Automatic parking control method and system based on image segmentation

Abstract

The invention discloses an automatic parking control method and system based on image segmentation. The method comprises the following steps of 1, extracting a parking space boundary line; 2, constructing a boundary compression vector; 3, generating a compression trend vector; 4, when the compression trend in any direction meets a preset triggering condition, switching to a compression response control state; 5, in a compression response control state, generating a non-uniform BEV query point set based on an improved BEV Former model, performing time sequence correlation modeling, and outputting a final BEV feature; 6, generating a vehicle control signal; 7, if the oppression trends in all directions do not meet the preset triggering condition, the parking control state is switched back; and 8, when the posture of the vehicle is in a stable state, the automatic parking control process is ended. According to the method, the compression trend analysis and the improved BEVFormer model are combined, and fine sensing and stable control of the limited space in automatic parking are achieved.

Description

Technical Field

[0001] This invention relates to the field of vehicle intelligent control technology, and in particular to an automatic parking control method and system based on image segmentation. Background Technology

[0002] With the rapid development of autonomous driving technology and the increasing scarcity of parking spaces in urban areas, achieving safe, efficient, and stable automatic parking control has become a key research focus in the field of intelligent vehicles. Existing automatic parking systems primarily rely on rule-based path planning and sensor fusion perception strategies. They construct geometric models between the vehicle and parking space, combining perception methods such as infrared, ultrasonic, radar, or cameras for parking control. However, in practical applications, current technologies generally suffer from the following problems: On the one hand, traditional automatic parking methods mainly rely on rule-based modeling or learnable BEV space construction, which makes it difficult to fully utilize the dynamic spatial relationship between the vehicle and the parking space boundary. Especially when encountering scenarios with uneven spatial compression and frequent posture adjustments during parking, the control strategy is prone to instability and repeated corrections, affecting parking efficiency and safety. On the other hand, existing image recognition-based automatic parking systems usually use fixed-distribution BEV query points or averaging strategies for spatial modeling, lacking the ability to dynamically focus on the vehicle's restricted direction. This leads to problems such as uneven spatial sampling and fuzzy feature representation in narrow parking spaces or complex boundary environments, making it difficult to support the generation of high-precision control signals.

[0003] Furthermore, the relationship between multi-view image information and the vehicle's physical contours has not been structurally modeled. Existing methods struggle to incorporate dynamic information such as changes in pressure trends over time into the control strategy, thus limiting the system's ability to judge and respond to attitude change trends. This is particularly evident in parking processes that require multiple attitude adjustments, where the system exhibits perception delays and sluggish movements.

[0004] Therefore, how to provide an automatic parking control method and system based on image segmentation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an automatic parking control method and system based on image segmentation. This invention combines compression trend analysis and an improved BEVFormer model to introduce the dynamic compression relationship between the vehicle's physical contour and the parking space boundary during the automatic parking process. By generating non-uniform BEV query points and modeling the temporal correlation of compression patterns, it achieves high-resolution perception and continuous feature representation of confined space areas, and generates stable and controllable vehicle control signals accordingly. This effectively improves the control accuracy, process stability, and overall safety of automatic parking in narrow environments.

[0006] An automatic parking control method based on image segmentation according to an embodiment of the present invention includes the following steps: Step 1: Collect multi-view image data of the vehicle's surrounding environment, perform image segmentation on the multi-view image data, and extract the parking space boundary lines; Step 2: Calculate the compression values ​​between the vehicle's physical outline and the parking space boundary line in the left, right, front, and rear directions, and construct the boundary compression vector; Step 3: Continuously update the boundary compression vector over time, calculate the change in compression value in each direction, and generate a compression trend vector; Step 4: Determine the current control state based on the pressure trend vector. When the pressure trend in any direction meets the preset trigger condition, interrupt the parking control state and switch to the pressure response control state. Step 5: Under the compression response control state, input the current boundary compression vector and the corresponding compression trend vector into the improved BEVFormer model, generate a non-uniform BEV query point set through the boundary compression vector, and perform temporal correlation modeling through the compression morphology temporal memory unit to output the final BEV features; Step 6: Generate vehicle control signals based on the final BEV characteristics to adjust the vehicle attitude; Step 7: After completing the vehicle attitude adjustment, if the current pressure trend in all directions does not meet the preset trigger conditions, switch back to parking control state; otherwise, continue to execute pressure response control state. Step 8: In parking control mode, when the vehicle's attitude is stable, the automatic parking control process ends.

[0007] Optionally, step one specifically includes: By using vehicle-mounted cameras installed at the front, rear, left, and right sides of the vehicle, multi-view image data covering the surrounding environment of the vehicle is collected. The multi-view image data includes forward view images, rear view images, and left and right side view images. The multi-view image data is uniformly timestamped during collection to form a synchronized image sequence. The multi-view image data is input into an image segmentation network, which includes an encoder and a decoder; The encoder divides the input multi-view image into image blocks of a set size, performs linear embedding processing on each image block, and encodes the image blocks layer by layer in a multi-layer Transformer structure to extract multi-scale semantic features corresponding to different spatial resolutions. The multi-scale semantic features are input into the decoder, which aligns and fuses the semantic features at different scales and outputs pixel-level classification results. Extract the set of pixels labeled as parking line category from the pixel-level classification results, and perform connected component filtering and edge fitting on the pixel set to obtain the parking space boundary line.

[0008] Optionally, step two specifically includes: Establish a vehicle coordinate system with the vehicle's geometric center as the origin, the vehicle's forward direction as the vertical axis, and the direction perpendicular to the vertical axis as the horizontal axis. The vehicle physical contour line is determined in the vehicle's own coordinate system, and the vehicle physical contour line is used to characterize the outer contour range of the vehicle in the vehicle's own coordinate system. The parking space boundary line is transformed from the image coordinate system to the vehicle's own coordinate system, and the parking space boundary line is represented as a set of boundary points composed of multiple boundary pixels; In the vehicle's own coordinate system, with the vehicle's physical outline as the reference, the Euclidean distance between each boundary pixel in the set of boundary points and the vehicle's physical outline is calculated in the left, right, front, and rear directions, based on the positional relationship of the boundary pixel points relative to the horizontal and vertical axes in the vehicle's own coordinate system. In the left, right, front, and rear directions, the minimum distance value is selected from all the calculated Euclidean distances as the compression value for the corresponding direction, and they are combined in order to construct a boundary compression vector including the left compression value, right compression value, front compression value, and rear compression value.

[0009] Optionally, step three specifically includes: By continuously sampling the boundary compression vector at set time intervals, a sequence of boundary compression vectors arranged in chronological order is obtained; For the left, right, front, and rear directions, the compression values ​​corresponding to two adjacent sampling times are extracted from the boundary compression vector sequence, respectively. Perform a differential operation on the pressure values ​​corresponding to two adjacent sampling times in each direction to obtain the pressure change in the corresponding direction at the current sampling time; The compression changes obtained from the left, right, front, and rear directions are combined in sequence to generate a compression trend vector.

