WO2022222036A1 - 车位确定方法及装置 - Google Patents

车位确定方法及装置 Download PDF

Info

Publication number
WO2022222036A1
WO2022222036A1 PCT/CN2021/088440 CN2021088440W WO2022222036A1 WO 2022222036 A1 WO2022222036 A1 WO 2022222036A1 CN 2021088440 W CN2021088440 W CN 2021088440W WO 2022222036 A1 WO2022222036 A1 WO 2022222036A1
Authority
WO
WIPO (PCT)
Prior art keywords
parking space
instance
area
target
parking
Prior art date
Application number
PCT/CN2021/088440
Other languages
English (en)
French (fr)
Inventor
裘索
陈超
徐吉睿
陈晓智
王谢兵
Original Assignee
深圳市大疆创新科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳市大疆创新科技有限公司 filed Critical 深圳市大疆创新科技有限公司
Priority to PCT/CN2021/088440 priority Critical patent/WO2022222036A1/zh
Publication of WO2022222036A1 publication Critical patent/WO2022222036A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition

Definitions

  • the present application relates to the field of computer technology, and in particular, to a parking space determination method, device, and computer-readable storage medium.
  • parking space detection is a key link.
  • the parking system can perceive the specific location of the surrounding parking spaces through the parking space detection function. It can be seen that the parking space detection function is the basis for realizing automatic parking.
  • the parking space detection first needs to extract the local features of the corner points in the area near the corner points of the parking space, and identify the position of the corner points of the parking space through the analysis of the local features of the corner points, and finally identify the specific position of the parking space according to the position of the corner points of the parking space. .
  • the present application provides a method, device, camera module and movable device for determining a parking space, which can solve the problem in the prior art that when the corners of the parking spaces are blocked, the local features of the corners cannot be extracted, resulting in the failure of the recognition of the parking spaces.
  • an embodiment of the present application provides a method for determining a parking space, including:
  • each group of said candidate sets includes the positions of 4 parking space corners;
  • a target candidate set is determined from the N groups of candidate sets for each parking space instance, and a target parking space corresponding to the parking space instance is determined according to the target candidate set.
  • an embodiment of the present application provides a device for determining a parking space, including: an acquisition module and a processor;
  • the obtaining module is used for: obtaining a bird's-eye view of the parking area, and extracting image features of the bird's-eye view of the parking area;
  • the processing module is used for: inputting the image features into a preset prediction model to obtain an image segmentation result including a parking space instance and a background instance, and the difference between the position point in the image segmentation result and the parking space corner point. offset information;
  • each group of said candidate sets includes the positions of 4 parking space corners;
  • a target candidate set is determined from the N groups of candidate sets for each parking space instance, and a target parking space corresponding to the parking space instance is determined according to the target candidate set.
  • the present application provides a computer-readable storage medium, the computer-readable storage medium comprising instructions that, when executed on a computer, cause the computer to perform the method described in the above aspects.
  • the present application provides a computer program product comprising instructions which, when run on a computer, cause the computer to perform the method described in the above aspects.
  • the present application is processed based on the global features of the parking spaces, which is different from the tedious extraction and processing of local features near the corners of the parking spaces in the related art, so the amount of calculation is reduced, and the four The corners of the parking spaces are calculated according to the offset, so even if the corners of the parking spaces are blocked in the actual environment, the corresponding offset information will still be obtained, so that the blocking of the corners of the parking spaces will not affect the calculation of the corners of the parking spaces. , which improves the success rate of parking space recognition.
  • FIG. 1 is an architecture diagram of a parking space determination system provided by an embodiment of the present application
  • FIG. 2 is a flowchart of a method for determining a parking space provided by an embodiment of the present application
  • FIG. 3 is a specific flowchart of a method for determining a parking space provided by an embodiment of the present application
  • FIG. 4 is a block diagram of an apparatus for determining a parking space provided by an embodiment of the present application.
  • the parking space determination system includes: a feature extraction module, a prediction module and a result output module, wherein, the feature extraction module can use the parking area bird's eye view collected in the parking scene 10 is the input, and outputs the image features of the bird's-eye view of the parking area.
  • the image segmentation operation can be performed based on the image features, and three result branches are obtained: the instance result branch, the corner offset branch, and the corner offset branch. Point attribute branch.
  • the instance result branch includes the image segmentation result 20 obtained based on the image feature extraction of the bird's-eye view of the parking area 10, and the image segmentation result 20 includes background examples, various types of parking space instances (such as parking spaces, non-parking spaces, etc.); image segmentation results 20 only includes each object example obtained by image segmentation, and for the parking space instances, the positions and attributes of specific parking space corners have not been assigned.
  • the corner offset branch contains the offset information between the position point in the image segmentation result 20 and the corner point of the parking space; for example, since a parking space consists of 4 parking space corner points, for a position point p, the angle
  • the point offset branch has the offset ⁇ p1 of the position point p relative to the corner point of the upper left parking space; the offset amount ⁇ p2 of the position point p relative to the corner point of the upper right parking space; the position point p relative to the corner point of the lower left parking space.
  • the offset ⁇ p3; the offset ⁇ p4 of the position point p relative to the corner point of the lower right parking space is the offset information between the position point in the image segmentation result 20 and the corner point of the parking space.
  • the corner attribute branch contains the specific attributes of each parking space corner, such as visible, invisible, parking space entrance corner, confidence, etc.
  • the area where the parking space instance is located may be extracted from the image segmentation result 20, and the distance between N position points in the area where each parking space instance is located and the parking space corner points may be extracted from the corner offset branch. Then, N groups of candidate sets corresponding to each parking space instance are established, and each group of candidate sets includes the positions of 4 parking space corner points; that is, for each parking space instance, N candidate parking space areas are established, and further, A target candidate set may be determined from the N groups of candidate sets of each parking space instance based on a strategy, and the parking space area established by the target candidate set may be used as the target parking space corresponding to the parking space instance.
  • the attribute information of the four parking space corner points in the target candidate set can be extracted from the corner attribute branch, and the parking space type corresponding to the parking space instance can be extracted from the instance result branch.
  • the attribute type and parking space type of the parking space corner point can be determined.
  • the solution provided by the present application is processed based on the global features of the parking space, which is different from the tedious extraction and processing of local features near the corners of the parking space in the related art, so the amount of calculation is reduced, and the parking space
  • the four corners of the parking spaces are calculated according to the offsets, so even if the corners of the parking spaces are blocked in the actual environment, the corresponding offset information will still be obtained, so that the blocked corners of the parking spaces will not affect the corners of the parking spaces.
  • the calculation has an impact and improves the success rate of parking space identification.
  • FIG. 2 is a flowchart of a method for determining a parking space provided by an embodiment of the present application. As shown in FIG. 2 , the method may include:
  • Step 101 Obtain a bird's-eye view of the parking area, and extract image features of the bird's-eye view of the parking area.
  • the bird's-eye view of the parking area may be collected by a multi-camera module disposed on the top of the vehicle, and the multi-camera module includes multiple cameras.
  • the multi-camera module is used for shooting to collect a bird's-eye view of the parking area in the area where the vehicle is located.
  • the data dimension of the bird's-eye view of the parking area is high, it is difficult to process it directly, so the image features of the bird's-eye view of the parking area can be extracted, thereby reducing the data dimension, and the difficulty of subsequent processing of the image features is low.
  • features are the corresponding characteristics or characteristics of a certain type of objects that are different from other types of objects, or a collection of these characteristics and characteristics.
  • Features are data that can be extracted through measurement or processing.
  • the representation of the corresponding features or characteristics of the aerial view of the area, and its main idea is to project the image visual characteristics or patterns of the aerial view of the parking area into a feature space, and get the best reflection of the nature of the aerial view of the parking area or to distinguish the aerial view of the parking area. image features.
  • Common image feature extraction methods include: (1) Geometric method feature extraction. The geometric method is to establish A texture feature analysis method based on image texture primitive theory. (2) Model method feature extraction. The model method is based on the structural model of the image, and uses the parameters of the model as texture features, such as a convolutional neural network model. (3) Feature extraction by signal processing. The extraction and matching of texture features mainly include: gray level co-occurrence matrix, autoregressive texture model, wavelet transform and so on.
  • Step 102 Input the image features into a preset prediction model to obtain an image segmentation result including a parking space instance and a background instance, and the offset information between the position point in the image segmentation result and the parking space corner point .
  • the prediction model can perform image segmentation and parking space corner information prediction based on image features, thereby outputting an image segmentation result including a parking space instance and a background instance, as well as the position point in the image segmentation result and the parking space corner point. offset information between.
  • the output result of the prediction model includes three branches, the instance result branch includes an image segmentation result 20 obtained based on the image feature extraction of the bird's-eye view of the parking area 10, and the image segmentation result 20 includes background examples, various types of parking space instances (such as parking spaces, non-parking spaces, etc.); the image segmentation result 20 only includes each object instance obtained by image segmentation, and for the parking space instances, the positions and attributes of specific parking spaces corners have not been assigned.
  • the instance result branch includes an image segmentation result 20 obtained based on the image feature extraction of the bird's-eye view of the parking area 10
  • the image segmentation result 20 includes background examples, various types of parking space instances (such as parking spaces, non-parking spaces, etc.); the image segmentation result 20 only includes each object instance obtained by image segmentation, and for the parking space instances, the positions and attributes of specific parking spaces corners have not been assigned.
