Intelligent parking auxiliary system based on image acquisition

Abstract

The invention discloses an intelligent parking auxiliary system based on image acquisition, and relates to the technical field of intelligent parking, the intelligent parking auxiliary system uses a multispectral visual system to acquire environment images of different wave bands around a vehicle in real time, and combines an HDR imaging module to carry out detail enhancement on highlight and low light areas, so as to improve the accuracy of parking. A cross-spectrum fusion algorithm based on a convolutional neural network carries out feature alignment and deep synthesis on a multi-band image, the visibility of target features is enhanced, a YOLO target detection algorithm and a boundary detection technology are introduced at the same time, a dynamic obstacle and the type and position thereof are identified, a parking space contour is extracted to generate a parking space model, and the parking space model is optimized. Distributing priority labels according to dynamic behaviors of the obstacles; in the aspect of path planning, a Bezier curve algorithm is adopted to generate a parking path, path control points are dynamically adjusted through obstacle priorities, and a vehicle control module dynamically adjusts a control strategy through discretization of the parking path and real-time correction of a kinetic model based on a closed-loop feedback control mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent parking, and in particular to an intelligent parking assistance system based on image acquisition. Background Art

[0002] With the popularization of intelligent driving technology, intelligent parking assistance systems have gradually become an important part of the automotive field. Such systems use on-board cameras to capture images of the vehicle's surroundings in real time, rely on computer vision algorithms to identify parking lines, obstacles and driving paths, and provide parking guidance for drivers.

[0003] However, when using this assistance system in low-light environments such as at night or in underground parking lots, the images captured by the camera are easily blurred and lack sufficient details due to the influence of light. In open areas or under direct sunlight, strong light spots, overexposed areas, or even complete loss of key features may appear on the camera screen, which directly affects the judgment ability of the visual algorithm and may lead to misidentification of obstacles or inability to detect the exact location of the parking space, causing deviations in parking path planning or parking failures.

[0004] To address the above issues, some traditional parking assistance solutions integrate high dynamic range (HDR) technology into the camera to balance the contrast between light and dark. Some algorithms dynamically adjust image brightness and contrast to enhance target recognition in low-light scenes, which improves the adaptability of the system to a certain extent. However, in environments with drastic lighting changes or complex light sources, the recognition effect still cannot meet the needs. Therefore, an intelligent parking assistance solution based on image acquisition is urgently needed to solve such problems. Summary of the invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides an intelligent parking assistance system based on image acquisition to solve the problem that traditional image-based parking methods perform poorly in low-light and high dynamic range scenes, such as parking lots at night or parking spaces with strong sunlight directly shining on them, which easily leads to obstacle recognition failure or misdetection.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] An embodiment of the present invention provides an intelligent parking assistance system based on image acquisition, which includes:

[0009] Multispectral vision system, used to collect environmental images of different bands around the vehicle in real time, including visible light, near infrared and far infrared band data;

[0010] High dynamic range (HDR) imaging module, which is used to enhance the details of bright and dark areas of the environmental images collected by the multispectral vision system, optimize the brightness performance in low-light scenes and the contrast balance under strong light conditions;

[0011] The spectral fusion processing module is used to input multi-band environmental image data into the fusion algorithm, perform cross-spectral feature alignment and deep integration through the convolutional neural network, and generate a fused image;

[0012] The path planning module is used to generate a parking path using the Bezier curve algorithm based on the obstacle detection results of the fused image, the parking space contour information and the current position of the vehicle, and dynamically adjust the path according to the priority of the obstacles;

[0013] The vehicle control module is used to receive the parking path instructions generated by the path planning module, output vehicle speed control and steering control signals, and adjust the control strategy in real time through closed-loop feedback.

[0014] As a preferred solution of the intelligent parking assistance system based on image acquisition described in the present invention, the assistance method of the intelligent parking assistance system based on image acquisition is:

[0015] Step S1, the multispectral vision system collects environmental images of different bands around the vehicle in real time, and combines the high dynamic range (HDR) imaging module to enhance the details of the high light and low light areas;

[0016] Step S2, inputting the multi-band environmental image data obtained in step S1 into the spectral fusion processing module, using a cross-spectral fusion algorithm based on a convolutional neural network to perform feature alignment and depth synthesis on the image, and outputting a fused image;

[0017] Step S3, based on the fused image, the YOLO target detection algorithm is used to identify dynamic obstacles in the environment, including their type and location information; at the same time, the parking space contour is extracted through the boundary detection algorithm, and a parking space model is generated. In combination with the dynamic behavior characteristics of the obstacles, priority labels are assigned to different targets;

[0018] Step S4, generating a parking path using a Bezier curve path planning algorithm according to the obstacle recognition result obtained in step S3, the parking space contour information and the current posture data of the vehicle;

[0019] Step S5, inputting the planned path into the vehicle control module to generate vehicle speed control and steering control instructions.

