Lightweight parking lot parking space identification method, apparatus and device, and storage medium

Abstract

The invention discloses a lightweight parking lot parking space recognition method, device and equipment and a storage medium, relates to the technical field of image recognition, and discloses a lightweight parking lot parking space recognition method, which comprises the following steps: splicing images collected by cameras to obtain a parking lot map, and determining parking lot image data according to the parking lot map; performing multi-stage image compression processing on the parking lot image data to obtain processed image data; the processed image data are input into a preset image recognition model for edge end lightweight reasoning, an initial parking space recognition result and recognition result confidence are obtained, and the preset image recognition model comprises a depth separable convolution layer and a pyramid compression layer; and performing parking space identification according to the initial parking space identification result and the identification result confidence to obtain a target parking space identification result. And through edge end lightweight reasoning, the accurate identification effect of the parking space of the parking lot on the premise of lightweight is realized.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, and in particular to lightweight parking space recognition methods, devices, equipment and storage media. Background Technology

[0002] With the continuous and rapid growth in the number of motor vehicles in cities, the difficulty in finding parking spaces and the low management efficiency caused by the shortage of parking resources have become the core pain points in the intelligent development of parking lots.

[0003] Against this backdrop, traditional technical solutions typically employ stand-alone camera acquisition and fixed network transmission, relying on a single camera to capture images at a fixed frame rate and uploading them via wired networks or unoptimized wireless methods. On the recognition side, they often use general visual models to determine parking space occupancy, without considering the unique lighting changes, viewing angle differences, and occlusion issues specific to parking lot scenarios. At the map and management level, most solutions rely on preset fixed benchmark maps or manually surveyed data, which cannot be updated according to real-time changes in parking lots. Furthermore, the interfaces between different platforms are not standardized, making it difficult to form stable data linkage.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this application is to provide a lightweight parking space identification method, device, equipment, and storage medium, aiming to solve the technical problem of accurately identifying the occupancy status of parking spaces in parking lots with complex parking space distribution, frequent environmental changes, and limited communication and deployment conditions.

[0006] To achieve the above objectives, this application proposes a lightweight parking space identification method, the method comprising: Images captured by the camera are stitched together to obtain a parking lot map, and parking lot image data is determined based on the parking lot map; The parking lot image data is subjected to multi-level image compression processing to obtain processed image data; The processed image data is input into a preset image recognition model for edge-end lightweight inference to obtain initial parking space recognition results and recognition result confidence. The preset image recognition model includes depth-separable convolutional layers and pyramid compression layers. Based on the initial parking space identification result and the confidence level of the identification result, parking space identification is performed to obtain the target parking space identification result.

[0007] In one embodiment, the step of identifying a target parking space based on the initial parking space identification result and the confidence level of the identification result includes: Obtain the preset reliability threshold; When the confidence level of the recognition result is less than the preset confidence threshold, obtain the neighboring frame image of the image corresponding to the initial parking space recognition result; The initial parking space recognition result is supplemented with complete semantic features based on the neighboring frame images to obtain a supplementary recognition result; The supplementary identification results are uploaded to the cloud so that the cloud can review them and provide feedback on the target parking space identification results.

[0008] In one embodiment, after the step of identifying parking spaces based on the initial parking space identification result and the confidence level of the identification result to obtain the target parking space identification result, the method further includes: Based on the target parking space recognition results, it is determined whether there are any abnormal scenarios in the parking lot. The abnormal scenarios include vehicles parking beyond the boundaries, parking spaces being obstructed by debris, or the camera acquisition device being offset. When an abnormal scenario exists in the parking lot, the warning level of the abnormal scenario is determined, and the warning level identification result is obtained; Based on the warning level identification result, the target anomaly is determined, and the anomaly is resolved in order to identify the parking space.

[0009] In one embodiment, the step of inputting the processed image data into a preset image recognition model for edge-side lightweight inference to obtain an initial parking space recognition result and a recognition result confidence level, wherein the preset image recognition model includes depthwise separable convolutional layers and pyramid compression layers, includes: The processed image data is input into a depth-separable convolutional layer for convolution to obtain convolutional image data. The convolutional image data is input into the pyramid compression layer for edge-end lightweight inference to obtain the initial recognition result; The initial identification results are subjected to category loss calculation and localization loss calculation to obtain the initial parking space identification result and the confidence level of the identification result.