  • the corner offset branch contains the offset information between the position point in the image segmentation result 20 and the corner point of the parking space; for example, since a parking space consists of 4 parking space corner points, for a position point p, the angle
  • the point offset branch has the offset ⁇ p1 of the position point p relative to the corner point of the upper left parking space; the offset amount ⁇ p2 of the position point p relative to the corner point of the upper right parking space; the position point p relative to the corner point of the lower left parking space.
  • the offset ⁇ p3; the offset ⁇ p4 of the position point p relative to the corner point of the lower right parking space is the offset information between the position point in the image segmentation result 20 and the corner point of the parking space.
  • the image segmentation result is an output result in a 3 ⁇ H ⁇ W dimension, where H ⁇ W is the resolution of the output result, and H ⁇ W can be the same as the resolution of the bird’s-eye view of the parking area, or it can be the parking area.
  • the resolution size of the bird's-eye view is 1/2, 1/4, 1/8, etc., which is not limited in this embodiment of the present application.
  • the dimension of the corner offset branch is 8 ⁇ H ⁇ W.
  • the 8 channels refer to: the offset from a position point p inside the parking space to the four corner points of the parking space, and the component sizes on the x-axis and the y-axis respectively.
  • Step 103 Extract the area where the parking space instance is located from the image segmentation result, and establish each parking space instance according to the offset information between the N position points and the parking space corner points in the area where each parking space instance is located.
  • N groups of candidate sets corresponding to the parking space instances, each group of candidate sets includes the positions of four parking space corners.
  • the area where the parking space instance is located can be extracted from the image segmentation result 20 , and N position points and parking spaces in the area where each parking space instance is located can be extracted from the corner offset branch.
  • the offset information between the corner points, and then N groups of candidate sets corresponding to each parking space instance are established, and each group of candidate sets includes the positions of 4 parking space corner points; that is, for each parking space instance, N candidate parking space areas are established .
  • the correspondence between the position point 1 and the four offsets in the area there is a correspondence between the position point 1 and the four offsets in the area, the correspondence between the position point 2 and the four offsets....
  • the position point N and the four offsets The correspondence between the offsets, each position point and its corresponding four offsets can obtain a group candidate set, in the process of establishing each group candidate set, according to the coordinates of each position point and the four offsets According to the size of the shift amount, the position coordinates of the four corner points (upper left, lower left, upper right, and lower right) of the four parking spaces can be obtained, that is, a set of candidate sets can be obtained.
  • Step 104 Determine a target candidate set from the N groups of candidate sets for each parking space instance, and determine a target parking space corresponding to the parking space instance according to the target candidate set.
  • a target candidate set may be determined from the N groups of candidate sets of each parking space instance based on a strategy, and the target candidate set established by the target candidate set may be determined.
  • the parking space area as the target parking space corresponding to the parking space instance.
  • the strategy may specifically include: weighted average strategy, voting strategy, non-maximum value suppression (NMS,
  • Non-maximum Suppression algorithm strategy, etc.
  • the specific selection of the strategy is not limited in this embodiment of the present application.
  • the method for determining a parking space provided by the embodiment of the present application is processed based on the global features of the parking space, which is different from the tedious extraction and processing of local features near the corners of the parking space in the related art, thus reducing the amount of calculation.
  • the four corner points of the parking space are calculated according to the offset, so even if the corner points of the parking space are blocked in the actual environment, the corresponding offset information will still be obtained, so that the blocking of the corner points of the parking space will not affect the parking space.
  • the calculation of the corner points has an impact and improves the success rate of parking space recognition.
  • FIG. 3 is a specific flowchart of a method for determining a parking space provided by an embodiment of the present application. As shown in FIG. 3 , the method may include:
  • Step 201 Obtain a bird's-eye view of the parking area, and extract image features of the bird's-eye view of the parking area.
  • step 201 For details of step 201, reference may be made to the above-mentioned 101, which will not be repeated here.
  • Step 202 Input the image features into a preset prediction model to obtain an image segmentation result including a parking space instance and a background instance, and the offset information between the position point in the image segmentation result and the parking space corner point .
  • step 202 For details of step 202, reference may be made to the above 102, which will not be repeated here.
  • Step 203 Distinguish parking space instances and background instances in the image segmentation result.
  • step 203 may specifically include:
  • Sub-step 2031 normalize the image segmentation result, determine the class probability value of each position point in the image segmentation result, so as to obtain a probability map; the class probability value includes the probability value of the corresponding parking space class, The probability value corresponding to the background class.
  • the image segmentation result may be normalized first, and the class probability value of each location point in the image segmentation result is determined, that is, the probability of each location point belonging to the parking space class and the probability of the background class to which each location point belongs. probability. Based on the obtained probability and the image segmentation result, a probability map can be established, and each pixel position in the probability map has a probability value corresponding to the parking space category and a probability value corresponding to the background category.
  • normalization processing softmax
  • Normalization processing is used to map the input to a real number between 0 and 1, and the normalization guarantee sum is 1, so that the sum of the probabilities of multiple classifications is also exactly 1.
  • Sub-step 2032 Perform maximum value independent variable point set processing on the probability map to obtain a label map, in which the area in the parking space instance is set with a parking space label, and the area in the background instance is set with a background Label.
  • the probability map is obtained, and each pixel position in the probability map has a probability value corresponding to a parking space category and a probability value corresponding to a background category. Therefore, in this step, the probability map can be processed for the maximum independent variable point set to obtain a label map.
  • Each pixel position in the label map has a label of the corresponding category. It has a parking space label; if a pixel position belongs to the area of the background instance, it has a background label.
  • the probability map is processed for the maximum independent variable point set
  • the area of the parking space instance and the area of the background instance in the probability map can be determined according to the processed probability value, and then the two areas can be determined. Add the corresponding labels respectively, that is, the label map of 1 ⁇ H ⁇ W dimension is obtained.
  • Sub-step 2033 based on the parking space label and the background label, distinguish the parking space instance and the background instance in the image segmentation result.
  • each pixel position in the label map has a label of a corresponding category
  • the parking space instance and the background instance in the image segmentation result can be distinguished based on the parking space label and the background label in the label map.
  • a mask that is not a background label can be obtained by calculation according to the label map.
  • Step 204 Screen out the parking space instances whose area is less than or equal to the preset area threshold from all the parking space instances, and extract the area where the remaining parking space instances are located.
  • M parking space instances can be determined based on the obtained mask map that is not a background label, and based on a preset instance (connected domain) area, the area of all parking spaces instances is removed that is less than or equal to the preset area The parking space instance of the area threshold, so as to eliminate the unqualified parking space instance.
  • Step 205 According to the offset information between the N position points in the area where each parking space instance is located and the parking space corner points, establish N groups of candidate sets corresponding to each of the parking space instances, each group of the candidate sets.
  • the set includes the locations of 4 parking corner corners.
  • step 205 For details of step 205, reference may be made to the above-mentioned 103, which will not be repeated here.
  • the order of the four parking space corner points may be any predefined order.
  • the positions of the four corner points of the parking spaces may be the absolute distance or relative distance in pixels from the current position point to the corner points of the four parking spaces, or may be the relative distance or absolute distance in pixels under a predefined scale.
  • the N position points include: all the position points in the area where the parking space instance is located, or N position points that are respectively in N preset positions in the area where the parking space instance is located, or A position point in at least one preset area in the area where the parking space instance is located.
  • the N location points in the area where the parking space instance is located may include: all location points in the area where the parking space instance is located, or N pre-defined location points in the area where the parking space instance is located, or the parking space instance
  • the set of location points in at least one preset area in the area where the area is located is not limited in this embodiment of the present application.
  • Step 206 Determine a target candidate set from the N groups of candidate sets for each parking space instance, and determine a target parking space corresponding to the parking space instance according to the target candidate set.
  • step 206 reference may be made to 104 above, and details are not repeated here.
  • Sub-step 2061 according to the weight value of each of the four parking space corner points, perform a weighted average calculation on the positions of the parking space corner points in the N groups of candidate sets for each of the parking space instances, and obtain a parking space including four parking spaces.
  • the corresponding weight values can be set for the four parking space corner points according to their importance requirements, and then the weighted average of the four parking space corner points in all N groups of candidate sets is performed to obtain the final one. Group target candidate set.
  • the weighted average of the positions of the N upper left parking space corner points in all N groups of candidate sets is performed to obtain the position of the target upper left parking space corner point; Average, get the position of the corner of the target upper right parking space; perform the weighted average of the positions of the N lower left parking space corner points in all N groups of candidate sets to obtain the position of the target lower left parking space corner; The position of the lower right parking space corner point is weighted to obtain the position of the target lower right parking space corner point; finally based on the position of the target upper left parking space corner point, the position of the target lower left parking space corner point, the position of the target upper right parking space corner point, and the target lower right parking space corner point position The position of the corner of the parking space is obtained, and the target candidate set is obtained.