[0020] As a preferred solution of the intelligent parking assistance system based on image acquisition described in the present invention, the step of enhancing the details of the highlight and low light areas by combining the high dynamic range HDR imaging module is as follows:

[0021] Let the input image be I, use the brightness decomposition function to process I, and generate the highlight area I h and low light area I l :

[0022] I h =ReLU(IT h ),I l =ReLU(T l -I),

[0023] Where I represents the input multispectral image, T h Represents the highlight threshold, which is used to distinguish the highlight area, T l Indicates the low light threshold, used to distinguish low light areas, I h Represents the highlight area image, I l represents the low-light area image, ReLU(x) represents the linear correction activation function,

[0024] The local details of the highlight and low light areas are enhanced respectively, and the enhancement formula is:

[0025] I′ h =I h ·(1+α h ·log(1+I h )),I′ l =I l ·(1+α l ·log(1+I l )),

[0026] Among them, I′ h and I′ l Represent the enhanced highlight and low-light images, α h and α l They represent the highlight and low light enhancement coefficients respectively, and log(x) is a logarithmic function;

[0027] Merge the enhanced highlights with the low-light areas to get the enhanced image:

[0028] I′=I′ h +I l ′,

[0029] Where I′ represents the output image after detail enhancement.

[0030] As a preferred solution of the intelligent parking assistance system based on image acquisition described in the present invention, the steps of using a cross-spectral fusion algorithm based on a convolutional neural network to perform feature alignment and depth synthesis on the image and output a fused image are as follows:

[0031] Input multispectral image sequence Use the initial feature extraction network to extract features for each band:

[0032] F i =Conv(I i ;w i ,b i ),

[0033] Among them, I i represents the input image of the i-th band, n represents the number of bands, F i represents the feature map of the i-th band, Conv(I;w,b) represents the convolution operation, the weight is w, the bias is b,

[0034] Cross-spectral registration is used to correct the spatial offset of each band feature. The correction formula is:

[0035] F i ′=Warp(F i ,H i ),

[0036] Among them, F i ′ represents the i-th band feature after alignment, H i represents the perspective transformation matrix of the i-th band, and Warp(F,H) represents the feature transformation function based on H;

[0037] The aligned features are fused by multi-layer convolution, and the fused features are calculated using the attention mechanism. The calculation formula is:

[0038]

[0039] Among them, F fusion represents the fused global features, Attention(x) represents the attention mechanism function, and w i Represents the weight coefficient of the i-th band feature;

[0040] The fused features are restored to the output image through the decoder:

[0041] I out =Decoder(F fusion ),

[0042] Among them, I out represents the output fused image, and Decoder(x) represents the feature decoder function.

[0043] As a preferred solution of the intelligent parking assistance system based on image acquisition described in the present invention, the step of using the YOLO target detection algorithm to identify dynamic obstacles in the environment is:

[0044] Based on the fusion image I out, extract the feature vector of the candidate region:

[0045] F YOLO =Backbone(I out ),

[0046] Among them, I out represents the input fused image, F YOLO represents the extracted feature vector, Backbone(x) represents the backbone network of YOLO;

[0047] Use YOLO's anchor box mechanism to predict candidate boxes Location and Category:

[0048] P k ,C k ,S k =Head(F YOLO ,A k ),

[0049] in, represents m candidate boxes, P k represents the center position of the kth candidate box, C k represents the category probability of the kth candidate box, S k represents the size of the kth candidate box, A k represents the kth anchor box parameter, Head(x,A) represents the detection head of YOLO, which is used to predict the position, category and size;

[0050] Use non-maximum suppression NMS to filter the final detection results The screening formula is:

[0051]

[0052] in, are the n detection boxes outputted in the end, and NMS(·) is the non-maximum suppression algorithm.

