Search

CN-122009155-A - Automatic parking control method and system based on image segmentation

CN122009155ACN 122009155 ACN122009155 ACN 122009155ACN-122009155-A

Abstract

The invention discloses an automatic parking control method and system based on image segmentation, which comprises the following steps of firstly extracting a parking space boundary line, secondly constructing a boundary compression vector, thirdly generating a compression trend vector, fourthly switching to a compression response control state when the compression trend in any direction meets a preset trigger condition, fifth generating a non-uniform BEV query point set based on an improved BEVFormer model in the compression response control state, performing time sequence association modeling, and outputting final BEV characteristics, sixth generating a vehicle control signal, seventh switching back to the parking control state if the compression trend in all directions does not meet the preset trigger condition, and eighth ending the automatic parking control process when the vehicle posture is in a stable state. The invention combines compression trend analysis and improvement BEVFormer model to realize fine sensing and stable control of limited space in automatic parking.

Inventors

  • WANG XI
  • SONG JIANMING
  • WANG RUOYU
  • PENG XINLIANG

Assignees

  • 苏州天瞳威视电子科技股份有限公司

Dates

Publication Date
20260512
Application Date
20260225

Claims (10)

  1. 1. An automatic parking control method based on image segmentation is characterized by comprising the following steps: Collecting multi-view image data of the surrounding environment of a vehicle, performing image segmentation on the multi-view image data, and extracting a parking space boundary line; calculating the compression values of the physical contour line of the vehicle and the parking space boundary line in the left measuring, right measuring, front measuring and rear measuring directions, and constructing a boundary compression vector; step three, continuously updating the boundary compression vector for a time, calculating the variation of compression values in all directions, and generating a compression trend vector; judging a current control state according to the compression trend vector, interrupting the parking control state when the compression trend in any direction meets a preset trigger condition, and switching to a compression response control state; In a compression response control state, inputting a current boundary compression vector and a corresponding compression trend vector into an improved BEVFormer model, generating a non-uniform BEV query point set through the boundary compression vector, performing time sequence association modeling through a compression form time sequence memory unit, and outputting final BEV characteristics; Step six, generating a vehicle control signal based on the final BEV characteristics, and adjusting the vehicle posture; step seven, after the vehicle posture adjustment is completed, if the current compression trend in all directions does not meet the preset triggering condition, switching back to the parking control state, otherwise, continuing to execute the compression response control state; and step eight, ending the automatic parking control process when the vehicle posture is in a stable state in the parking control state.
  2. 2. The automatic parking control method based on image segmentation according to claim 1, wherein the first step is specifically: collecting multi-view image data covering the surrounding environment of a vehicle through vehicle-mounted cameras arranged at the front part, the rear part, the left side and the right side of the vehicle, wherein the multi-view image data comprises a forward view image, a backward view image and left and right side view images, and timestamps are uniformly added to the multi-view image data when the multi-view image data is collected so as to form a synchronous image sequence; inputting the multi-view image data into an image segmentation network, the image segmentation network comprising an encoder and a decoder; The encoder divides an input multi-view image into image blocks with set sizes, performs linear embedding processing on each image block, and performs layer-by-layer encoding on the image blocks in a multi-layer transform structure to extract multi-scale semantic features corresponding to different spatial resolutions; inputting the multi-scale semantic features into the decoder, and performing alignment and fusion processing on the semantic features with different scales by the decoder and outputting a pixel classification result; and extracting a pixel set marked as a parking space line category from the pixel classification result, and carrying out connected region screening and edge fitting treatment on the pixel set to obtain a parking space boundary line.
  3. 3. The automatic parking control method based on image segmentation according to claim 1, wherein the step two is specifically: Establishing a vehicle self-coordinate system, wherein the vehicle self-coordinate system takes a vehicle geometric center as an origin, takes a vehicle advancing direction as a longitudinal axis direction and takes a direction perpendicular to the longitudinal axis direction as a transverse axis direction; Determining a vehicle physical contour line in the vehicle self-coordinate system, wherein the vehicle physical contour line is used for representing the outer contour range of a vehicle in the vehicle self-coordinate system; Converting a parking space boundary line from an image coordinate system into the vehicle coordinate system, and representing the parking space boundary line as a boundary point set consisting of a plurality of boundary pixel points; Respectively calculating Euclidean distances between each boundary pixel point in the boundary point set and the vehicle physical contour line in the left side, right side, front side and rear side directions according to the position relation of the boundary pixel point relative to a transverse axis and a longitudinal axis in the vehicle coordinate system by taking the vehicle physical contour line as a reference in the vehicle coordinate system; And respectively selecting minimum distance values from all the calculated Euclidean distances in the left side, right side, front side and rear side directions as compression values in corresponding directions, and sequentially combining to construct a boundary compression vector comprising a left side compression value, a right side compression value, a front side compression value and a rear side compression value.