[0010] In one embodiment, the step of stitching together images captured by the camera to obtain a parking lot map includes: Stable feature points are extracted from images captured by the camera based on a scale-invariant feature transformation strategy; The adjacent images captured by the camera are input into the fast nearest neighbor search packet processor to obtain overlapping feature points; Mismatched points among the overlapping feature points are deleted according to the random sampling consistency strategy to obtain corrected feature points; The parking lot map is determined based on the stable feature points and the corrected feature points.

[0011] In one embodiment, the step of determining parking lot image data based on the parking lot map includes: Obtain a first sampling frequency and a second sampling frequency, wherein the first sampling frequency is less than the second sampling frequency; When the parking space status of the parking lot map is not changed, the parking area image is collected according to the first collection frequency to obtain parking lot image data. When a vehicle is detected entering or leaving a parking space area on the parking lot map, parking area images are acquired according to a second acquisition frequency to obtain parking lot image data.

[0012] In one embodiment, the step of performing multi-level image compression processing on the parking lot image data to obtain processed image data includes: The parking lot image data is cropped to obtain an effective image region; The effective image region is subjected to image compression processing to obtain a compressed image region; The compressed image region is encoded to obtain processed image data.

[0013] In addition, to achieve the above objectives, this application also proposes a lightweight parking space recognition device, which includes: a data acquisition module for stitching together images captured by a camera to obtain a parking lot map, and determining parking lot image data based on the parking lot map; The image preprocessing module is used to perform multi-level image compression processing on the parking lot image data to obtain processed image data; The edge inference module is used to input the processed image data into a preset image recognition model for edge-end lightweight inference to obtain the initial parking space recognition result and the confidence level of the recognition result. The preset image recognition model includes a depth-separable convolutional layer and a pyramid compression layer. The parking space recognition module is used to recognize parking spaces based on the initial parking space recognition result and the confidence level of the recognition result, and obtain the target parking space recognition result.

[0014] In addition, to achieve the above objectives, this application also proposes a lightweight parking space recognition device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the lightweight parking space recognition method described above.

[0015] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the lightweight parking space recognition method described above.

[0016] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the lightweight parking space recognition method described above.

[0017] One or more technical solutions proposed in this application have at least the following technical effects: By employing a technique that stitches together images captured by cameras to obtain a parking lot map and determine parking lot image data, the effective image areas related to parking spaces within the parking lot can be focused on, reducing the processing of invalid data. The technique of multi-level image compression of parking lot image data reduces the amount of image data, thereby reducing the burden of data transmission and storage. The technique of inputting processed image data into a pre-defined image recognition model containing depthwise separable convolutional layers and pyramid compression layers for lightweight inference at the edge is also employed. The depthwise separable convolutional layers reduce the number of model parameters and computational load, while the pyramid compression layers reduce computational load while retaining key semantic information, adapting to the limited computing power at the edge to achieve lightweight inference and improve inference efficiency. The technique of combining initial parking space recognition results with the confidence level of the recognition results for parking space identification can filter high-confidence recognition results to improve recognition accuracy. These techniques work together to solve the problems of high computing power requirements, large data transmission volume, and low recognition efficiency in existing parking lot parking space recognition technologies. Compared with existing technologies, this achieves accurate and efficient parking space recognition under lightweight conditions. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art are briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a lightweight parking space identification method according to this application; Figure 2 A flowchart illustrating Embodiment 2 of the lightweight parking space recognition method of this application; Figure 3 A simplified flowchart illustrating the lightweight parking space identification method provided in Embodiment 2 of this application; Figure 4 This is a schematic diagram of the module structure of a lightweight parking space recognition device according to an embodiment of this application; Figure 5 This is a schematic diagram of the hardware operating environment involved in the lightweight parking space identification method in this application embodiment.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0023] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0024] The main solution of this application embodiment is as follows: stitching images captured by a camera to obtain a parking lot map, and determining parking lot image data based on the parking lot map; performing multi-level image compression processing on the parking lot image data to obtain processed image data; inputting the processed image data into a preset image recognition model for edge-end lightweight inference to obtain an initial parking space recognition result and a recognition result confidence level, wherein the preset image recognition model includes a depthwise separable convolutional layer and a pyramid compression layer; and performing parking space recognition based on the initial parking space recognition result and the recognition result confidence level to obtain the target parking space recognition result.