  • step 206 may specifically include:
  • Sub-step 2062 For each of the N groups of candidate sets of the parking space instance, obtain the number of votes obtained for each group of the candidate sets.
  • Sub-step 2063 Determine the candidate set with the largest number of votes as the target candidate set.
  • the number of votes obtained for each group of the candidate sets can also be obtained, and the candidate set with the largest number of votes is determined as the target candidate set, wherein the voting process for each group of candidate sets can be performed by Computer-designed voting.
  • the prediction model also outputs the confidence level of the parking space corner; each candidate set has a total confidence level; step 206 may specifically include:
  • Sub-step 2064 Sort the candidate set in descending order of total confidence.
  • Sub-step 2065 Calculate the overlap ratio between the quadrilateral formed by each candidate set with a larger total confidence and the quadrilateral formed by all candidate sets with a smaller total confidence.
  • Sub-step 2066 delete the candidate set with the largest overlap ratio and the lower total confidence.
  • Sub-step 2067 based on the remaining candidate sets, enter the step of sorting the candidate sets in descending order of total confidence, until only one set of the target candidate sets remains.
  • NMS Non-maximum Suppression
  • the idea of NMS is to construct a corresponding quadrilateral according to the positions of the four corner points of the parking spaces in the candidate set. , after sorting the candidate sets according to the total confidence, from the candidate set with a lower score than a certain candidate set, exclude the candidate set with a higher quadrilateral overlap rate, and then repeat the process of sorting, overlapping rate calculation and deletion, and finally Take the remaining set of candidate sets as the target candidate set.
  • the prediction model also outputs attribute information of the parking space corner; after step 206, the method may further include:
  • Step 207 Correspondingly add the attribute information of the four corners of the parking spaces included in the target candidate set to the four corners of the parking spaces included in the target parking space.
  • the attribute information includes one or more of visible type, invisible type, parking space entrance corner type, and confidence level.
  • the corner attribute branch includes specific attributes of each parking space corner, such as visible, invisible , parking space entrance corner, confidence, etc.
  • the attribute information of the four corner points of the parking spaces can be extracted from the corner attribute branch, and the attribute information can be added to Display in the four corners of the target parking space to improve the valuable information content in the target parking space.
  • the corner point of the entrance of the parking space has a great influence on the parking process, and the parking space is accurately marked.
  • the attribute of the corner point of the parking space at the entrance of the parking space helps to improve the accuracy of automatic parking.
  • the parking space instances include at least: a parking space instance in a state of being able to park, and an instance of a parking space being in a state of not being able to park; after step 206, the method may further include:
  • Step 208 In the case that the parking space instance corresponding to the target candidate set is a parking space instance in a parking space state, determine that the state of the target parking space is a parking space state.
  • Step 209 In the case that the parking space instance corresponding to the target candidate set is a parking space instance in a parking space unavailable state, determine that the target parking space state is a parking space unavailable state.
  • the instance result branch includes background examples, various types of parking space instances (such as parking spaces, non-parking spaces) etc.), when the target parking space is finally obtained, according to the target parking space instance corresponding to the target candidate set corresponding to the target parking space, the type corresponding to the target parking space instance can be found in the instance result branch (for example, parking space, non-parking space etc.), and take the parking space status corresponding to the found type as the parking space status of the target parking space, so that the target parking space has more abundant status information.
  • various types of parking space instances such as parking spaces, non-parking spaces etc.
  • the method for determining a parking space provided by the embodiment of the present application is processed based on the global features of the parking space, which is different from the tedious extraction and processing of local features near the corners of the parking space in the related art, thus reducing the amount of calculation.
  • the four corner points of the parking space are calculated according to the offset, so even if the corner points of the parking space are blocked in the actual environment, the corresponding offset information will still be obtained, so that the blocking of the corner points of the parking space will not affect the parking space.
  • the calculation of the corner points has an impact and improves the success rate of parking space recognition.
  • FIG. 4 is a block diagram of an apparatus for determining a parking space provided by an embodiment of the present application.
  • the apparatus for determining a parking space 400 may include: an acquisition module 401 and a processing module 402;
  • the acquiring module 401 is configured to perform: acquiring a bird's-eye view of the parking area, and extracting image features of the bird's-eye view of the parking area;
  • the processing module 402 is used to execute:
  • each group of said candidate sets includes the positions of 4 parking space corners;
  • a target candidate set is determined from the N groups of candidate sets for each parking space instance, and a target parking space corresponding to the parking space instance is determined according to the target candidate set.
  • processing module is specifically used for:
  • processing module is specifically configured to execute:
  • the image segmentation result is normalized to determine the class probability value of each position point in the image segmentation result, so as to obtain a probability map;
  • the class probability value includes the probability value of the corresponding parking space class, the probability value of the corresponding background class. probability value;
  • the parking space instance and the background instance in the image segmentation result are distinguished.
  • the N position points include: all the position points in the area where the parking space instance is located, or N position points that are respectively in N preset positions in the area where the parking space instance is located, or A position point in at least one preset area in the area where the parking space instance is located.
  • the prediction model also outputs attribute information of the corner of the parking space; the processing module is further configured to execute:
  • the attribute information of the four corners of parking spaces included in the target candidate set is correspondingly added to the four corners of parking spaces included in the target parking space.
  • the attribute information includes one or more of a visible type, an invisible type, a parking space entrance corner type, and a confidence level.
  • the parking space instances include at least: a parking space instance in a parking space state, a parking space instance in a non-parking space state;
  • the processing module is also used to execute:
  • the parking space instance corresponding to the target candidate set is a parking space instance in a parking space state, determining that the target parking space state is a parking space state;
  • the state of the target parking space is determined to be a parking space unavailable state.
  • each corner of the parking space is provided with a corresponding weight value; the processing module is specifically configured to execute:
  • the position of the parking space corner points in the N groups of candidate sets of each said parking space instance is weighted and averaged to obtain the weighted value including the four parking space corner points.
  • the set of target candidates for the average position is weighted and averaged to obtain the weighted value including the four parking space corner points.
  • processing module is specifically configured to execute:
  • the candidate set with the largest number of votes is determined as the target candidate set.
  • processing module is specifically configured to execute:
  • Determining a target candidate set from the N groups of candidate sets for each parking space instance includes:
  • the step of sorting the candidate sets in descending order of total confidence is entered, until only one set of the target candidate sets remains.
  • the parking space determination device performs processing based on the global characteristics of the parking space, which is different from the tedious extraction and processing of local features near the corners of the parking space in the related art, so the amount of calculation is reduced, and The four corner points of the parking space are calculated according to the offset, so even if the corner points of the parking space are blocked in the actual environment, the corresponding offset information will still be obtained, so that the blocked corner points of the parking space will not affect the corner points of the parking space.
  • the calculation has an impact and improves the success rate of parking space identification.
  • Embodiments of the present application further provide a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium.
  • a computer program is stored on the computer-readable storage medium.
  • the computer program is executed by a processor, each process of the above embodiment of the parking space determination method is implemented, and the same technology can be achieved. The effect, in order to avoid repetition, is not repeated here.
  • the computer-readable storage medium such as read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), magnetic disk or optical disk and so on.
  • the acquiring module may be an interface connecting the external control terminal with the parking space determining device.
  • the external control terminal may include a wired or wireless headset port, an external power (or battery charger) port, a wired or wireless data port, a memory card port, a port for connecting a control terminal with an identification module, an audio input /Output (I/O) ports, video I/O ports, headphone ports, and more.
  • the acquisition module may be used to receive input (eg, data information, power, etc.) from an external control terminal and transmit the received input to one or more elements within the parking space determination device or may be used to communicate between the parking space determination device and external Data transfer between control terminals.
  • At least one magnetic disk storage device For example at least one magnetic disk storage device, flash memory device, or other volatile solid state storage device.
  • the processor is the control center of the control terminal. It uses various interfaces and lines to connect various parts of the entire control terminal, and executes control by running or executing the software programs and/or modules stored in the memory and calling the data stored in the memory. Various functions of the terminal and processing data, so as to carry out overall monitoring of the control terminal.
  • the processor may include one or more processing units; preferably, the processor may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface and application programs, etc., and the modem processor Mainly deals with wireless communication. It can be understood that, the above-mentioned modulation and demodulation processor may not be integrated into the processor.
  • the embodiments of the present application may be provided as a method, a control terminal, or a computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.
  • computer-usable storage media including, but not limited to, disk storage, CD-ROM, optical storage, etc.
  • These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer readable memory result in an article of manufacture comprising the instruction to control the terminal,
  • the instruction controls the terminal to implement the function specified in one flow or multiple flows of the flowchart and/or one block or multiple blocks of the block diagram.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Traffic Control Systems (AREA)

