used:

[0134]

[0135] Calculate the spectral intensity I;

[0136] Where P(x,y) is the value of pixel P at coordinate (x,y), x and y are the pixel coordinates in the perception area, and the range of x and y is from 1 to G;

[0137] S122: According to the perception area size G of each pixel, the formula is used:

[0138]

[0139] Calculate the edge direction D;

[0140] S123: According to the perception area size G of each pixel, the formula is used:

[0141]

[0142] Calculate the quantized value T of the texture gradient and integrate all the information to obtain the perceptual feature set;

[0143] Where A is the preset average brightness level;

[0144] The calculation process is as follows:

[0145] For the formula:

[0146]

[0147] Set the parameter G to 50 (representing that the side length of the perception area is 50 pixels), P(x,y) is extracted directly from the actual image data through image processing software, set a 50×50 pixel area, and all pixel values ​​are uniformly set to 100 (simulating a consistent spectral intensity scene):

[0148]

[0149] For the formula:

[0150]

[0151] Set the parameter G to 50, P(x,y) is obtained by contrasting the difference in the actual image, and the pixel value gradually increases from 10 to 100 on the boundary:

[0152]

[0153] For the formula:

[0154]

[0155] Set the parameter G to 50, set A to 50 calculated based on the image average (the average brightness actually obtained from the image data), and set the pixel values ​​to be randomly distributed between 40 and 60 to simulate a certain texture change:

[0156]

[0157] The mean of the squared differences for each pixel is 100.

[0158] Spectral intensity I = 100 means that all pixels in the perception area have the same intensity value of 100. This consistency indicates that the illumination or color distribution in the area is very uniform, without obvious spectral changes, and is suitable for identifying large-area monochromatic objects or stable light sources. Edge intensity D = 90 shows that at the edge of the perception area, the pixel value increases from 10 to 100, and the high value of edge intensity (90) reflects that there is a significant pixel change in this area. This significant change indicates that the perception area contains obvious edge features, which is very important for identifying and segmenting objects in the image (such as vehicle borders, road markings, etc.). Texture gradient T = 100 reveals that the variance of pixel values ​​in the perception area is large (if the average variance is 100), indicating that there is more texture information in the area. High texture gradients are usually associated with complex surface structures or patterns, which helps unmanned driving systems identify road conditions, ground materials or other environmental features in images.

[0159] See also Figure 4 , the specific steps for obtaining feature weight allocation and influence evaluation are:

[0160] S211: According to the spectral intensity I, the formula is:

[0161]

[0162] Calculate the weight w of the spectral intensity I I ;

[0163] Among them, α1 is the adjustment coefficient, which is determined by data analysis and is used to adjust the sensitivity of the weight. is the average value of the spectral intensity, is the variance of the spectral intensity, indicating the divergence of the spectral intensity distribution, ∈ is a constant, usually taking a small value close to zero but not zero;

[0164] S212: According to the edge direction D, the formula is used:

[0165]

[0166] Calculate the weight w of edge direction D D ;

[0167] Among them, k2 is the adjustment coefficient, which is obtained through data analysis or model tuning process to adjust the slope of the weight response function. It is the exponential function part, which provides a smooth nonlinear mapping and ensures that the weight value is between 0 and 1. This term ensures that the denominator is non-zero and is the standard form of the logistic function used for normalized weight calculation;

[0168] S213: According to the quantized value T of the texture gradient, the formula is adopted:

[0169]

[0170] Calculate the weight w of the quantized value T of the texture gradient T ;

[0171] Among them, w T is the weight of the texture gradient, indicating the relative importance of texture features in target recognition, γ is the adjustment coefficient, obtained through data analysis, used to adjust the influence of the weight, CV(T) is the coefficient of variation of the texture gradient, calculated as the ratio of the standard deviation of the texture gradient to the mean, indicating the relative standard deviation of the texture variation, 1+CV(T) ensures that the denominator is non-zero and provides a normalized form;

[0172] The calculation process is as follows:

[0173] For the formula:

[0174]

[0175] The parameter α1 is set to 0.5 according to the previous data analysis. is 150, is 300, and ∈ is set to 0.01.

[0176]

[0177] For the formula:

[0178]

[0179] Set parameter D to 5, and k2 is determined by model tuning and set to 0.1.