[0053] As a preferred solution of the intelligent parking assistance system based on image acquisition described in the present invention, the steps of extracting the parking space contour by using a boundary detection algorithm, generating a parking space model, and assigning priority labels to different targets in combination with the dynamic behavior characteristics of obstacles are as follows:

[0054] For the fused image I out Perform edge detection and generate a binary boundary image:

[0055] I edge =Canny(I out ),

[0056] Among them, I edgerepresents the binary image after boundary detection, Canny(x) represents the Canny edge detection algorithm,

[0057] Use contour extraction algorithm to determine the parking space contour:

[0058] C slot =Contour(I edge ),

[0059] Among them, C slot represents the detected parking space contour, Contour(x) represents the contour extraction algorithm;

[0060] Generate a parking space model based on the parking space outline:

[0061] M slot =Fit(C slot ),

[0062] Among them, M slot represents the generated parking space model, Fit(x) represents the fitting algorithm used to convert the contour into a parking space geometry model,

[0063] Analyze the behavior characteristics of the detected dynamic obstacles and calculate the target priority label L based on the priority rule k , the calculation formula is:

[0064] L k =Priority(D k ,V k ,R k ),

[0065] Among them, L k represents the priority label of the kth target, D k Indicates the distance between the target and the parking space, V k Indicates the target's moving speed, R k It represents the danger factor of the target, and Priority(x,y,z) represents the priority calculation function.

[0066] As a preferred solution of the intelligent parking assistance system based on image acquisition described in the present invention, in step S4, the path planning module dynamically adjusts the path according to the obstacle priority label.

[0067] As a preferred solution of the intelligent parking assistance system based on image acquisition described in the present invention, the step of generating a parking path using a Bezier curve path planning algorithm is as follows:

[0068] Determine the parking starting point P start and parking end point P end :

[0069] Pstart =(x s ,y s ),P end =(x e ,y e ),

[0070] Among them, P start represents the initial position of the current vehicle, P end Indicates the center point of the parking space, (x s ,y s ) and (x e ,y e ) are the coordinates of the starting point and the end point,

[0071] Determine the set of control points Constrain curve shape:

[0072] P i =(x i ,y i ),

[0073] Among them, P i represents the i-th control point, (x i ,y i ) are the coordinates of the control points, and m represents the number of control points;

[0074] Generate a Bezier path:

[0075]

[0076] Among them, B(t) represents the curve point corresponding to parameter t, t is the path parameter, and its value range is [0,1].

[0077] represents the number of combinations, defined as

[0078] According to the obstacle priority label L k Dynamically adjust the path control points, the adjustment formula is:

[0079] P i ′=P i +ΔP i ,

[0080] Among them, P i ′ represents the adjusted control point, ΔP i Represents the adjustment amount, which is calculated based on the obstacle priority. The calculation formula is:

[0081]

[0082] Among them, L k is the priority label of the kth target, fk (P i ) represents the kth target control point P i The influence function of n is the number of obstacles detected.

[0083] Generate the final parking path B′(t):

[0084]

[0085] Among them, B′(t) is the dynamically adjusted path.

[0086] As a preferred solution of the intelligent parking assistance system based on image acquisition described in the present invention, in which: in step S5, the parking task is performed through closed-loop control: during the execution process, the vehicle control module obtains feedback data in real time and dynamically corrects the control strategy.

[0087] As a preferred solution of the intelligent parking assistance system based on image acquisition described in the present invention, the step of inputting the planned path into the vehicle control module to generate the vehicle speed control and steering control instructions is as follows:

[0088] The dynamically adjusted parking path B′(t) is discretized to obtain a set of discrete path points. The discretization formula is:

[0089]

[0090] Among them, {P t} represents a set of discrete path points, T represents the total duration of the parking path,

[0091] is the kth discrete moment, N is the discrete number of path points,

[0092] According to the discrete path points, calculate the speed v of the vehicle at each moment k and steering angle θ k , the calculation formula is:

[0093]

[0094] Among them, v k represents the speed of the vehicle at the kth moment, θ k represents the steering angle of the vehicle at the kth moment, (x k ,y k ) and (x k+1 ,y k+1 ) is the coordinate of the path point, Δt is the time interval between adjacent path points;

[0095] Generate control instructions based on the vehicle dynamics model:

[0096] u k =(v k ,ω k ),

[0097]

[0098] Among them, u k represents the kth control instruction, ω k represents the angular velocity of the vehicle at the kth moment,

[0099] Feedback correction is performed on the execution process through closed-loop control, and control instructions are adjusted in real time:

[0100] u′ k =u k +K eek ,

[0101] Among them, u′ k Indicates the modified control instruction, K e is the feedback gain matrix, e k represents the path deviation at the kth moment and is defined as:

[0102]

[0103] Among them, P k is the kth path point of the planned path, is the actual position of the vehicle.