Abstract

一种车位确定方法、装置、摄像模组及可移动设备，所述方法包括：获取停车区域鸟瞰图，并提取停车区域鸟瞰图的图像特征(101)；将图像特征输入预设的预测模型，得到图像分割结果以及偏移量信息(102)；从图像分割结果中提取车位实例，并建立每个车位实例对应的N组候选集合(103)；从每个车位实例的N组候选集合中确定一个目标候选集合，并根据目标候选集合，确定与车位实例对应的目标车位(104)。本申请是基于车位的全局特征进行处理的，不同于相关技术中对车位角点附近的局部特征进行繁琐的提取和处理，所以降低了计算量，在实际环境中即使车位角点被遮挡，也依然会得到相应的偏移量信息，提高了车位识别的成功率。

Description

车位确定方法及装置 技术领域
本申请涉及计算机技术领域，特别是涉及一种车位确定方法、装置及计算机可读存储介质。
背景技术
在自动泊车场景中，车位检测是一个关键的环节，泊车系统通过车位检测功能可以感知周围车位的具体位置，由此可知，车位检测功能是实现自动泊车的基础。
相关技术中，车位检测首先需提取车位角点附近区域的角点局部特征，并通过对角点局部特征的分析，识别车位角点的位置，最后根据车位角点的位置，识别车位的具体位置。
但是，目前方案中，当车位角点发生遮挡时，由于无法提取角点局部特征，会导致车位识别失败。
发明内容
本申请提供一种车位确定方法、装置、摄像模组及可移动设备，可以解决现有技术中当车位角点发生遮挡时，由于无法提取角点局部特征，会导致车位识别失败的问题。
第一方面，本申请实施例提供了一种车位确定方法，包括：
获取停车区域鸟瞰图，并提取所述停车区域鸟瞰图的图像特征；
将所述图像特征输入预设的预测模型，得到包括车位实例和背景实例的图像分割结果，以及所述图像分割结果中的位置点与所述车位角点之间的偏移量信息；
从所述图像分割结果中提取所述车位实例所处区域，并根据每个所述车位实例所处区域中的N个位置点与车位角点之间的偏移量信息，建立每个所述车位实例对应的N组候选集合，每组所述候选集合包括4个车位角点的位置；
从每个所述车位实例的N组候选集合中确定一个目标候选集合，并根据所述目标候选集合，确定与所述车位实例对应的目标车位。
第二方面，本申请实施例提供了一种车位确定装置，包括：获取模块和 处理器；
所述获取模块用于：获取停车区域鸟瞰图，并提取所述停车区域鸟瞰图的图像特征；
所述处理模块用于：将所述图像特征输入预设的预测模型，得到包括车位实例和背景实例的图像分割结果，以及所述图像分割结果中的位置点与所述车位角点之间的偏移量信息；
从所述图像分割结果中提取所述车位实例所处区域，并根据每个所述车位实例所处区域中的N个位置点与车位角点之间的偏移量信息，建立每个所述车位实例对应的N组候选集合，每组所述候选集合包括4个车位角点的位置；
从每个所述车位实例的N组候选集合中确定一个目标候选集合，并根据所述目标候选集合，确定与所述车位实例对应的目标车位。
第三方面，本申请提供一种计算机可读存储介质，所述计算机可读存储介质包括指令，当其在计算机上运行时，使得计算机执行上述方面所述的方法。
第四方面，本申请提供一种计算机程序产品，所述计算机程序产品包括指令，当其在计算机上运行时，使得计算机执行上述方面所述的方法。
在本申请实施例中，本申请是基于车位的全局特征进行处理的，不同于相关技术中对车位角点附近的局部特征进行繁琐的提取和处理，所以降低了计算量，并且车位的四个车位角点是根据偏移量计算得到的，所以实际环境中即使车位角点被遮挡，也依然会得到相应的偏移量信息，使得车位角点被遮挡不会对车位角点的计算产生影响，提高了车位识别的成功率。
附图说明
图1是本申请实施例提供的一种车位确定系统的架构图；
图2是本申请实施例提供的一种车位确定方法的流程图；
图3是本申请实施例提供的一种车位确定方法的具体流程图；
图4是本申请实施例提供的一种车位确定装置的框图。
具体实施方式
为了使得本发明的目的、技术方案和优点更为明显，下面将参照附图详细描述根据本发明的示例实施例。显然，所描述的实施例仅仅是本发明的一部分实施例，而不是本发明的全部实施例，应理解，本发明不受这里描述的示例实施例的限制。基于本发明中描述的本发明实施例，本领域技术人员在没有付出创造性劳动的情况下所得到的所有其它实施例都应落入本发明的保护范围之内。
在下文的描述中，给出了大量具体的细节以便提供对本发明更为彻底的理解。然而，对于本领域技术人员而言显而易见的是，本发明可以无需一个或多个这些细节而得以实施。在其他的例子中，为了避免与本发明发生混淆，对于本领域公知的一些技术特征未进行描述。
应当理解的是，本发明能够以不同形式实施，而不应当解释为局限于这里提出的实施例。相反地，提供这些实施例将使公开彻底和完全，并且将本发明的范围完全地传递给本领域技术人员。
在此使用的术语的目的仅在于描述具体实施例并且不作为本发明的限制。在此使用时，单数形式的“一”、“一个”和“所述/该”也意图包括复数形式，除非上下文清楚指出另外的方式。还应明白术语“组成”和/或“包括”，当在该说明书中使用时，确定所述特征、整数、步骤、操作、元件和/或部件的存在，但不排除一个或更多其它的特征、整数、步骤、操作、元件、部件和/或组的存在或添加。在此使用时，术语“和/或”包括相关所列项目的任何及所有组合。
为了彻底理解本发明，将在下列的描述中提出详细的结构，以便阐释本发明提出的技术方案。本发明的可选实施例详细描述如下，然而除了这些详细描述外，本发明还可以具有其他实施方式。
下面结合附图，对本申请的图像标注方法和装置、系统进行详细说明。在不冲突的情况下，下述的实施例及实施方式中的特征可以相互组合。
参照图1，其示出了一种车位确定系统的架构图，该车位确定系统包括：特征提取模块、预测模块和结果输出模块，其中，特征提取模块可以以停车场景下采集的停车区域鸟瞰图10为输入，输出停车区域鸟瞰图10的图像特征，将图像特征再输入预测模块后，可以基于图像特征进行图像分割操作，得到三个结果分支：实例结果分支、角点偏移量分支、角点属性分支。
实例结果分支包含基于停车区域鸟瞰图10的图像特征提取得到的图像 分割结果20，图像分割结果20包含背景示例、各种类型的车位实例(如可停车位、不可停车位等)；图像分割结果20仅包含图像分割得到的各个对象示例，而针对其中的车位实例，还未进行具体车位角点的位置和属性的赋予。
角点偏移量分支则包含了图像分割结果20中的位置点与车位角点之间的偏移量信息；如，由于一个车位由4个车位角点组成，则针对一位置点p，角点偏移量分支具有该位置点p相对于左上车位角点的偏移量△p1；该位置点p相对于右上车位角点的偏移量△p2；该位置点p相对于左下车位角点的偏移量△p3；该位置点p相对于右下车位角点的偏移量△p4。
角点属性分支则包含每个车位角点的具体属性，如，可见、不可见、车位入口角点、置信度等。
在本申请实施例中，可以从图像分割结果20中提取车位实例所处区域，并从角点偏移量分支中提取每个车位实例所处区域中的N个位置点与车位角点之间的偏移量信息，进而建立每个车位实例对应的N组候选集合，每组候选集合包括4个车位角点的位置；即针对每个车位实例，建立有N个候选车位区域，进一步的，可以基于一种策略，从每个车位实例的N组候选集合中确定一个目标候选集合，并将由目标候选集合建立的车位区域，作为车位实例对应的目标车位。最后可以从角点属性分支中提取目标候选集合中四个车位角点的属性信息，以及从实例结果分支中提取该车位实例对应的车位类型，由车位角点的属性类型和车位类型，可以确定目标车位的状态以及其中四个车位角点的属性，对所有的车位实例处理完毕后，得到最终的车位确定结果30，车位确定结果30包含了各个车位实例的具体位置、状态和车位角点属性，使得自动泊车系统可以基于车位确定结果30进行准确的泊车操作。
在本申请实施例中，本申请提供的方案是基于车位的全局特征进行处理的，不同于相关技术中对车位角点附近的局部特征进行繁琐的提取和处理，所以降低了计算量，并且车位的四个车位角点是根据偏移量计算得到的，所以实际环境中即使车位角点被遮挡，也依然会得到相应的偏移量信息，使得车位角点被遮挡不会对车位角点的计算产生影响，提高了车位识别的成功率。
图2是本申请实施例提供的一种车位确定方法的流程图，如图2所示，该方法可以包括：
步骤101、获取停车区域鸟瞰图，并提取所述停车区域鸟瞰图的图像特征。
在本申请实施例中，停车区域鸟瞰图可以为在车辆顶部设置的多摄模块采集得到的，多摄模块包括多个摄像头，在标定了各个摄像头透视投影到同一个鸟瞰图平面上后，可以通过多摄模块进行拍摄，采集车辆所处区域的停车区域鸟瞰图。
进一步的，由于停车区域鸟瞰图的数据维度较高，直接对其进行处理难度较大，因此可以提取停车区域鸟瞰图的图像特征，从而降低数据维度，后续对图像特征进行处理难度较低。
具体的，特征是某一类对象区别于其他类对象的相应特点或特性，或是这些特点和特性的集合，特征是通过测量或处理能够抽取的数据，特征提取的主要目的是提炼能代表停车区域鸟瞰图相应特点或特性的表示，且其主要思想是将停车区域鸟瞰图的图像视觉特点或模式投影到一个特征空间中，得到最能反应停车区域鸟瞰图本质或进行停车区域鸟瞰图区分的图像特征。
在本申请实施例中，图像特征可以通过特征向量表达式进行表达，如，f＝{x1，x2…xn}，常见的图像特征提取方法包括：(1)几何法特征提取，几何法是建立在图像纹理基元理论基础上的一种纹理特征分析方法。(2)模型法特征提取，模型法以图像的构造模型为基础，采用模型的参数作为纹理特征，例如卷积神经网络模型。(3)信号处理法特征提取，纹理特征的提取与匹配主要有：灰度共生矩阵、自回归纹理模型、小波变换等。
步骤102、将所述图像特征输入预设的预测模型，得到包括车位实例和背景实例的图像分割结果，以及所述图像分割结果中的位置点与所述车位角点之间的偏移量信息。
在本申请实施例中，预测模型可以基于图像特征进行图像分割以及车位角点信息预测，从而输出包括车位实例和背景实例的图像分割结果，以及图像分割结果中的位置点与所述车位角点之间的偏移量信息。
参照图1，预测模型的输出结果包括三个分支，实例结果分支包含基于停车区域鸟瞰图10的图像特征提取得到的图像分割结果20，图像分割结果20包含背景示例、各种类型的车位实例(如可停车位、不可停车位等)；图像分割结果20仅包含图像分割得到的各个对象实例，而针对其中的车位实例，还未进行具体车位角点的位置和属性的赋予。