[0025] In this embodiment, for ease of description, the following description will focus on a lightweight parking space identification device.

[0026] To address the limitations of existing technologies in accurately identifying parking space occupancy status in parking lots with complex layouts, frequent environmental changes, and restricted communication and deployment conditions, this application provides a solution. By employing a technique of stitching together images captured by cameras to obtain a parking lot map and determine parking lot image data, it can focus on effective image areas related to parking spaces within the parking lot, reducing the processing of invalid data. Furthermore, by using a multi-level image compression technique for the parking lot image data, it can reduce the amount of image data, thereby reducing the burden of data transmission and storage. Finally, it employs a pre-defined image recognition model containing depthwise separable convolutional layers and pyramid compression layers for edge detection. The proposed lightweight inference technique utilizes depthwise separable convolutional layers to reduce model parameters and computational load, and pyramid compression layers to reduce computational load while preserving key semantic information. This adapts to the limited computing power at the edge, enabling lightweight inference and improving inference efficiency. Furthermore, the technique combines initial parking space recognition results with the confidence level of those results to filter high-confidence recognition outcomes, thereby improving accuracy. These combined techniques address the issues of high computational demands, large data transmission volumes, and low recognition efficiency in existing parking space recognition methods. Compared to existing technologies, this approach achieves accurate and efficient parking space recognition while maintaining a lightweight design.

[0027] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, a lightweight parking space recognition device, or a local execution terminal. The following description uses a local execution terminal as an example to illustrate this embodiment and the subsequent embodiments.

[0028] Based on this, embodiments of this application provide a lightweight parking space identification method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the lightweight parking space recognition method of this application.

[0029] In this embodiment, the lightweight parking space identification method includes steps S10 to S40: Step S10: Stitch together the images captured by the camera to obtain a parking lot map, and determine the parking lot image data based on the parking lot map; It should be noted that the images captured by the cameras are taken by parking space recognition cameras installed in the parking lot. These cameras are selected according to different areas and scenarios in the parking lot. For example, 8mm focal length wide-angle cameras are used in dense areas with an installation height of 3 meters and a downward angle of 45 degrees, while 25mm focal length telephoto cameras are used in open areas with an installation height of 5 meters and a downward angle of 30 degrees. When started, the cameras also capture standard color chart images and correct color deviations through white balance algorithms to ensure consistency of image features.

[0030] In addition, the parking lot map is a complete parking lot plan image generated by stitching together images captured by cameras. It does not rely on pre-stored reference maps or vehicle data. It will be stitched again every 30 seconds to automatically correct map deviations caused by slight camera offsets, and can accurately reflect the distribution of parking spaces in the parking lot.

[0031] Furthermore, the parking lot image data is determined based on the generated parking lot map, mainly containing the image areas corresponding to all parking spaces in the parking lot, excluding invalid areas in the map that are unrelated to parking spaces, thus providing a targeted data foundation for subsequent image compression and parking space recognition.

[0032] Understandably, the process involves first acquiring partitioned images from each camera, then using a feature extraction algorithm to extract at least 50 stable feature points from these images, then using a matcher to find overlapping feature points in adjacent images and removing mismatched points, and finally stitching the images together to obtain a parking lot map using geometric correction techniques. After that, valid image areas containing parking spaces are selected from the map to determine the parking lot image data.

[0033] In one feasible implementation, step S10 may include steps A11 to A14: Step A11: Extract stable feature points from the camera-acquired images using a scale-invariant feature transformation strategy; It should be noted that the Scale-Invariant Feature Transform (SIFT) strategy is an algorithm that can extract stable features of images under different scales, rotations, and lighting conditions. This strategy captures features in images that are not easily affected by environmental changes by analyzing image grayscale changes and edge structures, and is suitable for image feature extraction in complex parking lot environments.

[0034] Additionally, stable feature points are feature points extracted from images captured by the camera using a scale-invariant feature transformation strategy that can reflect the spatial positional relationships of the image. In this embodiment, at least 50 stable feature points are extracted from each image captured by the camera. These feature points typically include points with obvious spatial characteristics, such as intersections of parking lines and corner points of pole shadows.