  4. 4. The automatic parking control method based on image segmentation according to claim 1, wherein the third step is specifically: Continuously sampling the boundary compression vectors at set time intervals to obtain a boundary compression vector sequence arranged in time sequence; Respectively extracting compression values corresponding to two adjacent sampling moments from the boundary compression vector sequence aiming at the left side, the right side, the front side and the rear side; performing differential operation on compression values corresponding to two adjacent sampling moments in each direction to obtain compression variation of the corresponding direction at the current sampling moment; compression variation amounts obtained in the left, right, front and rear directions are sequentially combined to generate a compression trend vector.
  5. 5. The automatic parking control method based on image segmentation according to claim 1, wherein the step four is specifically: For the compression trend vector, compression variation amounts corresponding to the left side, the right side, the front side and the rear side directions are respectively obtained; Setting a change threshold value, wherein the change threshold value is used for limiting the allowable compression change range between two adjacent sampling moments; In any direction, when the absolute value of the corresponding compression variation quantity at the current sampling moment is larger than the variation threshold value, judging that a preset triggering condition is met; and when at least one direction meets the preset triggering condition, interrupting the current parking control state and switching to the compression response control state.
  6. 6. The automatic parking control method based on image segmentation according to claim 1, wherein the modified BEVFormer model includes a compression state encoding unit, a non-uniform BEV query point generating unit, a compression guidance cross-view aggregation unit, a compression shape time sequence memory unit, and a final BEV feature output unit; The compression state coding unit is used for receiving the boundary compression vector and the compression trend vector, matching compression values corresponding to each direction with compression variation according to the direction sequence of the left side, the right side, the front side and the rear side, and splicing the compression values and the compression variation in sequence to form a compression state sequence; Constructing a position index based on the sequential positions of the elements in each direction in the compression state sequence, converting the position index into a position coding vector in a linear mapping mode, and adding the position coding vector and the compression state sequence element by element to obtain a compression embedded vector; The non-uniform BEV query point generation unit generates a regular grid query point set by taking a vehicle coordinate system as a reference in a BEV plane, wherein the regular grid query point set is uniformly sampled at set intervals in the longitudinal direction and the transverse direction; performing coordinate remapping processing on the regular grid query point set based on compression values corresponding to the left side, the right side, the front side and the rear side in the boundary compression vector, multiplying coordinate components of the regular grid query points by coordinate scaling factors, wherein the coordinate scaling factors are the inverse of the sum of the compression values and set smoothing factors, and generating a non-uniform BEV query point set; the compression guide cross-view aggregation unit takes the non-uniform BEV query point set as input, performs cross-view feature aggregation processing on multi-view image features, and generates BEV intermediate features; The compression form time sequence memory unit is used for performing time sequence association modeling on the BEV intermediate features at continuous moments, storing BEV intermediate features corresponding to historical moments according to time sequences, and storing boundary compression vectors of corresponding moments for the BEV intermediate features of each historical moment in an associated mode; performing element-by-element differential operation on the boundary compression vector corresponding to the current moment and each boundary compression vector stored in the historical moment, and calculating the differential distance between the current boundary compression vector and each historical boundary compression vector; Selecting at least one historical moment corresponding to the minimum differential distance from the differential distances, and taking BEV intermediate features corresponding to the historical moment as historical features matched with the current compression form; splicing the matched historical features and BEV intermediate features at the current moment in feature dimensions to form BEV features with enhanced time sequence; The final BEV feature output unit performs linear transformation on the time sequence enhanced BEV features in a channel dimension, and performs channel-by-channel normalization processing to output final BEV features.