[0180]

[0181] For the formula:

[0182]

[0183] The parameter γ is obtained through data analysis and is set to 2. CV(T): The coefficient of variation of the texture gradient, calculated as the ratio of the standard deviation to the mean. The mean value of the texture gradient obtained from the image data is set to 80 and the standard deviation is 20.

[0184]

[0185] The baseline weight value is set to 0.5, which represents the average importance of a feature in target detection and classification decisions. Spectral intensity weight w I =0.25, which is lower than the benchmark value, indicating that the spectral intensity in this data set has a small contribution to the classification decision due to its large variation. This means that the spectral intensity feature is not the most critical factor in distinguishing different categories. Edge direction weight w D =0.62, which is higher than the benchmark value, indicating that edge direction features are very important in images and are crucial for identifying objects with clear edges (such as vehicles and road signs), emphasizing their role in improving detection accuracy. Texture gradient weight w T =0.4, slightly lower than the benchmark value, indicating that although texture gradient has good support for target recognition in some cases, its influence is not as significant as that of edge direction.

[0186] See also Figure 5 , the steps to obtain the target recognition model are as follows:

[0187] S221: According to the weight w of the spectral intensity I I , the weight w of edge direction D D and the weight w of the quantized value T of the texture gradient T , using the formula:

[0188] B=σ(w I I+w D ·D+w T ·T)

[0189] Define classification criteria B;

[0190] Where σ is the Sigmoid activation function, which converts the linear combination into an output between 0 and 1;

[0191] S222: According to classification standard B, use the formula:

[0192]

[0193] Establish a target recognition model M;

[0194] Among them, ReLU is an activation function used to introduce nonlinearity and enhance the expressiveness of the model. Conv is a convolution operation used to extract features and learn local spatial relationships. j It is the feature map segmented according to B, used for different convolution kernels, θ j and b j are the weights and biases of the jth convolution kernel, which are adjusted by training data to optimize model performance. J is the number of convolutional layers, which determines the depth and complexity of the model. Each layer is responsible for capturing different features of the image.

[0195] The calculation process is as follows:

[0196] For the formula:

[0197] B=σ(w I I+w D ·D+w T ·T)

[0198] in, (Sigmoid function);

[0199] Set the parameter w I 、w D 、w T The values ​​are 0.25, 0.62, and 0.4 respectively, the spectral intensity I=150, the edge direction D=5, and the texture gradient T=80.

[0200] B=σ(0.25·150+0.62·5+0.4·80

[0201] B=σ(37.5+3.1+32)

[0202] B=σ(72.6)

[0203]

[0204] This result B≈1 does indicate that the classification model has a very high degree of confidence that the current detection object belongs to the target category. This is because the Sigmoid function outputs a value between 0 and 1, and a value close to 1 indicates a very high classification probability, that is, the model is very sure that the current input belongs to the positive class.

[0205] For the formula:

[0206]

[0207] Among them, ReLU(x)=max(0,x), and J is the number of convolutional layers.

[0208] Set the parameter conv to use the feature map directly derived from the image data by the image processing algorithm, θ j ,b j : Obtained through network training optimization, J has 3 layers of convolution.

[0209] The output of each convolution layer is simplified to:

[0210] conv(B j ,θ j )=0.5B j

[0211] M=ReLU(0.5·1+0.5·1+0.5·1+0)

[0212] M = ReLU(1.5) = 1.5

[0213] Assuming the benchmark value is 1.0, the model output M = 1.5, which is higher than the benchmark value of 1.0. This indicates that the model has a high accuracy of detecting that the target belongs to a specific category, which exceeds the average level. This means that the model is very confident in its classification decision, which may reduce the misjudgment rate and improve the reliability and safety of the unmanned driving system in actual operation.