[0010] Optionally, step four specifically includes: For the compression trend vector, the compression change amounts corresponding to the left, right, front, and rear directions are obtained respectively; A change threshold is set, which is used to limit the allowable range of pressure change between two adjacent sampling times; In any direction, when the absolute value of the corresponding pressure change at the current sampling time is greater than the change threshold, it is determined that the preset triggering condition is met; When at least one direction meets the preset triggering condition, the current parking control state is interrupted and switched to the pressure response control state.

[0011] Optionally, the improved BEVFormer model includes a compression state encoding unit, a non-uniform BEV query point generation unit, a compression-guided cross-view aggregation unit, a compression morphology temporal memory unit, and a final BEV feature output unit. The compression state encoding unit is used to receive the boundary compression vector and the compression trend vector, and pair the compression value corresponding to each direction with the compression change amount in the order of left, right, front and back, and splice them together in sequence to form a compression state sequence. Based on the sequential position of each element in the compression state sequence, a position index is constructed, and the position index is converted into a position encoding vector through a linear mapping method. The position encoding vector is then added element by element to the compression state sequence to obtain a compression embedding vector. The non-uniform BEV query point generation unit generates a set of regular grid query points in the BEV plane, with the vehicle's own coordinate system as a reference. The set of regular grid query points is uniformly sampled at set intervals in both the longitudinal and lateral directions. Based on the compression values ​​corresponding to the left, right, front, and rear directions in the boundary compression vector, the set of regular grid query points is remapped. The coordinate components of the regular grid query points are multiplied by a coordinate scaling factor, which is the reciprocal of the sum of the compression value and a set smoothing factor, to generate a non-uniform BEV query point set. The compression-guided cross-view aggregation unit takes the non-uniform BEV query point set as input and performs cross-view feature aggregation processing on multi-view image features to generate intermediate BEV features. The compression morphology temporal memory unit is used to perform temporal correlation modeling of the BEV intermediate features at continuous time intervals, store the BEV intermediate features corresponding to historical time intervals in chronological order, and associate and store the boundary compression vector of the corresponding time interval for the BEV intermediate features of each historical time interval. Perform element-wise difference operations between the current boundary compression vector and each historical boundary compression vector stored at the current time, and calculate the difference distance between the current boundary compression vector and each historical boundary compression vector. Select at least one historical moment corresponding to the minimum difference distance from the difference distances, and use the BEV intermediate feature corresponding to the historical moment as the historical feature that matches the current compression pattern; The matched historical features are concatenated with the intermediate BEV features at the current moment along the feature dimension to form time-enhanced BEV features. The final BEV feature output unit performs a linear transformation on the time-enhanced BEV features along the channel dimension and performs channel-by-channel normalization processing to output the final BEV features.

[0012] Optionally, the step of performing cross-view feature aggregation processing on multi-view image features to generate BEV intermediate features specifically involves: Based on the two-dimensional coordinate position of each non-uniform BEV query point in the vehicle's own coordinate system, the non-uniform BEV query point is mapped to the image plane of each viewpoint to obtain the corresponding projected coordinates of the non-uniform BEV query point in each viewpoint image. In each viewpoint image, local feature vectors of a set size are sampled from the image feature map of the corresponding viewpoint, with the projection coordinates as the center, and the local feature vectors from different viewpoints are spliced ​​together in viewpoint order to form a cross-viewpoint feature sequence corresponding to the non-uniform BEV query point. An attention calculation operation is performed on the cross-view feature sequence, the compression embedding vector is converted into an attention modulation vector through a linear mapping, and the attention modulation vector is multiplied element-wise with the cross-view feature sequence to perform weight modulation on the local feature vectors of each view in the cross-view feature sequence. The weighted cross-view feature sequence is summed to obtain the BEV spatial features corresponding to the non-uniform BEV query point. Based on the spatial location of the non-uniform BEV query point within the BEV plane, all BEV spatial features are arranged to form intermediate BEV features.

[0013] Optionally, step six specifically includes: The final BEV features are input into the vehicle control generation network, which includes a feature mapping layer and a control output layer. The feature mapping layer performs a linear transformation on the final BEV feature along the channel dimension, and applies a ReLU function to each element of the linearly transformed result for nonlinear mapping to obtain the control feature vector. The control feature vector is input to the control output layer, which includes two linear mapping units that generate steering control and displacement control quantities respectively through linear mapping. The steering control quantity and the displacement control quantity are combined as a vehicle control signal output to control the change of the vehicle's steering angle and longitudinal movement, thereby adjusting the vehicle's attitude.

[0014] Optionally, step eight specifically includes: In parking control mode, the vehicle's attitude parameters are continuously acquired over multiple control cycles, including the vehicle's heading angle and the vehicle's center of gravity position. Perform differential calculations on the vehicle heading angle and the vehicle center of gravity position within adjacent control cycles to obtain the corresponding changes in heading angle and center of gravity position. When the change in heading angle is less than the set heading angle change stabilization threshold for a set number of consecutive control cycles, and the change in position is less than the set center of mass position stabilization threshold for a set number of consecutive control cycles, the vehicle attitude is determined to be stable, a parking completion command is output, and the automatic parking control process ends.

[0015] An automatic parking control system based on image segmentation according to an embodiment of the present invention includes the following modules: The image acquisition and segmentation module is used to acquire multi-view image data of the environment around the vehicle, and to perform image segmentation processing on the multi-view image data to extract the parking space boundary lines. The compression calculation module is used to calculate the compression values ​​between the vehicle's physical outline and the parking space boundary line in the left, right, front, and rear directions, and to construct the boundary compression vector. The compression trend analysis module is used to continuously update the boundary compression vector over time, calculate the change in compression value in each direction, and generate a compression trend vector. The control state determination module is used to determine the current control state based on the pressure trend vector. When the pressure trend in any direction meets the preset trigger condition, the parking control state is interrupted and the pressure response control state is switched. The BEV feature generation module is used to input the current boundary compression vector and the corresponding compression trend vector into the improved BEVFormer model under the compression response control state. It generates a non-uniform BEV query point set through the boundary compression vector, performs temporal correlation modeling through the compression morphology temporal memory unit, and outputs the final BEV features. A vehicle control signal generation module is used to generate vehicle control signals based on the final BEV characteristics. The vehicle control signals include steering control quantities and displacement control quantities, which are used to adjust the vehicle attitude. The parking control execution module is used to control the vehicle to perform steering and longitudinal movement according to the vehicle control signal. After completing a vehicle posture adjustment, if the current pressure trend in all directions does not meet the preset trigger condition, it switches back to the parking control state; otherwise, it continues to execute the pressure response control state. The parking termination determination module is used to continuously acquire the vehicle's attitude parameters within multiple control cycles during parking control, determine whether the vehicle's attitude is in a stable state, and output a parking completion command when the vehicle's attitude is determined to be in a stable state, thereby ending the automatic parking control process.