[0104] The beneficial effects of the present invention are as follows: the present invention utilizes a multispectral vision system to collect environmental images of different bands around the vehicle in real time, and combines the HDR imaging module to enhance the details of high-light and low-light areas, significantly improving the image quality under weak-light and strong-light conditions; a cross-spectral fusion algorithm based on a convolutional neural network performs feature alignment and depth synthesis on multi-band images to generate a clear and consistent fused image, enhancing the visibility of target features; at the same time, the YOLO target detection algorithm and boundary detection technology are introduced to identify dynamic obstacles and their types and positions, and the parking space contour is extracted to generate a parking space model, and priority labels are assigned according to the dynamic behavior of the obstacles; in terms of path planning, a Bezier curve algorithm is used to generate a parking path, and the path control points are dynamically adjusted according to the obstacle priority; the vehicle control module is based on a closed-loop feedback control mechanism, and dynamically adjusts the control strategy by discretizing the parking path and correcting the dynamic model in real time.

[0105] The present invention significantly improves the reliability and adaptability of the parking assistance system in low-light, high-light or dynamic obstacle scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0107] Figure 1 Schematic diagram of the framework of the intelligent parking assistance system of the present invention.

[0108] Figure 2 Schematic diagram of the auxiliary process of the intelligent parking assistance system of the present invention. DETAILED DESCRIPTION

[0109] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0110] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0111] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0112] Example 1, reference Figure 1 and Figure 2 , this embodiment provides an intelligent parking assistance system based on image acquisition, including:

[0113] Multispectral vision system, used to collect environmental images of different bands around the vehicle in real time, including visible light, near infrared and far infrared band data;

[0114] High dynamic range (HDR) imaging module, which is used to enhance the details of bright and dark areas of the environmental images collected by the multispectral vision system, optimize the brightness performance in low-light scenes and the contrast balance under strong light conditions;

[0115] The spectral fusion processing module is used to input multi-band environmental image data into the fusion algorithm, perform cross-spectral feature alignment and deep integration through the convolutional neural network, and generate a fused image;

[0116] The path planning module is used to generate a parking path using the Bezier curve algorithm based on the obstacle detection results of the fused image, the parking space contour information and the current position of the vehicle, and dynamically adjust the path according to the priority of the obstacles;

[0117] The vehicle control module is used to receive the parking path instructions generated by the path planning module, output vehicle speed control and steering control signals, and adjust the control strategy in real time through closed-loop feedback.

[0118] This embodiment also provides an assisting method of the intelligent parking assist system based on image acquisition, including:

[0119] Step S1, the multispectral vision system collects environmental images of different bands around the vehicle in real time, and combines the high dynamic range (HDR) imaging module to enhance the details of the high light and low light areas;

[0120] Combined with the high dynamic range (HDR) imaging module, the steps to enhance the details of the highlight and low light areas are as follows:

[0121] Let the input image be I, use the brightness decomposition function to process I, and generate the highlight area I h and low light area I l :

[0122] I h =ReLU(IT h ),I l =ReLU(T l -I),

[0123] Where I represents the input multispectral image, T h Represents the highlight threshold, which is used to distinguish the highlight area, T l Indicates the low light threshold, used to distinguish low light areas, I h Represents the highlight area image, I l represents the low-light area image, ReLU(x) represents the linear correction activation function,

[0124] The local details of the highlight and low light areas are enhanced respectively, and the enhancement formula is:

[0125] I′ h =I h ·(1+α h ·log(1+I h )),I′ l ′=I l ·(1+α l ·log(1+I l )),

[0126] Among them, I′ h and I l′ represents the enhanced high-light and low-light images, respectively, and α h and α l They represent the highlight and low light enhancement coefficients respectively, and log(x) is a logarithmic function;

[0127] Merge the enhanced highlights with the low-light areas to get the enhanced image:

[0128] I′=I′ h +I l ′,

[0129] Where I′ represents the output image after detail enhancement,

[0130] Specifically, the input image is divided into highlight and low-light areas through brightness decomposition technology, and nonlinear detail enhancement is performed on both of them respectively, and the contrast is improved through a logarithmic function to generate an enhanced image.