角点偏移量分支则包含了图像分割结果20中的位置点与车位角点之间的偏移量信息；如，由于一个车位由4个车位角点组成，则针对一位置点p，角点偏移量分支具有该位置点p相对于左上车位角点的偏移量△p1；该位置点p相对于右上车位角点的偏移量△p2；该位置点p相对于左下车位角点的偏移量△p3；该位置点p相对于右下车位角点的偏移量△p4。
具体的，图像分割结果是一个3×H×W维度的输出结果，其中H×W是输出结果的分辨率，H×W可以是和停车区域鸟瞰图的分辨率大小一致，也可以是停车区域鸟瞰图的分辨率大小的1/2、1/4、1/8等，本申请实施例对此不作限定。
进一步的，角点偏移量分支的维度为8×H×W。8个通道是指：车位内部的一个位置点p到四个车位角点的偏移量，分别在x轴和y轴上的分量大小。
步骤103、从所述图像分割结果中提取所述车位实例所处区域，并根据每个所述车位实例所处区域中的N个位置点与车位角点之间的偏移量信息，建立每个所述车位实例对应的N组候选集合，每组所述候选集合包括4个车位角点的位置。
在本申请实施例中，参照图1，可以从图像分割结果20中提取车位实例所处区域，并从角点偏移量分支中提取每个车位实例所处区域中的N个位置点与车位角点之间的偏移量信息，进而建立每个车位实例对应的N组候选集合，每组候选集合包括4个车位角点的位置；即针对每个车位实例，建立有N个候选车位区域。
例如，针对一个车位实例，具有所处区域中的位置点1与四个偏移量之间的对应关系、位置点2与四个偏移量之间的对应关系….位置点N与四个偏移量之间的对应关系，每个位置点与其对应的四个偏移量可以求得一个组候选集合，在建立每组候选集合的过程中，根据每个位置点的坐标和四个偏移量的大小，可以求出四个车位角点(左上、左下、右上、右下)的位置坐标，即得到一组候选集合。
步骤104、从每个所述车位实例的N组候选集合中确定一个目标候选集合，并根据所述目标候选集合，确定与所述车位实例对应的目标车位。
在本申请实施例中，得到每个车位实例对应的N组候选集合之后，可以基于一种策略，从每个车位实例的N组候选集合中确定一个目标候选集合， 并将由目标候选集合建立的车位区域，作为车位实例对应的目标车位。其中，该策略具体可以包括：加权平均策略、投票策略、非极大值抑制(NMS，
Non-maximum Suppression)算法策略等。本申请实施例对该策略的具体选取不作限定。
综上，本申请实施例提供的一种车位确定方法，是基于车位的全局特征进行处理的，不同于相关技术中对车位角点附近的局部特征进行繁琐的提取和处理，所以降低了计算量，并且车位的四个车位角点是根据偏移量计算得到的，所以实际环境中即使车位角点被遮挡，也依然会得到相应的偏移量信息，使得车位角点被遮挡不会对车位角点的计算产生影响，提高了车位识别的成功率。
图3是本申请实施例提供的一种车位确定方法的具体流程图，如图3所示，该方法可以包括：
步骤201、获取停车区域鸟瞰图，并提取所述停车区域鸟瞰图的图像特征。
步骤201具体可以参照上述101，此处不再赘述。
步骤202、将所述图像特征输入预设的预测模型，得到包括车位实例和背景实例的图像分割结果，以及所述图像分割结果中的位置点与所述车位角点之间的偏移量信息。
步骤202具体可以参照上述102，此处不再赘述。
步骤203、区分所述图像分割结果中的车位实例和背景实例。
在本申请实施例中，基于预测模型输出的图像分割结果以及偏移量信息，建立每个车位实例的N组候选集合之前，需要基于图像分割结果，区分其中的车位实例和背景实例，并提取车位实例。
可选的，步骤203具体可以包括：
子步骤2031、对所述图像分割结果进行归一化处理，确定所述图像分割结果中每一个位置点的类别概率值，从而得到概率图；所述类别概率值包括对应车位类别的概率值、对应背景类别的概率值。
在本申请实施例中，可以首先对图像分割结果进行归一化处理，确定图像分割结果中每一个位置点的类别概率值，即每个位置点所属于车位类别的概率以及所属于背景类别的概率。基于求得的概率和图像分割结果，可以建立得到概率图，概率图中每个像素点位置具有对应车位类别的概率值和对应背景类别的概率值。其中，归一化处理(softmax)是机器学习尤其是深度学 习中常用的一种数据处理手段，归一化处理用于把输入映射为0-1之间的实数，并且归一化保证和为1，从而使得多分类的概率之和也刚好为1。
子步骤2032、对所述概率图进行求最大值自变量点集处理，得到标签图，所述标签图中处于所述车位实例的区域设置有车位标签，处于所述背景实例的区域设置有背景标签。
在本申请实施例中，由于得到了概率图，且概率图中每个像素点位置具有对应车位类别的概率值和对应背景类别的概率值。因此该步骤可以对概率图进行求最大值自变量点集处理，得到标签图，标签图中每个像素点位置具有相应类别的标签，如，一个像素点位置所属于车位实例的区域，则其具有车位标签；一个像素点位置所属于背景实例的区域，则其具有背景标签。其中，最大值自变量点集(argmax)的概念为寻找具有最大评分的参量，定义函数y＝f(x)，则x0＝argmax(f(x))的的含义包括：参数x0满足f(x0)为f(x)的最大值；换句话说就是argmax(f(x))是使得f(x)取得最大值所对应的变量x。本申请实施例中，对概率图进行求最大值自变量点集处理后，可以根据处理后的概率值，确定概率图中所属于车位实例的区域和背景实例的区域，之后对这两个区域分别添加对应的标签，即得到了1×H×W维度的标签图。
子步骤2033、基于所述车位标签和所述背景标签，区分所述图像分割结果中的车位实例和背景实例。
在本申请实施例中，由于标签图中每个像素点位置具有相应类别的标签，则可以基于标签图中的车位标签和所述背景标签，区分所述图像分割结果中的车位实例和背景实例，具体的，本申请实施例可以根据标签图，计算得到不是背景标签的掩码图(mask)。
步骤204、对所有所述车位实例中面积小于或等于预设面积阈值的车位实例进行筛除，并提取剩余的车位实例所处的区域。
在该步骤中，可以基于上述得到的不是背景标签的掩码图确定M个车位实例(连通域)，并基于预设的实例(连通域)面积，去除所有车位实例中面积小于或等于预设面积阈值的车位实例，从而剔除不合格的车位实例。
步骤205、根据每个所述车位实例所处区域中的N个位置点与车位角点之间的偏移量信息，建立每个所述车位实例对应的N组候选集合，每组所述候选集合包括4个车位角点的位置。
步骤205具体可以参照上述103，此处不再赘述。
其中，车位实例内部的位置点回归出车位的四个车位角点的过程中，四 个车位角点的顺序可以是预先定义好的任意顺序。且四个车位角点的位置，可以是当前位置点到四个车位角点的像素绝对距离或相对距离，也可以是预先定义尺度下的相对距离或者像素绝对距离。
可选的，所述N个位置点包括：所述车位实例所处区域中所有的位置点，或分别处于所述车位实例所处区域中的N预设位置的N个位置点，或在所述车位实例所处区域中的至少一个预设区域内的位置点。
在本申请实施例中，车位实例所处区域中的N个位置点可以包括：车位实例所处区域的所有位置点，或车位实例所处区域中预先定义好的N个位置点，或车位实例所处区域中的至少一个预设区域内的位置点的集合，本申请实施例对此不作限定。
步骤206、从每个所述车位实例的N组候选集合中确定一个目标候选集合，并根据所述目标候选集合，确定与所述车位实例对应的目标车位。
步骤206具体可以参照上述104，此处不再赘述。
可选的，在一种实现方式中，每个所述车位角点设置有对应的权重值；步骤206具体可以包括：
子步骤2061、根据4个车位角点中每个所述车位角点的权重值，对每个所述车位实例的N组候选集合中车位角点的位置进行加权平均计算，得到包括4个车位角点的加权平均位置的目标候选集合。
在本申请实施例中，可以对4个车位角点按照其重要性的要求，分别设置对应的权重值，之后对所有N组候选集合中的4个车位角点进行加权平均，得到最终的一组目标候选集合。
具体为，对所有N组候选集合中的N个左上车位角点进行位置的加权平均，得到目标左上车位角点的位置；对所有N组候选集合中的N个右上车位角点进行位置的加权平均，得到目标右上车位角点的位置；对所有N组候选集合中的N个左下车位角点进行位置的加权平均，得到目标左下车位角点的位置；对所有N组候选集合中的N个右下车位角点进行位置的加权平均，得到目标右下车位角点的位置；最终基于目标左上车位角点的位置、目标左下车位角点的位置、目标右上车位角点的位置、目标右下车位角点的位置，得到目标候选集合。
可选的，在另一种实现方式中，步骤206具体可以包括：
子步骤2062、针对每个所述车位实例的N组候选集合，获取对每组所述候选集合的得票数量。
子步骤2063、将得票数量最大的候选集合确定为所述目标候选集合。
在另一种实现方式中，也可以获取对每组所述候选集合的得票数量，并将得票数量最大的候选集合确定为所述目标候选集合，其中，针对每组候选集合的投票过程可以通过计算机设计的投票方式所实现。
可选的，在另一种实现方式中，所述预测模型还输出所述车位角点的置信度；每个所述候选集合具有总置信度；步骤206具体可以包括：
子步骤2064、将所述候选集合按照总置信度由大到小的顺序进行排序。
子步骤2065、计算每个总置信度较大的候选集合形成的四边形，与所有总置信度较小的候选集合形成的四边形之间的重叠率。
子步骤2066、将所述重叠率最大的总置信度较小候选集合删除。
子步骤2067、基于剩余的所述候选集合，进入所述将所述候选集合按照总置信度由大到小的顺序进行排序的步骤，直至仅剩一组所述目标候选集合。
在本申请实施例中，可以采用非极大值抑制(NMS，Non-maximum Suppression)算法来实现确定目标候选集合，NMS的思想是根据候选集合中的四个车位角点的位置构建对应的四边形，将候选集合按照总置信度排序后，从比某个候选集合得分低的候选集合里，把四边形重叠率较高的候选集合排除掉，然后重复进行排序、重叠率计算和删除的过程，最后将仅剩的一组候选集合作为目标候选集合。
可选的，所述预测模型还输出所述车位角点的属性信息；在步骤206之后，所述方法还可以包括：