[0016] The beneficial effects of this invention are: This invention addresses the problems of low perception resolution, unstable attitude adjustment, and lag in control response in confined spaces in existing automatic parking control methods by synergistically integrating a pressure trend analysis module with an improved BEVFormer model. It constructs a dynamic geometric state representation between the vehicle and the parking space boundary using boundary pressure vectors and pressure trend vectors, and combines this with a control state determination module to achieve real-time monitoring of attitude evolution trends and control state switching during parking. Furthermore, the BEV feature generation module introduces a non-uniform BEV query point generation unit and a pressure morphology temporal memory unit to adaptively adjust the BEV based on the vehicle's current pressure state. Spatial query distribution is used, and intermediate features of the matching BEV are retrieved based on the similarity of historical compression patterns to achieve high-resolution aggregation and enhanced temporal consistency of multi-view image features. In the vehicle control signal generation module, steering control and displacement control quantities are generated through the serial structure of the feature mapping layer and the control output layer, respectively, to achieve decoupling of vehicle lateral and longitudinal control and improve attitude adjustment accuracy. In the parking termination determination module, the stability of the heading angle and center of gravity position is judged by continuous periodic difference, and the parking completion command is reliably output, thereby realizing a closed loop of automatic parking control that is dynamically adjustable throughout the process, with timely compression response and stable control accuracy. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of an automatic parking control method based on image segmentation proposed in this invention; Figure 2 This is a schematic diagram of the structure of an automatic parking control system based on image segmentation proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 An automatic parking control method based on image segmentation includes the following steps: Step 1: Collect multi-view image data of the vehicle's surrounding environment, perform image segmentation on the multi-view image data, and extract the parking space boundary lines; Step 2: Calculate the compression values ​​between the vehicle's physical outline and the parking space boundary line in the left, right, front, and rear directions, and construct the boundary compression vector; Step 3: Continuously update the boundary compression vector over time, calculate the change in compression value in each direction, and generate a compression trend vector; Step 4: Determine the current control state based on the pressure trend vector. When the pressure trend in any direction meets the preset trigger condition, interrupt the parking control state and switch to the pressure response control state. Step 5: Under the compression response control state, input the current boundary compression vector and the corresponding compression trend vector into the improved BEVFormer model, generate a non-uniform BEV query point set through the boundary compression vector, and perform temporal correlation modeling through the compression morphology temporal memory unit to output the final BEV features; Step 6: Generate vehicle control signals based on the final BEV characteristics to adjust the vehicle attitude; Step 7: After completing the vehicle attitude adjustment, if the current pressure trend in all directions does not meet the preset trigger conditions, switch back to parking control state; otherwise, continue to execute pressure response control state. Step 8: In parking control mode, when the vehicle's attitude is stable, the automatic parking control process ends.

[0020] In this embodiment, step one specifically includes: By using vehicle-mounted cameras installed at the front, rear, left, and right sides of the vehicle, multi-view image data covering the surrounding environment of the vehicle is collected. The multi-view image data includes forward view images, rear view images, and left and right side view images. The multi-view image data is uniformly timestamped during collection to form a synchronized image sequence. The multi-view image data is input into an image segmentation network, which includes an encoder and a decoder; The encoder divides the input multi-view image into image blocks of a set size, performs linear embedding processing on each image block, and encodes the image blocks layer by layer in a multi-layer Transformer structure to extract multi-scale semantic features corresponding to different spatial resolutions. In this invention, after receiving the input multi-view image, the encoder first performs block processing on each input image, dividing the input image into multiple non-overlapping image blocks according to a preset size, wherein the preset size is set to 16×16 pixels.

[0021] For each image block, a linear embedding process is performed, specifically: the pixel values ​​in the image block are expanded in channel order to form a one-dimensional vector, and the one-dimensional vector is converted into an embedding vector through a linear mapping, wherein the dimension of the embedding vector is set to 64 dimensions to realize the mapping of the image block from the pixel space to the feature space.

[0022] The encoder includes a multi-layer Transformer encoding structure. Each Transformer layer sequentially encodes the embedding vector. During the encoding process of each Transformer layer, the feature relationships between different image patches are modeled through a self-attention mechanism, and a feedforward network is used to perform nonlinear transformation on the feature representation, thereby enhancing the semantic expressive power of the image patches layer by layer. As the number of Transformer layers increases, the encoder's feature representation of the image patches gradually transitions from local texture features to global semantic features.

[0023] In the Transformer encoding process at different levels, the encoder outputs feature representations corresponding to different spatial resolutions. The Transformer layer output closer to the input maintains a higher spatial resolution and is used to characterize the local structural information of the image, while the Transformer layer output closer to the output has a lower spatial resolution and is used to characterize the global semantic information of the image, thus forming multi-scale semantic features. The multi-scale semantic features are input into the decoder, which aligns and fuses the semantic features at different scales and outputs pixel-level classification results. In this invention, the decoder receives multi-scale semantic features output from the encoder and performs spatial alignment and feature fusion processing on the multi-scale semantic features. Since the semantic features of different scales have different spatial resolutions, the decoder first performs upsampling processing on the low-resolution semantic features to make their spatial size consistent with the high-resolution semantic features. The upsampling processing adopts the bilinear interpolation method. The aligned multi-scale semantic features form a correspondence in spatial position so as to facilitate subsequent fusion.

[0024] After spatial alignment is completed, the decoder concatenates semantic features of different scales along the feature channel direction and processes the concatenated features through a convolutional fusion module, which includes a convolutional operation and a ReLU activation function to generate a unified fused feature representation.

[0025] Based on the fused feature representation, the decoder generates pixel-level classification results through a classification head. The classification head includes a convolutional layer to map the fused feature representation to the number of channels corresponding to the number of categories. The softmax function is used to normalize the channel output of each pixel position to obtain the probability distribution of each pixel belonging to different categories. The category label corresponding to each pixel is determined according to the probability distribution to obtain the final pixel-level classification result. Extract the set of pixels labeled as parking line category from the pixel-level classification results, and perform connected component filtering and edge fitting on the set of pixels to obtain the parking space boundary line. Specifically, the pixel set is grouped using an eight-neighbor connectivity determination method, dividing spatially connected pixels into the same connected region. For each connected region, the number of pixels it contains is calculated. When the number of pixels in a connected region is less than a preset threshold, the connected region is determined to be a noise region and is removed. The preset threshold is set to 100 pixels to eliminate discrete pixel regions caused by image noise or missegmentation.