[0035] Understandably, for each image captured by the camera, a scale-invariant feature transformation strategy is applied to analyze the gray-level gradient and edge contours of each region of the image one by one, and at least 50 stable feature points such as the intersection of parking lines and the corner of the shadow of the pole are identified and extracted, providing feature basis for the matching and stitching of subsequent adjacent images.

[0036] Step A12: Input the adjacent images in the images captured by the camera into the fast nearest neighbor search packet processor to obtain overlapping feature points; It should be noted that the Fast Library for Approximate Nearest Neighbors (FLANN) is a tool for efficiently matching image feature points. This processor uses an approximate nearest neighbor search algorithm to quickly calculate the similarity between different image feature points, greatly improving the efficiency of feature point matching and avoiding the time-consuming problem of traditional matching methods.

[0037] In addition, overlapping feature points are the corresponding feature points in the overlapping areas of adjacent images after the adjacent images are input into the fast nearest neighbor search packet processor. In this embodiment, a matching threshold is set. When the matching degree between feature points is ≥0.7, they are determined to be overlapping feature points, so as to ensure the reliability of the matching results.

[0038] It is understandable that adjacent images are selected from the images captured by the camera, and the stable feature point set of each pair of adjacent images is input into the fast nearest neighbor search packet processor. The processor calculates the similarity between feature points and filters out feature point pairs with a matching degree ≥ 0.7. These feature point pairs are the overlapping feature points of adjacent images.

[0039] Step A13: Delete mismatched points among the overlapping feature points according to the random sampling consistency strategy to obtain corrected feature points; It should be noted that Random Sample Consensus (RANSAC) is an algorithm used to remove mismatched points in feature point matching. This strategy constructs a spatial geometric model by randomly sampling some overlapping feature points, calculates the deviation between the remaining overlapping feature points and the model, and determines points with deviations exceeding a threshold as mismatched points and deletes them, thus ensuring the accuracy of the remaining feature points.

[0040] Additionally, mismatched points are feature points among overlapping feature points that do not conform to the true spatial correspondence between adjacent images. These points are mostly generated by image noise and interference from similar textures, which will affect the image stitching accuracy. Corrected feature points are the overlapping feature points remaining after deleting mismatched points. In this embodiment, the mismatched point removal rate of this strategy is ≤10%, ensuring that the corrected feature points can accurately reflect the spatial relationship between adjacent images.

[0041] Understandably, for overlapping feature points, a random sampling consensus strategy is applied to randomly select some overlapping feature points to construct an image spatial geometric model, calculate the deviation between the remaining overlapping feature points and the model, remove mismatched points whose deviation exceeds a preset threshold, and finally obtain corrected feature points, with the removal rate controlled at ≤10%.

[0042] Step A14: Determine the parking lot map based on the stable feature points and the corrected feature points.

[0043] It should be noted that the parking lot map is a complete parking lot planar image generated by stitching together images from multiple cameras by integrating stable feature points and corrected feature points. This map does not rely on pre-stored benchmark maps or vehicle data and can reflect the distribution and spatial layout of parking spaces in real time and accurately.

[0044] Additionally, stable feature points are used to determine the independent spatial position of each image captured by a single camera, ensuring accurate initial positioning of each image during stitching; correction feature points are used to determine the stitching boundary and relative position of adjacent images, and geometrically correct adjacent images in conjunction with the homography matrix to avoid image shifting or distortion after stitching.

[0045] Understandably, the spatial coordinates of each camera image are determined using stable feature points, the stitching relationship between adjacent images is determined using corrected feature points, and the adjacent images are geometrically corrected using a homography matrix. All images are stitched together in spatial coordinate order to form a complete parking lot map. The above process is repeated every 30 seconds to automatically correct map deviations caused by slight camera offsets.

[0046] In one feasible implementation, step S10 may include steps B11 to B13: Step B11: Obtain a first sampling frequency and a second sampling frequency, wherein the first sampling frequency is less than the second sampling frequency; It should be noted that the first acquisition frequency is the frame rate used to acquire images of the parking area when the parking space status remains unchanged. In this embodiment, the frequency is set to 1 frame per second. Its purpose is to minimize the resource consumption of image acquisition and the amount of subsequent data processing while meeting the real-time monitoring requirements of parking space status (delay not exceeding 1 second).