[0131] Step S2, inputting the multi-band environmental image data obtained in step S1 into the spectral fusion processing module, using a cross-spectral fusion algorithm based on a convolutional neural network to perform feature alignment and depth synthesis on the image, and outputting a fused image;

[0132] The cross-spectral fusion algorithm based on convolutional neural network is used to align the features and perform depth synthesis on the images. The steps to output the fused image are as follows:

[0133] Input multispectral image sequence Use the initial feature extraction network to extract features for each band:

[0134] F i =Conv(I i ;w i ,b i ),

[0135] Among them, I i represents the input image of the i-th band, n represents the number of bands, F i represents the feature map of the i-th band, Conv(I;w,b) represents the convolution operation, the weight is w, the bias is b,

[0136] Cross-spectral registration is used to correct the spatial offset of each band feature. The correction formula is:

[0137] F i ′=Warp(F i ,H i ),

[0138] Among them, F i ′ represents the first i Band characteristics, H irepresents the perspective transformation matrix of the i-th band, and Warp(F,H) represents the feature transformation function based on H;

[0139] The aligned features are fused by multi-layer convolution, and the fused features are calculated using the attention mechanism. The calculation formula is:

[0140]

[0141] Among them, F fusion represents the fused global features, Attention(x) represents the attention mechanism function, and w i Represents the weight coefficient of the i-th band feature;

[0142] The fused features are restored to the output image through the decoder:

[0143] I out =Decoder(F fusion ),

[0144] Among them, I out represents the output fused image, Decoder(x) represents the feature decoder function;

[0145] Specifically, multispectral images are fused through feature extraction, cross-spectral registration and feature deep integration. The convolution layer extracts band features and aligns spatial deviations through the transformation matrix. The attention mechanism improves the expressiveness of the fused image. Finally, the decoder is used to restore the high-resolution, full-band consistent output image.

[0146] Step S3, based on the fused image, the YOLO target detection algorithm is used to identify dynamic obstacles in the environment, including their type and location information; at the same time, the parking space contour is extracted through the boundary detection algorithm, and a parking space model is generated. In combination with the dynamic behavior characteristics of the obstacles, priority labels are assigned to different targets;

[0147] The steps to use the YOLO target detection algorithm to identify dynamic obstacles in the environment are:

[0148] Based on the fusion image I out , extract the feature vector of the candidate region:

[0149] F YOLO =Backbone(I out ),

[0150] Among them, I out represents the input fused image, F YOLO represents the extracted feature vector, Backbone(x) represents the backbone network of YOLO;

[0151] Use YOLO's anchor box mechanism to predict candidate boxes Location and Category:

[0152] P k ,C k ,S k =Head(F YOLO ,A k ),

[0153] in, represents m candidate boxes, P k represents the center position of the kth candidate box, C k represents the category probability of the kth candidate box, S k represents the size of the kth candidate box, A k represents the kth anchor box parameter, Head(x,A) represents the detection head of YOLO, which is used to predict the position, category and size;

[0154] Use non-maximum suppression NMS to filter the final detection results The screening formula is:

[0155]

[0156] in, are the n detection boxes outputted in the end, and NMS(·) is the non-maximum suppression algorithm;

[0157] Specifically, in this step, detection is performed based on the YOLO model, multi-level features are extracted through the backbone network, the anchor box mechanism is used to predict the center position, category and size of the candidate box, and finally the non-maximum suppression method is used to screen the effective detection results to improve the detection efficiency; dynamic obstacles in multispectral images can be quickly and accurately identified, and the specific location and category information can be output.

[0158] The steps of extracting the parking space contour through the boundary detection algorithm, generating the parking space model, and assigning priority labels to different targets in combination with the dynamic behavior characteristics of obstacles are as follows:

[0159] For the fused image I out Perform edge detection and generate a binary boundary image:

[0160] I edge =Canny(I out ),

[0161] Among them, I edge represents the binary image after boundary detection, Canny(x) represents the Canny edge detection algorithm,

[0162] Use contour extraction algorithm to determine the parking space contour:

[0163] C slot=Contour(I edge ),

[0164] Among them, C slot represents the detected parking space contour, Contour(x) represents the contour extraction algorithm;

[0165] Generate a parking space model based on the parking space outline:

[0166] M slot =Fit(C slot ),

[0167] Among them, M slot represents the generated parking space model, Fit(x) represents the fitting algorithm used to convert the contour into a parking space geometry model,

[0168] Analyze the behavior characteristics of the detected dynamic obstacles and calculate the target priority label L based on the priority rule k , the calculation formula is:

[0169] L k =Priority(D k ,V k ,R k ),

[0170] Among them, L k represents the priority label of the kth target, D k Indicates the distance between the target and the parking space, V k Indicates the target's moving speed, R k represents the risk factor of the target, and Priority(x,y,z) represents the priority calculation function;

[0171] Specifically, this step identifies and geometrically models the parking area through edge detection and contour extraction techniques, and combines the motion behavior and dangerous characteristics of dynamic obstacles to assign priority labels to targets in the environment.