步骤207、将所述目标候选集合包含的4个车位角点的属性信息，对应添加至所述目标车位包含的四个车位角点。
可选的，属性信息包括可视类型、不可视类型、车位入口角点类型、置信度中的一种或多种。
在本申请实施例中，参照图1，预测模型的输出结果包括的三个分支中，具有角点属性分支，角点属性分支则包含每个车位角点的具体属性，如，可见、不可见、车位入口角点、置信度等。
本申请实施例在得到目标候选集合包含的用于构建目标车位的四个车位角点后，可以从角点属性分支中提取这四个车位角点的的属性信息，并将这些属性信息添加至目标车位的四个车位角点中进行展示，以提高目标车位中的有价值信息含量，如，在自动泊车时，车位入口的角点对泊车环节影响较大，则精确的标注了处于车位入口的车位角点的属性，有助于提高自动泊车的精度。
可选的，所述车位实例至少包括：处于可停车位状态的车位实例、处于不可停车位状态的车位实例；在步骤206之后，所述方法还可以包括：
步骤208、在所述目标候选集合对应的车位实例为处于可停车位状态的车位实例的情况下，确定所述目标车位的状态为可停车位状态。
步骤209、在所述目标候选集合对应的车位实例为处于不可停车位状态的车位实例的情况下，确定所述目标车位的状态为不可停车位状态。
在本申请实施例中，参照图1，预测模型的输出结果包括的三个分支中，具有实例结果分支，实例结果分支包含背景示例、各种类型的车位实例(如可停车位、不可停车位等)，在最终得到了目标车位时，可以根据目标车位对应的目标候选集合所对应的目标车位实例，在实例结果分支中查找该目标车位实例对应的类型(如，可停车位、不可停车位等)，并将查找到的类型对应的车位状态，作为目标车位的车位状态，从而使得目标车位具备更丰富的状态信息。
综上，本申请实施例提供的一种车位确定方法，是基于车位的全局特征进行处理的，不同于相关技术中对车位角点附近的局部特征进行繁琐的提取和处理，所以降低了计算量，并且车位的四个车位角点是根据偏移量计算得到的，所以实际环境中即使车位角点被遮挡，也依然会得到相应的偏移量信息，使得车位角点被遮挡不会对车位角点的计算产生影响，提高了车位识别的成功率。
图4是本申请实施例提供的一种车位确定装置的框图，如图4所示，该车位确定装置400可以包括：获取模块401和处理模块402；
所述获取模块401用于执行：获取停车区域鸟瞰图，并提取所述停车区域鸟瞰图的图像特征；
所述处理模块402用于执行：
将所述图像特征输入预设的预测模型，得到包括车位实例和背景实例的图像分割结果，以及所述图像分割结果中的位置点与所述车位角点之间的偏移量信息；
从所述图像分割结果中提取所述车位实例所处区域，并根据每个所述车位实例所处区域中的N个位置点与车位角点之间的偏移量信息，建立每个所述车位实例对应的N组候选集合，每组所述候选集合包括4个车位角点的位置；
从每个所述车位实例的N组候选集合中确定一个目标候选集合，并根据所述目标候选集合，确定与所述车位实例对应的目标车位。
可选的，所述处理模块具体用于：
区分所述图像分割结果中的车位实例和背景实例；
对所有所述车位实例中面积小于或等于预设面积阈值的车位实例进行筛除，并提取剩余的车位实例所处的区域。
可选的，所述处理模块具体用于执行：
对所述图像分割结果进行归一化处理，确定所述图像分割结果中每一个位置点的类别概率值，从而得到概率图；所述类别概率值包括对应车位类别的概率值、对应背景类别的概率值；
对所述概率图进行求最大值自变量点集处理，得到标签图，所述标签图中处于所述车位实例的区域设置有车位标签，处于所述背景实例的区域设置有背景标签；
基于所述车位标签和所述背景标签，区分所述图像分割结果中的车位实例和背景实例。
可选的，所述N个位置点包括：所述车位实例所处区域中所有的位置点，或分别处于所述车位实例所处区域中的N预设位置的N个位置点，或在所述车位实例所处区域中的至少一个预设区域内的位置点。
可选的，所述预测模型还输出所述车位角点的属性信息；所述处理模块还用于执行：
将所述目标候选集合包含的4个车位角点的属性信息，对应添加至所述 目标车位包含的四个车位角点。
可选的，所述属性信息包括可视类型、不可视类型、车位入口角点类型、置信度中的一种或多种。
可选的，所述车位实例至少包括：处于可停车位状态的车位实例、处于不可停车位状态的车位实例；
所述处理模块还用于执行：
在所述目标候选集合对应的车位实例为处于可停车位状态的车位实例的情况下，确定所述目标车位的状态为可停车位状态；
在所述目标候选集合对应的车位实例为处于不可停车位状态的车位实例的情况下，确定所述目标车位的状态为不可停车位状态。
可选的，每个所述车位角点设置有对应的权重值；所述处理模块具体用于执行：
根据4个车位角点中每个所述车位角点的权重值，对每个所述车位实例的N组候选集合中车位角点的位置进行加权平均计算，得到包括4个车位角点的加权平均位置的目标候选集合。
可选的，所述处理模块具体用于执行：
针对每个所述车位实例的N组候选集合，获取对每组所述候选集合的得票数量；
将得票数量最大的候选集合确定为所述目标候选集合。
可选的，所述处理模块具体用于执行：
所述从每个所述车位实例的N组候选集合中确定一个目标候选集合，包括：
将所述候选集合按照总置信度由大到小的顺序进行排序；
计算每个总置信度较大的候选集合形成的四边形，与所有总置信度较小的候选集合形成的四边形之间的重叠率；
将所述重叠率最大的总置信度较小候选集合删除；
基于剩余的所述候选集合，进入所述将所述候选集合按照总置信度由大到小的顺序进行排序的步骤，直至仅剩一组所述目标候选集合。
综上，本申请实施例提供的车位确定装置，是基于车位的全局特征进行 处理的，不同于相关技术中对车位角点附近的局部特征进行繁琐的提取和处理，所以降低了计算量，并且车位的四个车位角点是根据偏移量计算得到的，所以实际环境中即使车位角点被遮挡，也依然会得到相应的偏移量信息，使得车位角点被遮挡不会对车位角点的计算产生影响，提高了车位识别的成功率。
本申请实施例还提供一种计算机可读存储介质，计算机可读存储介质上存储有计算机程序，该计算机程序被处理器执行时实现上述车位确定方法实施例的各个过程，且能达到相同的技术效果，为避免重复，这里不再赘述。其中，所述的计算机可读存储介质，如只读存储器(Read-Only Memory，简称ROM)、随机存取存储器(Random Access Memory，简称RAM)、磁碟或者光盘等。
获取模块可以为外部控制终端与车位确定装置连接的接口。例如，外部控制终端可以包括有线或无线头戴式耳机端口、外部电源(或电池充电器)端口、有线或无线数据端口、存储卡端口、用于连接具有识别模块的控制终端的端口、音频输入/输出(I/O)端口、视频I/O端口、耳机端口等等。获取模块可以用于接收来自外部控制终端的输入(例如，数据信息、电力等等)并且将接收到的输入传输到车位确定装置内的一个或多个元件或者可以用于在车位确定装置和外部控制终端之间传输数据。
例如至少一个磁盘存储器件、闪存器件、或其他易失性固态存储器件。
处理器是控制终端的控制中心，利用各种接口和线路连接整个控制终端的各个部分，通过运行或执行存储在存储器内的软件程序和/或模块，以及调用存储在存储器内的数据，执行控制终端的各种功能和处理数据，从而对控制终端进行整体监控。处理器可包括一个或多个处理单元；优选的，处理器可集成应用处理器和调制解调处理器，其中，应用处理器主要处理操作系统、用户界面和应用程序等，调制解调处理器主要处理无线通信。可以理解的是，上述调制解调处理器也可以不集成到处理器中。
本说明书中的各个实施例均采用递进的方式描述，每个实施例重点说明的都是与其他实施例的不同之处，各个实施例之间相同相似的部分互相参见即可。
本领域内的技术人员应明白，本申请的实施例可提供为方法、控制终端、或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘 存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本申请是参照根据本申请的方法、终端设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理终端设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理终端设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的控制终端。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理终端设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令控制终端的制造品，该指令控制终端实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理终端设备上，使得在计算机或其他可编程终端设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程终端设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
尽管已描述了本申请的优选实施例，但本领域内的技术人员一旦得知了基本创造性概念，则可对这些实施例做出另外的变更和修改。所以，所附权利要求意欲解释为包括优选实施例以及落入本申请范围的所有变更和修改。
最后，还需要说明的是，在本文中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者终端设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者终端设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者终端设备中还存在另外的相同要素。
以上对本申请进行了详细介绍，本文中应用了具体个例对本申请的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本申请的方法及其核心思想；同时，对于本领域的一般技术人员，依据本申请的思想，在 具体实施方式及应用范围上均会有改变之处，综上所述，本说明书内容不应理解为对本申请的限制。