[0214] See also Figure 6 , the specific steps for calculating the recognition probability of the target in the perception area are:

[0215] S311: According to the target recognition model M, the formula is used:

[0216]

[0217] The extracted feature set F t ;

[0218] Among them, M j is the jth layer in the target recognition model M, which performs convolution operations on the image for feature extraction, R t is the real-time road environment image captured at time t;

[0219] S312: Based on the feature set F t , using the formula:

[0220]

[0221] Calculate the recognition probability P of the target t ;

[0222] Among them, P t is the target recognition probability calculated at time t, Ft,q It is from F t The qth feature in v q is the weight of the qth feature, obtained through training optimization, indicating the influence of the feature in the recognition process, Q is the total number of features, which determines the range of the summation operation, and Z is the normalization factor used to ensure that P t The value of is between 0 and 1;

[0223] The calculation process is as follows:

[0224] For the formula:

[0225]

[0226] The parameters are set to have 3 convolutional layers (J=3). The output features of each convolution operation are simplified to numerical examples. The numbers in the actual operation represent the feature strength:

[0227] M1(R t )=100

[0228] M2(R t )=150

[0229] M3(R t )=200

[0230] Then the feature set is calculated as:

[0231] F t =100+150+200=450

[0232] For the formula:

[0233]

[0234] Setting parameter F t contains 3 features (Q=3), weight v q is 0.1, 0.2, 0.3, F t,q It is 100, 200, 300.

[0235] calculate Values:

[0236] e 0.1·100 ≈e 10 ≈22026.47

[0237] e 0.2·200 ≈e 40 ≈2.3538532×10 17

[0238] e 0.3·300 ≈e 90 ≈1.2204033×1039

[0239] Calculate the normalization factor Z:

[0240]

[0241] Z = 22026.47 + 2.3538532 × 10 17 +1.2204033×10 39 ≈1.2204033×10 39

[0242] (The other terms are negligible due to the very large third term).

[0243] Calculate P t :

[0244]

[0245] (mainly determined by the largest item).

[0246] Set the benchmark value 0.5 as the comparison standard, and calculate the target recognition probability P t ≈1.0, which is much higher than the benchmark value. This indicates that the model is extremely confident about the object category it recognizes, with little doubt. Compared to the benchmark value, this high probability output indicates that the model uses the input features very effectively for accurate recognition, indicating high reliability and accuracy of the system.

[0247] See also Figure 7 , the specific steps for obtaining target detection results are:

[0248] S321: According to the recognition probability P of the target t , using the formula:

[0249] E t =|P t -τ|

[0250] Calculate the recognition probability P of the target t The deviation E from the preset threshold τ t ;

[0251] Among them, τ is used to compare P t , to determine whether the target is correctly identified;

[0252] S322: According to the recognition probability P of the target t The deviation E from the preset threshold τ t , using the formula:

[0253] ΔS t =(P t -P t-1 )·λ1

[0254] Calculate the change in the spatial position of the target at time t, ΔS t ;

[0255] Among them, P t is the target recognition probability at time t, P t-1 is the target recognition probability at time t-1, indicating the recognition probability at the previous moment, λ1 is an adjustment parameter used to amplify or reduce the amount of spatial position change caused by the probability change to adapt to different detection environments and target dynamics, based on experimental data or obtained through performance feedback adjustment during system optimization;

[0256] The calculation process is as follows:

[0257] For the formula:

[0258] E t =|P t -τ|

[0259] Setting Parameters P t is 1.0, and τ is 0.9, which are set based on industry standards or system performance requirements.

[0260] E t =|1.0-0.9|=0.1

[0261] Results E t = 0.1 indicates that there is only a small deviation between the target recognition probability and the preset threshold, indicating that the detection accuracy is very high.

[0262] For the formula:

[0263] ΔS t =(P t -P t-1 )·λ1

[0264] Setting parameter P t is 1.0, P t-1 is 0.95, and λ1 is 10, which are obtained through optimization simulation and actual testing.

[0265] ΔS t =(1.0-0.95)·10=0.5

[0266] Results ΔS t = 0.5 indicates that the difference in recognition probability at consecutive time points leads to a small change in the target's spatial position. This amount of change can help the system evaluate the dynamic position change of the target, which is crucial for unmanned vehicle navigation in dynamic environments.