[0172] Step S4, generating a parking path using a Bezier curve path planning algorithm according to the obstacle recognition result obtained in step S3, the parking space contour information and the current posture data of the vehicle;

[0173] In step S4, the path planning module dynamically adjusts the path according to the obstacle priority labels;

[0174] Using the Bezier curve path planning algorithm, the steps to generate the parking path are:

[0175] Determine the parking starting point P start and parking end point P end :

[0176] Pstart =(x s ,y s ),P end =(x e ,y e ),

[0177] Among them, P start represents the initial position of the current vehicle, P end Indicates the center point of the parking space, (x s ,y s ) and (x e ,y e ) are the coordinates of the starting point and the end point,

[0178] Determine the set of control points Constrain curve shape:

[0179] P i =(x i ,y i ),

[0180] Among them, P i represents the i-th control point, (x i ,y i ) are the coordinates of the control points, and m represents the number of control points;

[0181] Generate a Bezier path:

[0182]

[0183] Among them, B(t) represents the curve point corresponding to parameter t, t is the path parameter, and its value range is [0,1].

[0184] represents the number of combinations, defined as

[0185] According to the obstacle priority label L k Dynamically adjust the path control points, the adjustment formula is:

[0186] P i ′=P i +ΔP i ,

[0187] Among them, P i ′ represents the adjusted control point, ΔP i Represents the adjustment amount, which is calculated based on the obstacle priority. The calculation formula is:

[0188]

[0189] Among them, L k is the priority label of the kth target, fk (P i ) represents the kth target control point P i The influence function of n is the number of obstacles detected.

[0190] Generate the final parking path B′(t):

[0191]

[0192] Among them, B′(t) is the dynamically adjusted path,

[0193] Specifically, the parking path is generated by the Bezier curve. The curve shape is constrained by the starting point, end point and several control points. The control points are dynamically adjusted based on the obstacle priority labels to ensure the path obstacle avoidance effect and parking accuracy, and finally generate a smooth and controllable curve path.

[0194] Step S5, inputting the planned path into the vehicle control module to generate vehicle speed control and steering control instructions;

[0195] In step S5, the parking task is executed through closed-loop control: during the execution, the vehicle control module obtains feedback data in real time and dynamically modifies the control strategy;

[0196] The steps of inputting the planned path into the vehicle control module and generating the vehicle's speed control and steering control instructions are as follows:

[0197] The dynamically adjusted parking path B′(t) is discretized to obtain a set of discrete path points. The discretization formula is:

[0198]

[0199] Among them, {P t} represents a set of discrete path points, T represents the total duration of the parking path,

[0200] is the kth discrete moment, N is the discrete number of path points,

[0201] According to the discrete path points, calculate the speed v of the vehicle at each moment k and steering angle θ k , the calculation formula is:

[0202]

[0203] Among them, v k represents the speed of the vehicle at the kth moment, θ k represents the steering angle of the vehicle at the kth moment, (x k ,y k ) and (x k+1 ,yk+1 ) is the coordinate of the path point, Δt is the time interval between adjacent path points;

[0204] Generate control instructions based on the vehicle dynamics model:

[0205] u k =(v k ,ω k ),

[0206]

[0207] Among them, u k represents the kth control instruction, ω k represents the angular velocity of the vehicle at the kth moment,

[0208] Feedback correction is performed on the execution process through closed-loop control, and control instructions are adjusted in real time:

[0209] u′ k =u k +K e e k ,

[0210] Among them, u′ k Indicates the modified control instruction, K e is the feedback gain matrix, e k represents the path deviation at the kth moment and is defined as:

[0211]

[0212] Among them, P k is the kth path point of the planned path, is the actual position of the vehicle;

[0213] Specifically, the vehicle's speed and steering angle instructions are calculated by discretizing the parking path, and the control parameters are generated in combination with the vehicle dynamics model. Closed-loop control feedback is used to correct deviations in real time to ensure that the vehicle is parked accurately along the planned path.

[0214] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An intelligent parking assistance system based on image acquisition, characterized in that: include, Multispectral vision system, used to collect environmental images of different bands around the vehicle in real time, including visible light, near infrared and far infrared band data; High dynamic range (HDR) imaging module, which is used to enhance the details of bright and dark areas of the environmental images collected by the multispectral vision system, optimize the brightness performance in low-light scenes and the contrast balance under strong light conditions; The spectral fusion processing module is used to input multi-band environmental image data into the fusion algorithm, perform cross-spectral feature alignment and deep integration through the convolutional neural network, and generate a fused image; The path planning module is used to generate a parking path using the Bezier curve algorithm based on the obstacle detection results of the fused image, the parking space contour information and the current position of the vehicle, and dynamically adjust the path according to the priority of the obstacles; The vehicle control module is used to receive the parking path instructions generated by the path planning module, output vehicle speed control and steering control signals, and adjust the control strategy in real time through closed-loop feedback.