[0047] Additionally, the second acquisition frequency is the frame rate at which images of the parking area are acquired when a vehicle enters or leaves the parking space. In this embodiment, the frequency is set to 5 frames per second. Its function is to capture the dynamic process of a vehicle entering or leaving more accurately by increasing the frame rate during critical periods when the vehicle's status changes, thereby ensuring the accuracy of subsequent parking space status recognition.

[0048] Step B12: When it is detected that there is no change in the parking space status of the parking lot map, the parking area image is collected according to the first collection frequency to obtain parking lot image data. It should be noted that the parking area images are collected by parking space recognition cameras in the parking lot, covering the parking space areas within the parking lot. These cameras are selected according to the area and scene. For example, dense areas use 8mm focal length wide-angle cameras (100° field of view) installed at a height of 3 meters and a downward angle of 45°, while open areas use 25mm focal length telephoto cameras (35° field of view) installed at a height of 5 meters and a downward angle of 30° to ensure that the parking space situation in the parking area can be fully captured.

[0049] Understandably, real-time monitoring of parking space status on the parking lot map is used. When no change in parking space status is detected, the camera is controlled to collect images of the parking area at the first acquisition frequency (1 frame per second). The collected images are used to determine the parking lot image data, which satisfies the real-time requirement (delay of no more than 1 second) and avoids unnecessary high-frequency acquisition.

[0050] Step B13: When a vehicle is detected entering or leaving the parking space area of ​​the parking lot map, the parking area image is acquired according to the second acquisition frequency to obtain parking lot image data.

[0051] Understandably, when a vehicle is detected entering or leaving a parking space area on the parking lot map, the acquisition frequency is immediately switched, and the camera is controlled to acquire images of the parking area at the second acquisition frequency. This high-frequency acquisition state lasts for 3 seconds and then returns to the first acquisition frequency. The images acquired through this process are used to obtain parking lot image data, ensuring that the process of changes in parking space status when vehicles enter and exit is fully captured.

[0052] Step S20: Perform multi-level image compression processing on the parking lot image data to obtain processed image data; It should be noted that multi-level image compression processing is a staged compression operation on parking lot image data, including two stages: preprocessing compression and encoding compression. In the preprocessing compression stage, the region of interest of the parking lot image data is cropped, retaining only the effective area containing parking spaces. In the encoding compression stage, the cropped image is encoded using the H.265 Main10 encoding standard to further reduce the image data volume.

[0053] In addition, the image data processing involves multi-level image compression of parking lot image data, which significantly reduces the data volume compared to the original parking lot image data. For example, the image area is reduced by more than 60% after cropping, and 1080P images can be reduced to 10 kilobytes per frame after encoding and compression, which can effectively reduce resource consumption in subsequent data transmission and inference processes.

[0054] Understandably, the determined parking lot image data is first preprocessed and compressed. Invalid areas that do not contain parking spaces are removed by cropping the region of interest and the valid areas are retained. Then, the cropped image is encoded and compressed using the H.265 Main10 encoding standard. After completing multi-level image compression processing, the processed image data is obtained.

[0055] In one feasible implementation, step S20 may include steps S21 to S23: Step S21: Perform region cropping processing on the parking lot image data to obtain an effective image region; It should be noted that region cropping is a preprocessing operation performed on parking lot image data. Its core is to remove invalid areas in the parking lot image data that do not contain parking spaces, and only retain the image parts that are directly related to parking spaces. This operation can focus on the key image information required for parking space recognition and avoid invalid data from interfering with subsequent processing.

[0056] In addition, the effective image region is the image region obtained after region cropping. In this embodiment, the region only includes the parking space area in the parking lot. The cropped image area is reduced by more than 60% compared with the original parking lot image data, which greatly reduces the amount of data for subsequent image processing.

[0057] Understandably, for the established parking lot image data, invalid areas that do not involve parking spaces are removed through region cropping, such as blank areas in the parking lot passageway and wall areas, and only the image portion containing parking spaces is retained, thus obtaining the valid image area.