  7. 7. The automatic parking control method based on image segmentation according to claim 6, wherein the cross-view feature aggregation processing is performed on the multi-view image features to generate BEV intermediate features, specifically: mapping each non-uniform BEV query point to each view angle image plane according to the two-dimensional coordinate position of each non-uniform BEV query point in the vehicle coordinate system to obtain the corresponding projection coordinate of the non-uniform BEV query point in each view angle image; In each view angle image, taking the projection coordinates as the center, sampling local feature vectors with set sizes from the image feature images of the corresponding view angles, and splicing the local feature vectors from different view angles according to the view angle sequence to form a cross-view angle feature sequence corresponding to the non-uniform BEV query point; Performing attention calculation operation on the cross-view feature sequence, converting the compression embedded vector into an attention modulation vector through linear mapping, and multiplying the attention modulation vector with the cross-view feature sequence element by element so as to perform weight modulation on local feature vectors of all views in the cross-view feature sequence; Performing summation operation on the cross-view angle characteristic sequence subjected to weight modulation to obtain BEV space characteristics corresponding to the nonuniform BEV query points; all BEV spatial features are arranged to form BEV intermediate features according to the spatial position of the non-uniform BEV query point in the BEV plane.
  8. 8. The automatic parking control method based on image segmentation according to claim 1, wherein the sixth step is specifically: inputting the final BEV features into a vehicle control generation network, the vehicle control generation network comprising a feature mapping layer and a control output layer; The feature mapping layer performs linear transformation processing on the final BEV feature in a channel dimension, and performs nonlinear mapping on a result obtained after linear transformation by applying a ReLU function element by element to obtain a control feature vector; inputting the control feature vector to a control output layer, wherein the control output layer comprises two linear mapping units, and respectively generating steering control quantity and displacement control quantity in a linear mapping mode; and combining the steering control quantity and the displacement control quantity to be used as a vehicle control signal to be output, controlling the steering angle change and the longitudinal movement of the vehicle, and adjusting the vehicle posture.
  9. 9. The automatic parking control method based on image segmentation according to claim 1, wherein the step eight is specifically: Continuously acquiring attitude parameters of a vehicle in a plurality of control periods in a parking control state, wherein the attitude parameters comprise a vehicle course angle and a vehicle mass center position; respectively performing differential operation on the vehicle course angle and the vehicle centroid position in adjacent control periods to obtain corresponding course angle variation and centroid position variation; When the course angle variation is smaller than the set course angle variation stability threshold value in the continuously set number of control periods and the position variation is smaller than the set centroid position stability threshold value in the continuously set number of control periods, the vehicle posture is judged to be in a stable state, a parking completion instruction is output, and the automatic parking control process is ended.
  10. 10. An image segmentation-based automatic parking control system that performs an image segmentation-based automatic parking control method according to any one of claims 1 to 9, characterized by comprising the following modules: the image acquisition and segmentation module is used for acquiring multi-view image data of the surrounding environment of the vehicle, carrying out image segmentation processing on the multi-view image data and extracting a parking space boundary line; The compression calculation module is used for calculating compression values of the physical contour line of the vehicle and the parking space boundary line in the left measurement, right measurement, front measurement and rear measurement directions and constructing a boundary compression vector; The compression trend analysis module is used for continuously updating the boundary compression vector, calculating the variation of compression values in all directions and generating a compression trend vector; The control state judging module is used for judging the current control state based on the compression trend vector, interrupting the parking control state when the compression trend in any direction meets the preset triggering condition, and switching to the compression response control state; The BEV feature generation module is used for inputting the current boundary compression vector and the corresponding compression trend vector into the improved BEVFormer model under the compression response control state, generating a non-uniform BEV query point set through the boundary compression vector, carrying out time sequence association modeling through the compression form time sequence memory unit, and outputting the final BEV feature; A vehicle control signal generation module for generating a vehicle control signal based on the final BEV characteristics, the vehicle control signal including a steering control amount and a displacement control amount for performing vehicle attitude adjustment; The parking control execution module is used for controlling the vehicle to execute steering and longitudinal movement according to the vehicle control signal, switching back to a parking control state if the pressing trend in all the current directions does not meet the preset triggering condition after finishing one-time vehicle posture adjustment, and otherwise, continuing to execute the pressing response control state; the parking termination judging module is used for continuously acquiring the posture parameters of the vehicle in a plurality of control periods under the parking control state, judging whether the posture of the vehicle is in a stable state, outputting a parking completion instruction when the posture of the vehicle is judged to be in the stable state, and ending the automatic parking control process.