[0026] After completing the connected region filtering, edge extraction processing is performed on the retained connected regions. Specifically, the set of pixels located at the region boundary in the connected region is extracted to form the initial edge point set of the parking space line. Based on the initial edge point set, the edge points are fitted using the least squares curve fitting method to represent the discrete edge points as continuous line segments or curves. In the case of straight parking space lines, a straight line fitting method is used to generate the parking space boundary line. In the case of curved or perspective distortion, a quadratic polynomial fitting method is used to generate the parking space boundary line.

[0027] In this embodiment, step two specifically includes: Establish a vehicle coordinate system with the vehicle's geometric center as the origin, the vehicle's forward direction as the vertical axis, and the direction perpendicular to the vertical axis as the horizontal axis. The vehicle physical contour line is determined in the vehicle's own coordinate system, and the vehicle physical contour line is used to characterize the outer contour range of the vehicle in the vehicle's own coordinate system. The parking space boundary line is transformed from the image coordinate system to the vehicle's own coordinate system, and the parking space boundary line is represented as a set of boundary points composed of multiple boundary pixels; In the vehicle's own coordinate system, with the vehicle's physical outline as the reference, the Euclidean distance between each boundary pixel in the set of boundary points and the vehicle's physical outline is calculated in the left, right, front, and rear directions, based on the positional relationship of the boundary pixel points relative to the horizontal and vertical axes in the vehicle's own coordinate system. In the left, right, front, and rear directions, the minimum distance value is selected from all the calculated Euclidean distances as the compression value in the corresponding direction, and they are combined in order to construct a boundary compression vector including the left compression value, right compression value, front compression value, and rear compression value. In this invention, in order to achieve a unified calculation of the spatial relationship between the vehicle and the parking space boundary, a vehicle coordinate system is first established to provide a unified reference frame for the space around the vehicle. The vehicle's physical outline is determined in the vehicle coordinate system to represent the outer outline boundary of the vehicle in the current posture.

[0028] The parking space boundary line obtained by image segmentation is first transformed from the image coordinate system to the vehicle's own coordinate system and represented as a set of boundary points composed of multiple discrete boundary pixels. For the set of boundary points, according to the positional relationship of each boundary pixel relative to the horizontal and vertical axes in the vehicle's own coordinate system, the boundary pixels are respectively divided into the directional regions corresponding to the left, right, front and rear sides of the vehicle.

[0029] Within each directional region, using the vehicle's physical outline as the distance calculation benchmark, the Euclidean distance between each boundary pixel in its respective direction and the vehicle's physical outline is calculated. The smallest Euclidean distance within the same direction is then selected as the spatial compression value for that direction. Finally, the compression values ​​in the left, right, front, and rear directions are combined sequentially to form a boundary compression vector characterizing the spatial relationship between the vehicle and the parking space boundary, providing a basic input for compression trend analysis and control decisions.

[0030] In this embodiment, step three specifically includes: By continuously sampling the boundary compression vector at set time intervals, a sequence of boundary compression vectors arranged in chronological order is obtained; For the left, right, front, and rear directions, the compression values ​​corresponding to two adjacent sampling times are extracted from the boundary compression vector sequence, respectively. Perform a differential operation on the pressure values ​​corresponding to two adjacent sampling times in each direction to obtain the pressure change in the corresponding direction at the current sampling time; The compression changes obtained from the left, right, front, and rear directions are combined in sequence to generate a compression trend vector; In this invention, the boundary compression vector is continuously updated over time to generate a compression trend vector. Specifically, the system samples the boundary compression vector at a fixed time interval, which is set to 50 milliseconds, thereby forming a sequence of boundary compression vectors arranged in chronological order. Each boundary compression vector includes compression values ​​corresponding to the four directions: left, right, front, and back.

[0031] Between two adjacent sampling times, the pressure value corresponding to the same direction is extracted, and the pressure value of the previous sampling time and the pressure value of the current sampling time are differentially calculated to obtain the pressure change in that direction at the current sampling time.

[0032] The compression changes obtained from the left, right, front, and rear directions are combined in sequence to form a compression trend vector, which is used to characterize the changes of the boundary compression vector in the time dimension and serves as the basic data for control state judgment and model input.

[0033] In this embodiment, step four specifically includes: For the compression trend vector, the compression change amounts corresponding to the left, right, front, and rear directions are obtained respectively; A change threshold is set, which is used to limit the allowable range of pressure change between two adjacent sampling times; In any direction, when the absolute value of the corresponding pressure change at the current sampling time is greater than the change threshold, it is determined that the preset triggering condition is met; When at least one direction meets the preset triggering condition, the current parking control state is interrupted and switched to the pressure response control state; In this invention, the triggering of the pressure response control state is determined based on the amount of pressure change in the pressure trend vector. For the left, right, front, and rear directions, the system obtains the amount of pressure change in the corresponding direction from the pressure trend vector and sets a change threshold. The change threshold is a fixed value used to limit the allowable range of pressure change between two adjacent sampling times. The change threshold of this invention is set to 0.10 meters and can be adjusted according to the vehicle size and parking scenario. The amount of pressure change and the corresponding change threshold are compared under the same time reference.

[0034] When the absolute value of the pressure change in any direction at the current sampling time is greater than the change threshold, the system determines that the direction meets the triggering condition of the pressure response control state. If at least one direction meets the triggering condition, the system outputs a control state switching command to interrupt the current parking control state and enter the pressure response control state, so that the control process continues to be executed based on the pressure response mechanism. In this invention, the current parking control state is the basic control state during the automatic parking process, which is used to perform normal parking control of the vehicle when the pressure response control state is not triggered. In this parking control state, vehicle control is based on the parking space boundary, the current attitude of the vehicle, and multi-sensor perception information.

[0035] Specifically, the system continuously acquires perception data of the vehicle's surrounding environment while in parking control mode. This includes multi-view image data obtained from the vehicle-mounted camera, short-range ranging data obtained from ultrasonic sensors, and medium-to-long-range ranging data obtained from radar sensors. The ultrasonic sensor data is used to provide distance information between the vehicle and objects within a range of 0 to 3 meters, while the radar sensor data is used to provide distance and orientation information between the vehicle and targets within a range greater than 3 meters but not exceeding 30 meters.

[0036] In parking control mode, ultrasonic sensor data and radar sensor data are monitored in real time to verify whether there are physical obstacles around the vehicle and to obtain the spatial distance between the vehicle and the physical obstacles.

[0037] When the boundary compression trend detected in the parking control state does not meet the preset trigger condition, the system maintains the parking control state unchanged and continues to output control signals according to the established parking control logic; when the compression trend vector is subsequently detected to meet the preset trigger condition, the system interrupts the current parking control state and switches to the compression response control state to execute response control for changes in local spatial constraints.