Claims (21)

  1. 一种车位确定方法，其特征在于，所述方法包括：
    获取停车区域鸟瞰图，并提取所述停车区域鸟瞰图的图像特征；
    将所述图像特征输入预设的预测模型，得到包括车位实例和背景实例的图像分割结果，以及所述图像分割结果中的位置点与所述车位角点之间的偏移量信息；
    从所述图像分割结果中提取所述车位实例所处区域，并根据每个所述车位实例所处区域中的N个位置点与车位角点之间的偏移量信息，建立每个所述车位实例对应的N组候选集合，每组所述候选集合包括4个车位角点的位置；
    从每个所述车位实例的N组候选集合中确定一个目标候选集合，并根据所述目标候选集合，确定与所述车位实例对应的目标车位。
  2. 根据权利要求1所述的方法，其特征在于，所述从所述图像分割结果中提取所述车位实例所处区域，包括：
    区分所述图像分割结果中的车位实例和背景实例；
    对所有所述车位实例中面积小于或等于预设面积阈值的车位实例进行筛除，并提取剩余的车位实例所处的区域。
  3. 根据权利要求2所述的方法，其特征在于，所述区分所述图像分割结果中的车位实例和背景实例，包括：
    对所述图像分割结果进行归一化处理，确定所述图像分割结果中每一个位置点的类别概率值，从而得到概率图；所述类别概率值包括对应车位类别的概率值、对应背景类别的概率值；
    对所述概率图进行求最大值自变量点集处理，得到标签图，所述标签图中处于所述车位实例的区域设置有车位标签，处于所述背景实例的区域设置有背景标签；
    基于所述车位标签和所述背景标签，区分所述图像分割结果中的车位实例和背景实例。
  4. 根据权利要求1所述的方法，其特征在于，所述N个位置点包括：所述车位实例所处区域中所有的位置点，或分别处于所述车位实例所处区域 中的N预设位置的N个位置点，或在所述车位实例所处区域中的至少一个预设区域内的位置点。
  5. 根据权利要求1所述的方法，其特征在于，所述预测模型还输出所述车位角点的属性信息；
    在所述根据所述目标候选集合，确定与所述车位实例对应的目标车位之后，所述方法还包括：
    将所述目标候选集合包含的4个车位角点的属性信息，对应添加至所述目标车位包含的四个车位角点。
  6. 根据权利要求5所述的方法，其特征在于，所述属性信息包括可视类型、不可视类型、车位入口角点类型、置信度中的一种或多种。
  7. 根据权利要求1所述的方法，其特征在于，所述车位实例至少包括：处于可停车位状态的车位实例、处于不可停车位状态的车位实例；
    在所述根据所述目标候选集合，确定与所述车位实例对应的目标车位之后，所述方法还包括：
    在所述目标候选集合对应的车位实例为处于可停车位状态的车位实例的情况下，确定所述目标车位的状态为可停车位状态；
    在所述目标候选集合对应的车位实例为处于不可停车位状态的车位实例的情况下，确定所述目标车位的状态为不可停车位状态。
  8. 根据权利要求1所述的方法，其特征在于，每个所述车位角点设置有对应的权重值；
    所述从每个所述车位实例的N组候选集合中确定一个目标候选集合，包括：
    根据4个车位角点中每个所述车位角点的权重值，对每个所述车位实例的N组候选集合中车位角点的位置进行加权平均计算，得到包括4个车位角点的加权平均位置的目标候选集合。
  9. 根据权利要求1所述的方法，其特征在于，所述从每个所述车位实例的N组候选集合中确定一个目标候选集合，包括：
    针对每个所述车位实例的N组候选集合，获取对每组所述候选集合的得票数量；
    将得票数量最大的候选集合确定为所述目标候选集合。
  10. 根据权利要求1所述的方法，其特征在于，所述预测模型还输出所述车位角点的置信度；每个所述候选集合具有总置信度；
    所述从每个所述车位实例的N组候选集合中确定一个目标候选集合，包括：
    将所述候选集合按照总置信度由大到小的顺序进行排序；
    计算每个总置信度较大的候选集合形成的四边形，与所有总置信度较小的候选集合形成的四边形之间的重叠率；
    将所述重叠率最大的总置信度较小候选集合删除；
    从剩余的所述候选集合中，选取总置信度最大的候选集合作为所述目标候选集合。
  11. 一种车位确定装置，其特征在于，所述装置包括：获取模块和处理模块；
    所述获取模块用于：获取停车区域鸟瞰图，并提取所述停车区域鸟瞰图的图像特征；
    所述处理模块用于：将所述图像特征输入预设的预测模型，得到包括车位实例和背景实例的图像分割结果，以及所述图像分割结果中的位置点与所述车位角点之间的偏移量信息；
    从所述图像分割结果中提取所述车位实例所处区域，并根据每个所述车位实例所处区域中的N个位置点与车位角点之间的偏移量信息，建立每个所述车位实例对应的N组候选集合，每组所述候选集合包括4个车位角点的位置；
    从每个所述车位实例的N组候选集合中确定一个目标候选集合，并根据所述目标候选集合，确定与所述车位实例对应的目标车位。
  12. 根据权利要求11所述的装置，其特征在于，所述处理模块具体用于：
    区分所述图像分割结果中的车位实例和背景实例；
    对所有所述车位实例中面积小于或等于预设面积阈值的车位实例进行筛除，并提取剩余的车位实例所处的区域。
  13. 根据权利要求12所述的装置，其特征在于，所述处理模块具体用于执行：
    对所述图像分割结果进行归一化处理，确定所述图像分割结果中每一个位置点的类别概率值，从而得到概率图；所述类别概率值包括对应车位类别的概率值、对应背景类别的概率值；
    对所述概率图进行求最大值自变量点集处理，得到标签图，所述标签图中处于所述车位实例的区域设置有车位标签，处于所述背景实例的区域设置有背景标签；
    基于所述车位标签和所述背景标签，区分所述图像分割结果中的车位实例和背景实例。
  14. 根据权利要求11所述的装置，其特征在于，所述N个位置点包括：所述车位实例所处区域中所有的位置点，或分别处于所述车位实例所处区域中的N预设位置的N个位置点，或在所述车位实例所处区域中的至少一个预设区域内的位置点。
  15. 根据权利要求11所述的装置，其特征在于，所述预测模型还输出所述车位角点的属性信息；所述处理模块还用于执行：
    将所述目标候选集合包含的4个车位角点的属性信息，对应添加至所述目标车位包含的四个车位角点。
  16. 根据权利要求15所述的装置，其特征在于，所述属性信息包括可视类型、不可视类型、车位入口角点类型、置信度中的一种或多种。
  17. 根据权利要求11所述的装置，其特征在于，所述车位实例至少包括：处于可停车位状态的车位实例、处于不可停车位状态的车位实例；
    所述处理模块还用于执行：
    在所述目标候选集合对应的车位实例为处于可停车位状态的车位实例的情况下，确定所述目标车位的状态为可停车位状态；
    在所述目标候选集合对应的车位实例为处于不可停车位状态的车位实例的情况下，确定所述目标车位的状态为不可停车位状态。
  18. 根据权利要求11所述的装置，其特征在于，每个所述车位角点设置有对应的权重值；所述处理模块具体用于执行：
    根据4个车位角点中每个所述车位角点的权重值，对每个所述车位实例的N组候选集合中车位角点的位置进行加权平均计算，得到包括4个车位角点的加权平均位置的目标候选集合。
  19. 根据权利要求11所述的装置，其特征在于，所述处理模块具体用于执行：
    针对每个所述车位实例的N组候选集合，获取对每组所述候选集合的得票数量；
    将得票数量最大的候选集合确定为所述目标候选集合。
  20. 根据权利要求11所述的装置，其特征在于，所述处理模块具体用于执行：
    所述从每个所述车位实例的N组候选集合中确定一个目标候选集合，包括：
    将所述候选集合按照总置信度由大到小的顺序进行排序；
    计算每个总置信度较大的候选集合形成的四边形，与所有总置信度较小的候选集合形成的四边形之间的重叠率；
    将所述重叠率最大的总置信度较小候选集合删除；
    基于剩余的所述候选集合，进入所述将所述候选集合按照总置信度由大到小的顺序进行排序的步骤，直至仅剩一组所述目标候选集合。
  21. 一种计算机可读存储介质，其特征在于，包括指令，当其在计算机上运行时，使得所述计算机执行权利要求1至10中任一项所述的车位确定方法。
PCT/CN2021/088440 2021-04-20 2021-04-20 车位确定方法及装置 WO2022222036A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/088440 WO2022222036A1 (zh) 2021-04-20 2021-04-20 车位确定方法及装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/088440 WO2022222036A1 (zh) 2021-04-20 2021-04-20 车位确定方法及装置