[0058] Step S22: Perform image compression processing on the effective image region to obtain a compressed image region; It should be noted that image compression is a further data optimization operation performed on the effective image area. By using specific algorithms, the storage volume of image data is reduced while retaining the parking space feature information in the effective image area. This reduces the burden on subsequent encoding processing and data transmission, ensuring that key features are not lost while improving data processing efficiency.

[0059] In addition, the compressed image region is the image region obtained after image compression processing of the effective image region. It retains the key features related to parking space recognition in the effective image region, while the data volume is further reduced compared to the effective image region, providing a lighter image foundation for subsequent encoding processing.

[0060] It is understandable that image compression processing is performed on the effective image area to reduce the amount of data in the effective image area while ensuring that the parking space feature information is completely preserved. After processing, a compressed image area is obtained.

[0061] Step S23: Encode the compressed image region to obtain processed image data.

[0062] It should be noted that the encoding process is an operation of encoding the compressed image area using a specific encoding standard. In this embodiment, the encoding standard used is H.265 Main10. Compared with the traditional H.264 encoding standard, this standard improves the compression rate by 50% and can further compress the 1080P resolution compressed image area to 10 kilobytes per frame, which greatly reduces the amount of image data.

[0063] In addition, image data processing involves encoding the compressed image region to obtain the final image data. This data is characterized by its small size and complete preservation of key features. It can be directly input into a preset image recognition model for lightweight edge-end inference, meeting the requirements for data transmission and inference efficiency.

[0064] Understandably, the compressed image region obtained in step S22 is encoded using the H.265 Main10 encoding standard, compressing the 1080P resolution compressed image region to 10 kilobytes per frame, and finally obtaining the processed image data, which prepares the data for subsequent lightweight inference at the edge.

[0065] Step S30: Input the processed image data into a preset image recognition model for edge-end lightweight inference to obtain the initial parking space recognition result and the confidence level of the recognition result. The preset image recognition model includes a depth-separable convolutional layer and a pyramid compression layer. It should be noted that the preset image recognition model is a deep learning model optimized for parking space recognition in parking lots. Its basic architecture is based on the YOLOv8 model. It achieves lightweight design by integrating depthwise separable convolutional layers and pyramid compression layers, and can be adapted to the edge computing chips built into the camera, such as the ARM Cortex-A76 with a computing power of 2 TOPS, thereby reducing the computing power requirement while ensuring recognition accuracy.

[0066] In addition, the depth-separable convolutional layer is a core component in the backbone network of the preset image recognition model. This component splits the traditional 3×3 convolution into two steps: depthwise convolution and pointwise convolution. Compared with traditional convolution, it can significantly reduce the number of model parameters and computation, reducing the number of model parameters by 65% ​​and increasing the inference speed by 40%, thus meeting the lightweight requirements of edge computing.

[0067] Furthermore, the pyramid compression layer is a component in the neck network of the pre-defined image recognition model. Its function is to compress the number of channels in the multi-scale feature maps extracted by the model, halving the number of channels. This reduces the computational load while retaining key semantic information, retaining only the feature maps of two key scales for multi-scale feature fusion, and further optimizing the lightweight effect of the model.

[0068] In addition, the confidence score of the recognition result is a value generated along with the initial parking space recognition result. It is used to represent the reliability of the initial parking space recognition result. This value is calculated by the preset image recognition model based on factors such as the clarity and completeness of the parking space features in the processed image data, and provides a basis for subsequent screening of accurate parking space recognition results.

[0069] Understandably, the processed image data is input into a preset image recognition model, which initiates lightweight inference at the edge on the edge computing chip built into the camera. During inference, the model first extracts parking space features from the processed image data through a depthwise separable convolutional layer, then compresses the feature map and fuses multi-scale features through a pyramid compression layer, judges the status of each parking space based on the fused features, and finally outputs the initial parking space recognition result and the corresponding recognition result confidence level.