[0267] See also Figure 8 , the specific steps for obtaining dynamic target prediction results are:

[0268] S411: Based on the change in the spatial position of the target at time t, ΔS t , using the formula:

[0269] L t+1 =L t +v t ·Δt+ΔS t

[0270] Calculate the future position L of the target t+1 ;

[0271] Among them, L t+1 is the target position predicted at time t+1, which is used to determine the future spatial position of the target. t is the target position at the current time t, indicating the known position of the target at the current time point, v t is the target speed at the current time t, indicating the target's moving speed, Δt is the time interval, indicating the time difference from time t to t+1, which is used to predict the future position;

[0272] S412: Based on the future position L of the target t+1 , using the formula:

[0273] O new =O current +α2·(L t+1 -O current )

[0274] Calculate the adjusted route O new ;

[0275] Among them, O current is the current vehicle route, which indicates the path of the vehicle at the current time point; α2 is the route adjustment coefficient, which is used to control the adjustment amplitude of the vehicle route. It is obtained through experiments or based on the feedback of the vehicle dynamic control system to ensure that the adjustment amplitude is within a safe range and adapts to the current traffic conditions;

[0276] The calculation process is as follows:

[0277] For the formula:

[0278] L t+1 =L t +v t ·Δt+ΔS t

[0279] Setting parameter L t For 100 meters, v t is 10 m / s, Δt is 1 second, ΔS t is 0.5 m.

[0280] L t+1=100+10·1+0.5=110.5

[0281] Results t+1 =110.5 meters represents the expected position of the target at the next moment.

[0282] For the formula:

[0283] O new =O current +α2·(L t+1 -O current )

[0284] Set Parameters current is 105 m, and α2 is 0.3, which are adjusted based on simulation and experimental data.

[0285] O new =105+0.3·(110.5-105)=105+0.3·5.5=106.65

[0286] Results new =106.65 meters represents the new route position after considering the target movement.

[0287] Route change ΔO = 1.65 meters, indicating that the route adjustment is small, which means that the adjustment is gentle and will not cause drastic changes in vehicle behavior, helping to maintain driving stability and safety. Such results show that the system makes reasonable path adjustments based on future target locations to adapt to changes in the surrounding environment, ensuring that the vehicle can effectively respond to dynamic obstacles ahead while maintaining route safety.

[0288] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A target detection method for unmanned driving, characterized in that: The following steps are involved: Collect image data of the road environment, set the perception area size of each pixel according to the complexity of the road environment and the expected size of the target to be identified, extract image features within the perception area, including spectral intensity, edge direction and texture gradient, and generate a set of perception features; Based on the perceptual feature set, the spectral intensity, edge direction and texture gradient features are weighted and evaluated for influence, and the features are used to establish a classification standard, and the feature parameters are introduced to initialize the convolutional neural network to obtain a target recognition model; Based on the target recognition model, the road environment image data captured in real time is analyzed, the recognition probability of the target in the perception area is calculated, and the recognition probability is compared with the preset threshold to determine the detection accuracy of the target, and the change of the target spatial position is calculated to obtain the target detection result; Based on the target detection results, combined with the vehicle's current position and speed information, the target's potential movement trajectory and future behavior changes are analyzed, and the integrated information is used to assist in adjusting the vehicle's driving route to obtain dynamic target prediction results.

2. The target detection method for unmanned driving according to claim 1, characterized in that: The steps of obtaining the size of the sensing area of ​​each pixel are specifically as follows: According to the collected road environment image data, the formula is used: Calculate the environmental complexity C of the road image; Where n is the total number of object categories in the image, x i is the number of objects of the i-th class, w i is the weight factor of the i-th type of object, which is set in advance according to the difficulty of detecting the object type, d i is the average distance from the i-th class of objects to the center of the image; According to the collected road environment image data, the formula is used: Calculate the expected size S of the target to be identified; Where n′ is the number of similar objects in the image, b f is the area of ​​the fth target, h f is the size coefficient of the fth target, indicating the display ratio of the target in the image; According to the environmental complexity C of the road image and the expected size S of the target to be identified, the formula is adopted: Calculate the perception area size G of each pixel; Among them, k1 is the adjustment coefficient, which is obtained according to the data in the scene, p1 is the exponential attenuation factor of the environmental complexity, which is determined by experiments, ln is the natural logarithmic function, and v1 is the environmental sensitivity parameter, which is adjusted by driving data.