[0038] In this embodiment, the improved BEVFormer model includes a compression state encoding unit, a non-uniform BEV query point generation unit, a compression-guided cross-view aggregation unit, a compression morphology temporal memory unit, and a final BEV feature output unit. The compression state encoding unit is used to receive the boundary compression vector and the compression trend vector, and pair the compression value corresponding to each direction with the compression change amount in the order of left, right, front and back, and splice them together in sequence to form a compression state sequence. Based on the sequential position of each element in the compression state sequence, a position index is constructed, and the position index is converted into a position encoding vector through a linear mapping method. The position encoding vector is then added element by element to the compression state sequence to obtain a compression embedding vector. The non-uniform BEV query point generation unit generates a set of regular grid query points in the BEV plane, with the vehicle's own coordinate system as a reference. The set of regular grid query points is uniformly sampled at set intervals in both the longitudinal and lateral directions. Based on the compression values ​​corresponding to the left, right, front, and rear directions in the boundary compression vector, the set of regular grid query points is remapped. The coordinate components of the regular grid query points are multiplied by a coordinate scaling factor, which is the reciprocal of the sum of the compression value and a set smoothing factor, to generate a non-uniform BEV query point set. In this invention, the BEV plane covers a range of -10 meters to 10 meters in the longitudinal direction and a range of -5 meters to 5 meters in the transverse direction. The BEV plane is uniformly sampled with a spatial resolution of 0.5 meters, thereby generating multiple regularly arranged query points within the BEV plane. Each query point corresponds to a two-dimensional coordinate position, forming a set of regular grid query points. Non-uniform processing is performed on the set of query points in the regular grid. The coordinates of the query points are remapped based on the compression values ​​corresponding to the left, right, front, and rear directions in the boundary compression vector. For each direction, the compression value corresponding to that direction is obtained, and the compression value is added to a smoothing factor. The smoothing factor is used to avoid the coordinate scaling coefficient being abnormal due to the compression value being too small, and the value is set to 0.05. The reciprocal of the sum of the compression value and the smoothing factor is taken to obtain the coordinate scaling coefficient corresponding to that direction. For each query point in the set of query points in the regular grid, its corresponding direction component is determined according to its position in the vehicle's own coordinate system, and the coordinate component corresponding to that direction is multiplied by the coordinate scaling coefficient to perform a linear scaling transformation on the query point coordinates.

[0039] Through the above coordinate remapping process, the regular grid query points are distributed in a non-uniform direction with a correlation in the BEV plane, resulting in a non-uniform set of BEV query points. The compression-guided cross-view aggregation unit takes the non-uniform BEV query point set as input and performs cross-view feature aggregation processing on multi-view image features to generate intermediate BEV features. The compression morphology temporal memory unit is used to perform temporal correlation modeling of the BEV intermediate features at continuous time intervals, store the BEV intermediate features corresponding to historical time intervals in chronological order, and associate and store the boundary compression vector of the corresponding time interval for the BEV intermediate features of each historical time interval. Perform element-wise difference operations between the current boundary compression vector and each historical boundary compression vector stored at the current time, and calculate the difference distance between the current boundary compression vector and each historical boundary compression vector. The differential distance is calculated as follows: subtract the corresponding elements of the current boundary compression vector from a certain historical boundary compression vector one by one, take the absolute value, and sum all the absolute values. Select at least one historical moment corresponding to the minimum difference distance from the difference distances, and use the BEV intermediate feature corresponding to the historical moment as the historical feature that matches the current compression pattern; The matched historical features are concatenated with the intermediate BEV features at the current moment along the feature dimension to form time-enhanced BEV features. The final BEV feature output unit performs a linear transformation on the time-enhanced BEV features along the channel dimension and performs channel-by-channel normalization processing to output the final BEV features. The improved BEVFormer model in this invention retains the basic framework of existing BEVFormer models that construct BEV representations based on multi-view image features. It introduces geometric compression information highly relevant to the automatic parking scenario, structurally improving the BEV query mechanism, cross-view feature aggregation method, and temporal correlation modeling method. Unlike existing BEVFormer models that use fixed regular grids or learnable parameters to generate BEV query points, this invention introduces a boundary compression vector. Based on the spatial relationship between the vehicle's physical contour and the parking space boundary, it dynamically reconstructs the spatial distribution of BEV query points, resulting in a directionally correlated non-uniform distribution of BEV query points within the BEV plane. This allows the model to achieve higher spatial sampling resolution in confined spaces during parking.

[0040] Furthermore, this invention introduces the compression state encoding result as modulation information during the cross-view feature aggregation process, directly linking the aggregation process of multi-view image features to the current geometric compression state of the vehicle. This avoids the problem in existing BEVFormer where cross-view feature aggregation relies solely on spatial projection relationships while ignoring parking constraint information. Further, this invention manages and retrieves historical BEV intermediate features through a compression morphology temporal memory unit. Instead of relying solely on temporal adjacency for temporal modeling, it selects historical features based on the similarity of boundary compression vectors to participate in current feature construction. This enables the model to maintain continuous perception of similar geometric states during parking in narrow spaces and when multiple attitude adjustments are required.

[0041] Through the above structural improvements, the improved BEVFormer model of this invention can more accurately reflect the spatial relationship between the vehicle and the parking space boundary in the automatic parking scenario, improve the ability of BEV features to express local restricted areas, and thus provide a more stable and consistent perception basis for the generation of vehicle control signals.

[0042] In this embodiment, the step of performing cross-view feature aggregation processing on multi-view image features to generate BEV intermediate features specifically involves: Based on the two-dimensional coordinate position of each non-uniform BEV query point in the vehicle's own coordinate system, the non-uniform BEV query point is mapped to the image plane of each viewpoint to obtain the corresponding projected coordinates of the non-uniform BEV query point in each viewpoint image. In each viewpoint image, local feature vectors of a set size are sampled from the image feature map of the corresponding viewpoint, with the projection coordinates as the center. The local feature vectors from different viewpoints are then spliced ​​together in viewpoint order to form a cross-viewpoint feature sequence corresponding to the non-uniform BEV query point. Specifically, the size of the local sampling window is set to 7×7 pixels, that is, with the projection coordinates as the center, it extends 3 pixels up, down, left, and right. Within the local sampling window, feature vectors at each pixel position are read sequentially from the image feature map of the corresponding viewpoint. An attention calculation operation is performed on the cross-view feature sequence, the compression embedding vector is converted into an attention modulation vector through a linear mapping, and the attention modulation vector is multiplied element-wise with the cross-view feature sequence to perform weight modulation on the local feature vectors of each view in the cross-view feature sequence. The weighted cross-view feature sequence is summed to obtain the BEV spatial features corresponding to the non-uniform BEV query point. Based on the spatial location of the non-uniform BEV query point in the BEV plane, all BEV spatial features are arranged to form BEV intermediate features. In the compression-guided cross-view feature aggregation process, the compression embedding vector is used to modulate the attention allocation of the cross-view feature sequence. Specifically, the compression embedding vector is a one-dimensional vector whose elements correspond to the compression state information in the left, right, front, and rear directions of the vehicle. The compression embedding vector is input to a linear mapping layer, and it is mapped to an attention modulation vector with the same number of channels as the cross-view feature sequence through matrix multiplication. The dimension of the weight matrix of the linear mapping layer matches the feature dimension of the cross-view feature sequence. Subsequently, the attention modulation vector is multiplied element-wise with the local feature vectors of each view in the cross-view feature sequence, so that the features of different views are subject to the numerical constraints of the corresponding compression state in the channel dimension.