2. The intelligent parking assistance system based on image acquisition according to claim 1, characterized in that: The auxiliary modes of the intelligent parking assistance system based on image acquisition are: Step S1, the multispectral vision system collects environmental images of different bands around the vehicle in real time, and combines the high dynamic range (HDR) imaging module to enhance the details of the high light and low light areas; Step S2, inputting the multi-band environmental image data obtained in step S1 into the spectral fusion processing module, using a cross-spectral fusion algorithm based on a convolutional neural network to perform feature alignment and depth synthesis on the image, and outputting a fused image; Step S3, based on the fused image, the YOLO target detection algorithm is used to identify dynamic obstacles in the environment, including their type and location information; at the same time, the parking space contour is extracted through the boundary detection algorithm, and a parking space model is generated. In combination with the dynamic behavior characteristics of the obstacles, priority labels are assigned to different targets; Step S4, generating a parking path using a Bezier curve path planning algorithm according to the obstacle recognition result obtained in step S3, the parking space contour information and the current posture data of the vehicle; Step S5, inputting the planned path into the vehicle control module to generate vehicle speed control and steering control instructions.

3. The intelligent parking assistance system based on image acquisition as claimed in claim 2, characterized in that: The step of enhancing the details of the highlight and low light areas by combining the high dynamic range HDR imaging module is as follows: Let the input image be I, use the brightness decomposition function to process I, and generate the highlight area I h and low light areas I l : AND h =ReLU(IT h ),AND l =ReLU(T l -AND), Where I represents the input multispectral image, T h Represents the highlight threshold, which is used to distinguish the highlight area, T l Indicates the low light threshold, used to distinguish low light areas, I h Represents the highlight area image, I l represents the low-light area image, ReLU(x) represents the linear correction activation function, The local details of the highlight and low light areas are enhanced respectively, and the enhancement formula is: I′ h =I h ·(1+α h ·log(1+I h )),I′ l =I l ·(1+α l ·log(1+I l )), Among them, I′ h and I′ l Represent the enhanced highlight and low-light images, α h and α l They represent the highlight and low light enhancement coefficients respectively, and log(x) is a logarithmic function; Merge the enhanced highlights with the low-light areas to get the enhanced image: I′=I′ h +I′ l , Where I′ represents the output image after detail enhancement.

4. The intelligent parking assistance system based on image acquisition as claimed in claim 3, characterized in that: The cross-spectral fusion algorithm based on convolutional neural network is used to perform feature alignment and depth synthesis on the image, and the steps of outputting the fused image are as follows: Input multispectral image sequence Use the initial feature extraction network to extract features for each band: F i =Conv(I i ;W i ,b i ), Among them, I i represents the input image of the i-th band, n represents the number of bands, F i represents the feature map of the i-th band, Conv(I;w,b) represents the convolution operation, the weight is w, the bias is b, Cross-spectral registration is used to correct the spatial offset of each band feature. The correction formula is: F′ i =Warp(F i ,H i ), Among them, F′ i represents the i-th band feature after alignment, H i represents the perspective transformation matrix of the i-th band, and Warp(F,H) represents the feature transformation function based on H; The aligned features are fused by multi-layer convolution, and the fused features are calculated using the attention mechanism. The calculation formula is: Among them, F fusion represents the fused global features, Attention(x) represents the attention mechanism function, and w i Represents the weight coefficient of the i-th band feature; The fused features are restored to the output image through the decoder: I out =Decoder(F fusion ), Among them, I out represents the output fused image, and Decoder(x) represents the feature decoder function.

5. The intelligent parking assistance system based on image acquisition as claimed in claim 4, characterized in that: The steps of using the YOLO target detection algorithm to identify dynamic obstacles in the environment are: Based on the fusion image I out , extract the feature vector of the candidate region: F YOLO =Backbone(I out ), Among them, I out represents the input fused image, F YOLO represents the extracted feature vector, Backbone(x) represents the backbone network of YOLO; Use YOLO's anchor box mechanism to predict candidate boxes Location and Category: P k ,C k ,S k =Head(F YOLO ,A k ), in, represents m candidate boxes, P k represents the center position of the kth candidate box, C k represents the category probability of the kth candidate box, S k represents the size of the kth candidate box, A k represents the kth anchor box parameter, Head(x,A) represents the detection head of YOLO, which is used to predict the position, category and size; Use non-maximum suppression NMS to filter the final detection results The screening formula is: in, are the n detection boxes outputted in the end, and NMS(·) is the non-maximum suppression algorithm.