3. The target detection method for unmanned driving according to claim 2, characterized in that: The steps of acquiring the perceptual feature set are specifically as follows: According to the perception area size G of each pixel, the formula is adopted: Calculate the spectral intensity I; Where P(x,y) is the value of pixel P at coordinate (x,y), x and y are the pixel coordinates in the perception area, and the range of x and y is from 1 to G; According to the perception area size G of each pixel, the formula is adopted: Calculate the edge direction D; According to the perception area size G of each pixel, the formula is adopted: Calculate the quantized value T of the texture gradient and integrate all the information to obtain the perceptual feature set; Where A is the preset average brightness level.

4. The target detection method for unmanned driving according to claim 3, characterized in that: The steps for obtaining feature weight allocation and influence evaluation are specifically as follows: According to the spectral intensity I, the formula is adopted: Calculate the weight w of the spectral intensity I I ; Among them, α1 is the adjustment coefficient, which is determined by data analysis. is the average value of the spectral intensity, is the variance of the spectral intensity, ∈ is a constant; According to the edge direction D, the formula is adopted: Calculate the weight w of edge direction D D ; Among them, k2 is the adjustment coefficient, which is obtained through data analysis or model tuning process. It is the exponential function part, which makes the weight value between 0 and 1; According to the quantized value T of the texture gradient, the formula is adopted: Calculate the weight w of the quantized value T of the texture gradient T ; Among them, w T is the weight of the texture gradient, γ is the adjustment coefficient, which is obtained through data analysis, and CV(T) is the coefficient of variation of the texture gradient.

5. The target detection method for unmanned driving according to claim 4, characterized in that: The steps of acquiring the target recognition model are specifically as follows: According to the weight w of the spectral intensity I I , the weight w of edge direction D D and the weight w of the quantized value T of the texture gradient T , using the formula: B=σ(w I ·I+w D ·D+w T ·T) Define classification criteria B; Where σ is the activation function that transforms the linear combination into an output between 0 and 1; According to the classification standard B, the formula is used: Establish a target recognition model M; Among them, ReLU is the activation function used to introduce nonlinearity, conv is the convolution operation used to extract features and learn local spatial relationships, and B j It is the feature map segmented according to B, used for convolution kernel, θ j and b j are the weights and biases of the jth convolution kernel, adjusted by the training data, and J is the number of convolutional layers.

6. The target detection method for unmanned driving according to claim 5, characterized in that: The steps of obtaining the recognition probability of the target in the sensing area are specifically as follows: According to the target recognition model M, the formula is adopted: The extracted feature set F t ; Among them, M j is the jth layer in the target recognition model M, which performs convolution operations on the image for feature extraction, R t is the real-time road environment image captured at time t; According to the feature set F t , using the formula: Calculate the recognition probability P of the target t ; Among them, P t is the target recognition probability calculated at time t, F t,q It is from F t The qth feature in v q is the weight of the qth feature, which is obtained through training optimization, Q is the total number of features, and Z is the normalization factor.

7. The target detection method for unmanned driving according to claim 6, characterized in that: The steps for obtaining the target detection result are specifically as follows: According to the recognition probability P of the target t , using the formula: E t =|P t -t| Calculate the recognition probability P of the target t The deviation E from the preset threshold τ t ; According to the recognition probability P of the target t The deviation E from the preset threshold τ t , using the formula: ΔS t =(P t -P t-1 )·λ1 Calculate the change in the spatial position of the target at time t, ΔS t ; Among them, P t is the target recognition probability at time t, P t-1 is the target recognition probability at time t-1, indicating the recognition probability at the previous moment, and λ1 is an adjustment parameter used to amplify or reduce the change in spatial position caused by the probability change, which is obtained based on experimental data or performance feedback adjustment during the optimization process.

8. The target detection method for unmanned driving according to claim 7, characterized in that: The steps for obtaining the dynamic target prediction result are specifically as follows: According to the spatial position change ΔS of the target at time t t , using the formula: L t+1 =L t +v t ·Δt+ΔS t Calculate the future position L of the target t+1 ; Among them, L t is the target position at the current time t, v t is the target speed at the current time t, Δt is the time interval, which represents the time difference from time t to t+1; According to the future position L of the target t+1 , using the formula: O new =O current +α2·(L t+1 -O current ) Calculate the adjusted route O new ; Among them, O current is the current vehicle route, which represents the path of the vehicle at the current time point, and α2 is the route adjustment coefficient, which is obtained through experiments or based on feedback from vehicle dynamic control.

Images