[0043] By using the above method, the compression embedding vector is directly involved in the attention calculation process of cross-view features, so that the multi-view feature fusion at a specific query location in the BEV space is consistent with the current geometric compression state of the vehicle, providing a more reliable perception basis for control decisions.

[0044] In this embodiment, step six specifically includes: The final BEV features are input into the vehicle control generation network, which includes a feature mapping layer and a control output layer. The feature mapping layer performs a linear transformation on the final BEV feature along the channel dimension, and applies a ReLU function to each element of the linearly transformed result for nonlinear mapping to obtain the control feature vector. The control feature vector is input to the control output layer, which includes two linear mapping units that generate steering control and displacement control quantities respectively through linear mapping. Specifically, the control feature vector output by the feature mapping layer is input to the control output layer, which consists of two independent linear mapping units. One linear mapping unit generates the vehicle's steering control quantity, and the other linear mapping unit generates the vehicle's displacement control quantity. Each linear mapping unit performs a linear transformation on the control feature vector through matrix multiplication and introduces a corresponding bias term during the transformation to obtain the corresponding scalar control output. The steering control quantity is used to characterize the steering angle change command of the vehicle's front wheels, and the displacement control quantity is used to characterize the vehicle's longitudinal movement command within the current control cycle. By using independent linear mapping units, the lateral control and longitudinal control are structurally decoupled, thereby ensuring the stability and continuity of the vehicle's attitude adjustment during automatic parking. The steering control quantity and the displacement control quantity are combined as a vehicle control signal output to control the change of the vehicle's steering angle and longitudinal movement, thereby adjusting the vehicle's attitude.

[0045] In this embodiment, step eight specifically includes: In parking control mode, the vehicle's attitude parameters are continuously acquired over multiple control cycles, including the vehicle's heading angle and the vehicle's center of gravity position. Perform differential calculations on the vehicle heading angle and the vehicle center of gravity position within adjacent control cycles to obtain the corresponding changes in heading angle and center of gravity position. When the change in heading angle is less than the set heading angle change stabilization threshold for a set number of consecutive control cycles, and the change in position is less than the set center of mass position stabilization threshold for a set number of consecutive control cycles, the vehicle attitude is determined to be stable, a parking completion command is output, and the automatic parking control process ends. In this invention, the heading angle change stabilization threshold is set to 0.5 degrees, the center of gravity position stabilization threshold is set to 0.05 meters, and the number of consecutive control cycles is set to 5 control cycles. When the heading angle change and the center of gravity position change are both less than the corresponding stabilization threshold within 5 consecutive control cycles, the vehicle attitude is determined to be in a stable state, and a parking completion command is output, thereby ending the automatic parking control process and achieving reliable determination of the parking termination state.

[0046] refer to Figure 2 An automatic parking control system based on image segmentation includes the following modules: The image acquisition and segmentation module is used to acquire multi-view image data of the environment around the vehicle, and to perform image segmentation processing on the multi-view image data to extract the parking space boundary lines. The compression calculation module is used to calculate the compression values ​​between the vehicle's physical outline and the parking space boundary line in the left, right, front, and rear directions, and to construct the boundary compression vector. The compression trend analysis module is used to continuously update the boundary compression vector over time, calculate the change in compression value in each direction, and generate a compression trend vector. The control state determination module is used to determine the current control state based on the pressure trend vector. When the pressure trend in any direction meets the preset trigger condition, the parking control state is interrupted and the pressure response control state is switched. The BEV feature generation module is used to input the current boundary compression vector and the corresponding compression trend vector into the improved BEVFormer model under the compression response control state. It generates a non-uniform BEV query point set through the boundary compression vector, performs temporal correlation modeling through the compression morphology temporal memory unit, and outputs the final BEV features. A vehicle control signal generation module is used to generate vehicle control signals based on the final BEV characteristics. The vehicle control signals include steering control quantities and displacement control quantities, which are used to adjust the vehicle attitude. The parking control execution module is used to control the vehicle to perform steering and longitudinal movement according to the vehicle control signal. After completing a vehicle posture adjustment, if the current pressure trend in all directions does not meet the preset trigger condition, it switches back to the parking control state; otherwise, it continues to execute the pressure response control state. The parking termination determination module is used to continuously acquire the vehicle's attitude parameters within multiple control cycles during parking control, determine whether the vehicle's attitude is in a stable state, and output a parking completion command when the vehicle's attitude is determined to be in a stable state, thereby ending the automatic parking control process. Example

[0047] To verify the feasibility of this invention in practice, it was deployed on a test vehicle equipped with a 360-degree surround-view camera system and an autonomous driving control platform. The experiment simulated a side parking scenario in a densely populated urban area. The parking space width was set between 2.2m and 2.5m, and the length was about 5.0m. The distance between the test vehicle and the parking space was within 0.5m. The environment had partial obstruction and structured light interference. This scenario has typical difficulties such as small compression space, rapid changes in compression trend, and low perception redundancy. It is one of the key scenarios for evaluating the stability and intelligent response capability of the parking system.

[0048] During the experiment, the system first acquires multi-view image data from four directions using an onboard camera. The image segmentation network extracts the parking space boundary lines in real time and calculates the four-way boundary pressure values ​​based on the vehicle's physical contour, constructing a boundary pressure vector. In continuous time frames, a pressure trend vector is further generated to monitor the rate of change of the pressure values, thereby determining whether to enter the pressure response control state. In this state, the vehicle control system inputs the boundary pressure vector and the pressure trend vector into the improved BEVFormer model, dynamically generating a non-uniform BEV query point set. It aggregates multi-view image features and introduces a pressure morphology temporal memory unit to improve feature stability under complex scenarios, generating the final BEV features.

[0049] To test the response of attitude adjustment during parking tasks, the system inputs the final BEV features into the vehicle control generation network, extracts steering control and displacement control variables, and obtains vehicle control signals. These vehicle control signals are applied to the vehicle steering and drive units in real time through the controller, enabling the vehicle to smoothly complete the parking operation in narrow parking spaces. When the system detects that the vehicle attitude remains stable for five consecutive control cycles (approximately 200ms per cycle), it outputs a parking completion command and terminates the control process.

[0050] To compare the performance differences between the present invention and the traditional uniform BEV sampling BEVFormer model scheme, multiple rounds of test data under the same scenario conditions were recorded, as shown in Table 1 below.