6. The intelligent parking assistance system based on image acquisition as claimed in claim 5, characterized in that: The steps of extracting the parking space contour through the boundary detection algorithm, generating the parking space model, and assigning priority labels to different targets in combination with the dynamic behavior characteristics of the obstacles are as follows: For the fused image I out Perform edge detection and generate a binary boundary image: I edge =Canny(I out ), Among them, I edge represents the binary image after boundary detection, Canny(x) represents the Canny edge detection algorithm, Use contour extraction algorithm to determine the parking space contour: C slot =Contour(I edge ), Among them, C slot represents the detected parking space contour, Contour(x) represents the contour extraction algorithm; Generate a parking space model based on the parking space outline: M slot =Fit(C slot ), Among them, M slot represents the generated parking space model, Fit(x) represents the fitting algorithm used to convert the contour into a parking space geometry model, Analyze the behavior characteristics of the detected dynamic obstacles and calculate the target priority label L based on the priority rule k , the calculation formula is: L k =Priority(D k ,V k ,R k ), Among them, L k represents the priority label of the kth target, D k Indicates the distance between the target and the parking space, V k Indicates the target's moving speed, R k It represents the danger factor of the target, and Priority(x,y,z) represents the priority calculation function.

7. The intelligent parking assistance system based on image acquisition as claimed in claim 6, characterized in that: In step S4, the path planning module dynamically adjusts the path according to the obstacle priority labels.

8. The intelligent parking assistance system based on image acquisition as claimed in claim 7, characterized in that: The steps of using the Bezier curve path planning algorithm to generate a parking path are: Determine the parking starting point P start and parking end point P end : P start =(x s ,y s ),P end =(x e ,y e ), Among them, P start represents the initial position of the current vehicle, P end Indicates the center point of the parking space, (x s ,y s ) and (x e ,y e ) are the coordinates of the starting point and the end point, Determine the set of control points Constrain curve shape: P i =(x i ,y i ), Among them, P i represents the i-th control point, (x i ,y i ) are the coordinates of the control points, and m represents the number of control points; Generate a Bezier path: Among them, B(t) represents the curve point corresponding to parameter t, t is the path parameter, and its value range is [0,1]. represents the number of combinations, defined as According to the obstacle priority label L k Dynamically adjust the path control points, the adjustment formula is: P′ i =P i +ΔP i , Among them, P′ i represents the adjusted control point, ΔP i Represents the adjustment amount, which is calculated based on the obstacle priority. The calculation formula is: Among them, L k is the priority label of the kth target, f k (P i ) represents the kth target control point P i The influence function of n is the number of obstacles detected. Generate the final parking path B′(t): Among them, B′(t) is the dynamically adjusted path.

9. The intelligent parking assistance system based on image acquisition as claimed in claim 8, characterized in that: In step S5, the parking task is executed through closed-loop control: during the execution process, the vehicle control module obtains feedback data in real time and dynamically corrects the control strategy.

10. The intelligent parking assistance system based on image acquisition as claimed in claim 9, characterized in that: The steps of inputting the planned path into the vehicle control module and generating the speed control and steering control instructions of the vehicle are as follows: The dynamically adjusted parking path B′(t) is discretized to obtain a set of discrete path points. The discretization formula is: Among them, {P t } represents a set of discrete path points, T represents the total duration of the parking path, is the kth discrete moment, N is the discrete number of path points, According to the discrete path points, calculate the speed v of the vehicle at each moment k and steering angle θ k , the calculation formula is: Among them, v k represents the speed of the vehicle at the kth moment, θ k represents the steering angle of the vehicle at the kth moment, (x k ,y k ) and (x k+1 ,y k+1 ) is the coordinate of the path point, Δt is the time interval between adjacent path points; Generate control instructions based on the vehicle dynamics model: u k =(v k ,ω k ), Among them, u k represents the kth control instruction, ω k represents the angular velocity of the vehicle at the kth moment, Feedback correction is performed on the execution process through closed-loop control, and control instructions are adjusted in real time: in' k =in k +K e yes k , Among them, u′ k Indicates the modified control instruction, K e is the feedback gain matrix, e k represents the path deviation at the kth moment and is defined as: Among them, P k is the kth path point of the planned path, is the actual position of the vehicle.

Images