Description

Automatic parking control method and system based on image segmentation Technical Field The invention relates to the technical field of intelligent control of vehicles, in particular to an automatic parking control method and system based on image segmentation. Background Along with the rapid development of automatic driving technology and the increasingly tense urban road parking space, how to realize safe, efficient and stable automatic parking control becomes the research focus in the field of intelligent vehicles, the existing automatic parking system mainly carries out parking control by constructing a geometric model between a vehicle and a parking space and combining sensing means such as infrared, ultrasonic, radar or cameras based on a regular path planning and sensor fusion sensing strategy. However, in practical applications, the following problems are common in the current technology: on one hand, the traditional automatic parking method mainly relies on rule modeling or a learnable BEV space construction means, so that the dynamic space relation between a vehicle and a parking space boundary is difficult to fully utilize, particularly when a scene with uneven space compression and frequent posture adjustment is encountered in the parking process, the control strategy is easy to generate unstable and repeated correction and the like, the parking efficiency and the safety are influenced, and on the other hand, the traditional automatic parking system based on image recognition generally adopts a BEV query point or an average strategy with fixed distribution to carry out space modeling, lacks the dynamic attention capability on the limited direction of the vehicle, causes the problems of uneven space sampling, fuzzy characteristic expression and the like in a narrow parking space or complex boundary environment, and is difficult to support high-precision control signal generation. In addition, the association relationship between the multi-view image information and the physical contour of the vehicle is not modeled in a structuring way, and the dynamic information of the time dimension such as the compression trend change is difficult to be introduced into the control strategy by the existing method, so that the judgment and response capability of the system to the gesture change trend are limited, and particularly, the problems of perception delay and action delay are presented in a parking process requiring multiple gesture adjustment. Therefore, how to provide an automatic parking control method and system based on image segmentation is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide an automatic parking control method and system based on image segmentation, wherein the method combines compression trend analysis and a BEVFormer model improvement, introduces a dynamic compression relation between a physical outline of a vehicle and a parking space boundary in an automatic parking process, realizes high-resolution sensing and continuous feature expression of a limited space region through non-uniform BEV query point generation and compression form time sequence associated modeling, and generates a stable and controllable vehicle control signal according to the high-resolution sensing and continuous feature expression, so that the control precision, the process stability and the overall safety of automatic parking in a narrow environment are effectively improved. According to the embodiment of the invention, the automatic parking control method based on image segmentation comprises the following steps: Collecting multi-view image data of the surrounding environment of a vehicle, performing image segmentation on the multi-view image data, and extracting a parking space boundary line; calculating the compression values of the physical contour line of the vehicle and the parking space boundary line in the left measuring, right measuring, front measuring and rear measuring directions, and constructing a boundary compression vector; step three, continuously updating the boundary compression vector for a time, calculating the variation of compression values in all directions, and generating a compression trend vector; judging a current control state according to the compression trend vector, interrupting the parking control state when the compression trend in any direction meets a preset trigger condition, and switching to a compression response control state; In a compression response control state, inputting a current boundary compression vector and a corresponding compression trend vector into an improved BEVFormer model, generating a non-uniform BEV query point set through the boundary compression vector, performing time sequence association modeling through a compression form time sequence memory unit, and outputting final BEV characteristics; Step six, generating a vehicle control signal based on the final BEV characteristics, a