Publications (1)

Publication Number Publication Date
WO2022222036A1 true WO2022222036A1 (zh) 2022-10-27

Family

ID=83723635

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/088440 WO2022222036A1 (zh) 2021-04-20 2021-04-20 车位确定方法及装置

Country Status (1)

Country Link
WO (1) WO2022222036A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116161040A (zh) * 2023-04-26 2023-05-26 禾多科技（北京）有限公司 车位信息生成方法、装置、电子设备和计算机可读介质
CN117068145A (zh) * 2023-10-17 2023-11-17 北京茵沃汽车科技有限公司 泊车方法、泊车装置、计算设备和存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109685000A (zh) * 2018-12-21 2019-04-26 广州小鹏汽车科技有限公司 一种基于视觉的车位检测方法及装置
CN110705359A (zh) * 2019-09-05 2020-01-17 北京智行者科技有限公司 一种车位检测方法
EP3624000A1 (en) * 2018-09-13 2020-03-18 Volvo Car Corporation System and method for camera or sensor-based parking spot detection and identification
CN110969655A (zh) * 2019-10-24 2020-04-07 百度在线网络技术（北京）有限公司 用于检测车位的方法、装置、设备、存储介质以及车辆

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3624000A1 (en) * 2018-09-13 2020-03-18 Volvo Car Corporation System and method for camera or sensor-based parking spot detection and identification
CN109685000A (zh) * 2018-12-21 2019-04-26 广州小鹏汽车科技有限公司 一种基于视觉的车位检测方法及装置
CN110705359A (zh) * 2019-09-05 2020-01-17 北京智行者科技有限公司 一种车位检测方法
CN110969655A (zh) * 2019-10-24 2020-04-07 百度在线网络技术（北京）有限公司 用于检测车位的方法、装置、设备、存储介质以及车辆

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116161040A (zh) * 2023-04-26 2023-05-26 禾多科技（北京）有限公司 车位信息生成方法、装置、电子设备和计算机可读介质
CN116161040B (zh) * 2023-04-26 2023-06-27 禾多科技（北京）有限公司 车位信息生成方法、装置、电子设备和计算机可读介质
CN117068145A (zh) * 2023-10-17 2023-11-17 北京茵沃汽车科技有限公司 泊车方法、泊车装置、计算设备和存储介质
CN117068145B (zh) * 2023-10-17 2024-01-26 北京茵沃汽车科技有限公司 泊车方法、泊车装置、计算设备和存储介质

Similar Documents

Publication Publication Date Title
US11830230B2 (en) Living body detection method based on facial recognition, and electronic device and storage medium
CN105574550A (zh) 一种车辆识别方法及装置
CN109711416B (zh) 目标识别方法、装置、计算机设备和存储介质
CN112200045B (zh) 基于上下文增强的遥感图像目标检测模型建立方法及应用
CN105809651B (zh) 基于边缘非相似性对比的图像显著性检测方法
CN110991444B (zh) 面向复杂场景的车牌识别方法及装置
WO2022222036A1 (zh) 车位确定方法及装置
CN113128481A (zh) 一种人脸活体检测方法、装置、设备及存储介质
CN111539456B (zh) 一种目标识别方法及设备
CN112288758B (zh) 一种电力设备红外与可见光图像配准方法
CN113850136A (zh) 基于yolov5与BCNN的车辆朝向识别方法及系统
CN114331946A (zh) 一种图像数据处理方法、设备以及介质
CN117557784B (zh) 目标检测方法、装置、电子设备及存储介质
CN112819008B (zh) 实例检测网络的优化方法、装置、介质及电子设备
CN112419243B (zh) 一种基于红外图像分析的配电房设备故障识别方法
CN115984537A (zh) 图像处理方法、装置及相关设备
CN113704276A (zh) 地图更新方法、装置、电子设备及计算机可读存储介质
CN114973333B (zh) 人物交互检测方法、装置、设备以及存储介质
CN115937492A (zh) 一种基于特征识别的变电设备红外图像识别方法
CN115358981A (zh) 胶水缺陷的确定方法、装置、设备及存储介质
CN115457385A (zh) 一种基于轻量级网络的建筑物变化检测方法
CN111967579A (zh) 使用卷积神经网络对图像进行卷积计算的方法和装置
CN111882545A (zh) 基于双向信息传递及特征融合的织物疵点检测方法
CN111274894A (zh) 一种基于改进YOLOv3的人员在岗状态检测方法
CN116071625B (zh) 深度学习模型的训练方法、目标检测方法及装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21937281

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21937281

Country of ref document: EP

Kind code of ref document: A1