[0051] Table 1 Comparison of attitude adjustment time under pressure response control state Based on the test data shown in Table 1 above, the performance of the traditional BEVFormer model and the improved BEVFormer model in terms of vehicle attitude adjustment time was compared and analyzed. The results show that the improved BEVFormer model has a significant advantage in control efficiency.

[0052] In the 10 experiments from test numbers T01 to T10, the average attitude adjustment time of the traditional BEVFormer model was approximately 8.02 seconds, while the average adjustment time of the improved BEVFormer model was 5.38 seconds, a reduction of nearly 2.64 seconds, representing an average improvement of about 33%. The largest improvements were observed in test scenarios T07 and T09, reaching 34.6% and 34.5% respectively. This indicates that in complex or high-pressure sensing environments, the improved BEVFormer model responds more quickly to attitude adjustments, exhibiting stronger adaptability and stability.

[0053] Furthermore, in all test numbers, the adjustment time of the improved BEVFormer model was significantly shorter than that of the traditional model, with less data fluctuation, indicating that the method not only improves the adjustment speed but also maintains good stability and consistency. This performance improvement is mainly due to the synergistic effect of the non-uniform BEV query point generation mechanism, the pressure-guided cross-view aggregation process, and the pressure morphology temporal memory unit introduced in this invention, which more accurately perceives boundary pressure information and quickly generates highly adaptive control signals. Overall, the improved BEVFormer model effectively solves the problems of lag and unstable control in traditional models in confined parking spaces, providing stronger guarantees for the real-time performance and safety of automatic parking.

[0054] In this embodiment, by introducing pressure perception analysis and improving the BEVFormer model, a more refined understanding and control response of confined spaces is achieved during automatic parking. Through the introduction of a pressure state encoding unit and a pressure trend vector, the system can dynamically perceive changes in the relative relationship between the vehicle and the parking boundary, forming a non-uniform BEV query point distribution and enhancing attention to the boundary environment. Simultaneously, the pressure-guided cross-view aggregation unit can dynamically adjust feature fusion weights based on pressure information from different directions, making the model more focused on perceptual information of potential collision directions and improving the discriminative ability of feature extraction. Furthermore, the pressure morphology temporal memory unit effectively utilizes historical parking states, enhancing the model's ability to remember the environment during dynamic parking and improving the consistency of attitude prediction. Finally, the vehicle control generation network combines the final BEV features to output vehicle control signals, ensuring that the vehicle completes parking smoothly and safely in complex scenarios. This achieves more adaptive perception-decision integrated control, possessing advantages such as clear structure, high precision, and strong generalization.

[0055] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An automatic parking control method based on image segmentation, characterized in that, Includes the following steps: Step 1: Collect multi-view image data of the vehicle's surrounding environment, perform image segmentation on the multi-view image data, and extract the parking space boundary lines; Step 2: Calculate the compression values ​​between the vehicle's physical outline and the parking space boundary line in the left, right, front, and rear directions, and construct the boundary compression vector; Step 3: Continuously update the boundary compression vector over time, calculate the change in compression value in each direction, and generate a compression trend vector; Step 4: Determine the current control state based on the pressure trend vector. When the pressure trend in any direction meets the preset trigger condition, interrupt the parking control state and switch to the pressure response control state. Step 5: Under the compression response control state, input the current boundary compression vector and the corresponding compression trend vector into the improved BEVFormer model, generate a non-uniform BEV query point set through the boundary compression vector, and perform temporal correlation modeling through the compression morphology temporal memory unit to output the final BEV features; Step 6: Generate vehicle control signals based on the final BEV characteristics to adjust the vehicle attitude; Step 7: After completing the vehicle attitude adjustment, if the current pressure trend in all directions does not meet the preset trigger conditions, switch back to parking control state; otherwise, continue to execute pressure response control state. Step 8: In parking control mode, when the vehicle's attitude is stable, the automatic parking control process ends.

2. The automatic parking control method based on image segmentation according to claim 1, characterized in that, Step one specifically involves: By using vehicle-mounted cameras installed at the front, rear, left, and right sides of the vehicle, multi-view image data covering the surrounding environment of the vehicle is collected. The multi-view image data includes forward view images, rear view images, and left and right side view images. The multi-view image data is uniformly timestamped during collection to form a synchronized image sequence. The multi-view image data is input into an image segmentation network, which includes an encoder and a decoder; The encoder divides the input multi-view image into image blocks of a set size, performs linear embedding processing on each image block, and encodes the image blocks layer by layer in a multi-layer Transformer structure to extract multi-scale semantic features corresponding to different spatial resolutions. The multi-scale semantic features are input into the decoder, which aligns and fuses the semantic features at different scales and outputs pixel-level classification results. Extract the set of pixels labeled as parking line category from the pixel-level classification results, and perform connected component filtering and edge fitting on the pixel set to obtain the parking space boundary line.

3. The automatic parking control method based on image segmentation according to claim 1, characterized in that, Step two specifically involves: Establish a vehicle coordinate system with the vehicle's geometric center as the origin, the vehicle's forward direction as the vertical axis, and the direction perpendicular to the vertical axis as the horizontal axis. The vehicle physical contour line is determined in the vehicle's own coordinate system, and the vehicle physical contour line is used to characterize the outer contour range of the vehicle in the vehicle's own coordinate system. The parking space boundary line is transformed from the image coordinate system to the vehicle's own coordinate system, and the parking space boundary line is represented as a set of boundary points composed of multiple boundary pixels; In the vehicle's own coordinate system, with the vehicle's physical outline as the reference, the Euclidean distance between each boundary pixel in the set of boundary points and the vehicle's physical outline is calculated in the left, right, front, and rear directions, based on the positional relationship of the boundary pixel points relative to the horizontal and vertical axes in the vehicle's own coordinate system. In the left, right, front, and rear directions, the minimum distance value is selected from all the calculated Euclidean distances as the compression value for the corresponding direction, and they are combined in order to construct a boundary compression vector including the left compression value, right compression value, front compression value, and rear compression value.

4. The automatic parking control method based on image segmentation according to claim 1, characterized in that, Step three specifically involves: By continuously sampling the boundary compression vector at set time intervals, a sequence of boundary compression vectors arranged in chronological order is obtained; For the left, right, front, and rear directions, the compression values ​​corresponding to two adjacent sampling times are extracted from the boundary compression vector sequence, respectively. Perform a differential operation on the pressure values ​​corresponding to two adjacent sampling times in each direction to obtain the pressure change in the corresponding direction at the current sampling time; The compression changes obtained from the left, right, front, and rear directions are combined in sequence to generate a compression trend vector.

5. The automatic parking control method based on image segmentation according to claim 1, characterized in that, Step four specifically involves: For the compression trend vector, the compression change amounts corresponding to the left, right, front, and rear directions are obtained respectively; A change threshold is set, which is used to limit the allowable range of pressure change between two adjacent sampling times; In any direction, when the absolute value of the corresponding pressure change at the current sampling time is greater than the change threshold, it is determined that the preset triggering condition is met; When at least one direction meets the preset triggering condition, the current parking control state is interrupted and switched to the pressure response control state.

6. The automatic parking control method based on image segmentation according to claim 1, characterized in that, The improved BEVFormer model includes a compression state encoding unit, a non-uniform BEV query point generation unit, a compression-guided cross-view aggregation unit, a compression morphology temporal memory unit, and a final BEV feature output unit. The compression state encoding unit is used to receive the boundary compression vector and the compression trend vector, and pair the compression value corresponding to each direction with the compression change amount in the order of left, right, front and back, and splice them together in sequence to form a compression state sequence. Based on the sequential position of each element in the compression state sequence, a position index is constructed, and the position index is converted into a position encoding vector through a linear mapping method. The position encoding vector is then added element by element to the compression state sequence to obtain a compression embedding vector. The non-uniform BEV query point generation unit generates a set of regular grid query points in the BEV plane, with the vehicle's own coordinate system as a reference. The set of regular grid query points is uniformly sampled at a set interval in both the longitudinal and lateral directions. Based on the compression values ​​corresponding to the left, right, front, and rear directions in the boundary compression vector, the set of regular grid query points is subjected to coordinate remapping processing. The coordinate components of the regular grid query points are multiplied by the coordinate scaling factor, which is the reciprocal of the sum of the compression value and the set smoothing factor, to generate a non-uniform BEV query point set. The compression-guided cross-view aggregation unit takes the non-uniform BEV query point set as input and performs cross-view feature aggregation processing on multi-view image features to generate intermediate BEV features. The compression morphology temporal memory unit is used to perform temporal correlation modeling of the BEV intermediate features at continuous time intervals, store the BEV intermediate features corresponding to historical time intervals in chronological order, and associate and store the boundary compression vector of the corresponding time interval for the BEV intermediate features of each historical time interval. Perform element-wise difference operations between the current boundary compression vector and each historical boundary compression vector stored at the current time, and calculate the difference distance between the current boundary compression vector and each historical boundary compression vector. Select at least one historical moment corresponding to the minimum difference distance from the difference distances, and use the BEV intermediate feature corresponding to the historical moment as the historical feature that matches the current compression pattern; The matched historical features are concatenated with the intermediate BEV features at the current moment along the feature dimension to form time-enhanced BEV features. The final BEV feature output unit performs a linear transformation on the time-enhanced BEV features along the channel dimension and performs channel-by-channel normalization processing to output the final BEV features.

7. The automatic parking control method based on image segmentation according to claim 6, characterized in that, The process of performing cross-view feature aggregation on multi-view image features to generate BEV intermediate features is as follows: Based on the two-dimensional coordinate position of each non-uniform BEV query point in the vehicle's own coordinate system, the non-uniform BEV query point is mapped to the image plane of each viewpoint to obtain the corresponding projected coordinates of the non-uniform BEV query point in each viewpoint image. In each viewpoint image, local feature vectors of a set size are sampled from the image feature map of the corresponding viewpoint, with the projection coordinates as the center, and the local feature vectors from different viewpoints are spliced ​​together in viewpoint order to form a cross-viewpoint feature sequence corresponding to the non-uniform BEV query point. An attention calculation operation is performed on the cross-view feature sequence, the compression embedding vector is converted into an attention modulation vector through a linear mapping, and the attention modulation vector is multiplied element-wise with the cross-view feature sequence to perform weight modulation on the local feature vectors of each view in the cross-view feature sequence. The weighted cross-view feature sequence is summed to obtain the BEV spatial features corresponding to the non-uniform BEV query point. Based on the spatial location of the non-uniform BEV query point within the BEV plane, all BEV spatial features are arranged to form intermediate BEV features.

8. The automatic parking control method based on image segmentation according to claim 1, characterized in that, Step six specifically involves: The final BEV features are input into the vehicle control generation network, which includes a feature mapping layer and a control output layer. The feature mapping layer performs a linear transformation on the final BEV feature along the channel dimension, and applies a ReLU function to each element of the linearly transformed result for nonlinear mapping to obtain the control feature vector. The control feature vector is input to the control output layer, which includes two linear mapping units that generate steering control and displacement control quantities respectively through linear mapping. The steering control quantity and the displacement control quantity are combined as a vehicle control signal output to control the change of the vehicle's steering angle and longitudinal movement, thereby adjusting the vehicle's attitude.

9. The automatic parking control method based on image segmentation according to claim 1, characterized in that, Step eight specifically involves: In parking control mode, the vehicle's attitude parameters are continuously acquired over multiple control cycles, including the vehicle's heading angle and the vehicle's center of gravity position. Perform differential calculations on the vehicle heading angle and the vehicle center of gravity position within adjacent control cycles to obtain the corresponding changes in heading angle and center of gravity position. When the change in heading angle is less than the set heading angle change stabilization threshold for a set number of consecutive control cycles, and the change in position is less than the set center of mass position stabilization threshold for a set number of consecutive control cycles, the vehicle attitude is determined to be stable, a parking completion command is output, and the automatic parking control process ends.

10. An automatic parking control system based on image segmentation, comprising executing the automatic parking control method based on image segmentation as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The image acquisition and segmentation module is used to acquire multi-view image data of the environment around the vehicle, and to perform image segmentation processing on the multi-view image data to extract the parking space boundary lines. The compression calculation module is used to calculate the compression values ​​between the vehicle's physical outline and the parking space boundary line in the left, right, front, and rear directions, and to construct the boundary compression vector. The compression trend analysis module is used to continuously update the boundary compression vector over time, calculate the change in compression value in each direction, and generate a compression trend vector. The control state determination module is used to determine the current control state based on the pressure trend vector. When the pressure trend in any direction meets the preset trigger condition, the parking control state is interrupted and the pressure response control state is switched. The BEV feature generation module is used to input the current boundary compression vector and the corresponding compression trend vector into the improved BEVFormer model under the compression response control state. It generates a non-uniform BEV query point set through the boundary compression vector, performs temporal correlation modeling through the compression morphology temporal memory unit, and outputs the final BEV features. A vehicle control signal generation module is used to generate vehicle control signals based on the final BEV characteristics. The vehicle control signals include steering control quantities and displacement control quantities, which are used to adjust the vehicle attitude. The parking control execution module is used to control the vehicle to perform steering and longitudinal movement according to the vehicle control signal. After completing a vehicle posture adjustment, if the current pressure trend in all directions does not meet the preset trigger condition, it switches back to the parking control state; otherwise, it continues to execute the pressure response control state. The parking termination determination module is used to continuously acquire the vehicle's attitude parameters within multiple control cycles during parking control, determine whether the vehicle's attitude is in a stable state, and output a parking completion command when the vehicle's attitude is determined to be in a stable state, thereby ending the automatic parking control process.

Images