[0070] In one feasible implementation, step S30 may include steps S31 to S33: Step S31: Input the processed image data into a depth-separable convolutional layer for convolution to obtain convolutional image data; It should be noted that the depthwise separable convolutional layer is a core component in the backbone network of the preset image recognition model. This component splits the traditional 3×3 convolution into two independent steps: depthwise convolution and pointwise convolution. While completing image feature extraction, it can significantly reduce the number of model parameters and computational load. In this embodiment, this component can reduce the number of model parameters by 65% ​​and increase the inference speed by 40%. It is compatible with low-computing-power edge computing chips built into cameras, such as ARM Cortex-A76 with a computing power of 2 TOPS.

[0071] Additionally, convolutional image data is image data obtained by convolving processed image data through depth-separable convolutional layers. This data retains key features related to parking space recognition in the processed image data, and due to the optimization of the convolution process, the amount of data and the computational load of subsequent processing are further reduced compared to processed image data.

[0072] It is understandable that the image data to be processed is input into a depthwise separable convolutional layer, which then performs depthwise convolution and pointwise convolution operations in sequence to extract parking space features from the processed image data, and finally obtains convolutional image data.

[0073] Step S32: Input the convolutional image data into the pyramid compression layer for edge-end lightweight inference to obtain the initial recognition result; It should be noted that the pyramid compression layer is a component in the neck network of the preset image recognition model. Its function is to compress the number of channels of the multi-scale feature maps corresponding to the convolutional image data, halve the number of channels and retain only the feature maps of the two key scales for fusion, thereby further reducing the amount of computation while retaining key semantic information and adapting to the computing power limitations of the edge.

[0074] In addition, the lightweight edge inference is an inference process that runs on the edge computing chip built into the camera. This process does not rely on a lot of computing power in the cloud and can quickly process convolutional image data. In this embodiment, it can achieve millisecond-level response with a latency of no more than 100 milliseconds, meeting the real-time requirements of parking space recognition.

[0075] Furthermore, the initial identification result is a preliminary judgment of the parking space status obtained through lightweight inference at the edge, which includes information on whether the parking space may be vacant or occupied. However, it has not yet been optimized by loss calculation and its accuracy needs to be further verified in subsequent steps.

[0076] Understandably, convolutional image data is input into a pyramid compression layer, and after feature map compression and multi-scale fusion, lightweight inference is performed on an edge computing chip to finally obtain the initial recognition result.

[0077] Step S33: Perform category loss calculation and localization loss calculation on the initial recognition result to obtain the initial parking space recognition result and the confidence level of the recognition result.

[0078] It should be noted that the category loss calculation is an operation that optimizes the initial recognition results using an improved GFL loss. This operation introduces a parking space status weight factor, setting the weight of vacant parking spaces to 1.2 and the weight of occupied parking spaces to 0.8, in order to solve the model bias problem caused by the low proportion of vacant parking spaces in the parking lot and improve the accuracy of category judgment.

[0079] In addition, the positioning loss calculation is an operation that optimizes the parking space frame position in the initial recognition result by combining the DFL loss and the parking space geometric constraints. The parking space geometric constraints limit the aspect ratio of the rectangular frame to between 1.5 and 2.5. This operation can reduce the parking space frame positioning error to no more than 5 pixels, thereby improving the accuracy of parking space position judgment.

[0080] Furthermore, the confidence score of the recognition result is a value generated along with the initial parking space recognition result. It is used to represent the reliability of the initial parking space recognition result. The higher the value, the higher the matching degree between the initial parking space recognition result and the actual parking space status.

[0081] Understandably, category loss calculation and localization loss calculation are performed on the initial recognition results to correct the category judgment bias and parking space box localization deviation, and finally the initial parking space recognition result and the corresponding recognition result confidence are obtained.

[0082] Step S40: Based on the initial parking space recognition result and the confidence level of the recognition result, parking space recognition is performed to obtain the target parking space recognition result.

[0083] It should be noted that the target parking space recognition result is the final parking space status judgment result obtained after filtering and verifying the initial parking space recognition result. This result can eliminate the parts with low reliability in the initial parking space recognition result, ensuring that the output parking space status information is accurate and meets the practical needs of parking lot management.

[0084] Understandably, after obtaining the initial parking space recognition result and the corresponding recognition result confidence level, a confidence level threshold is set, and the recognition result confidence level is compared with the threshold. If the recognition result confidence level is higher than or equal to the threshold, the corresponding initial parking space recognition result is directly determined as the target parking space recognition result. If the recognition result confidence level is lower than the threshold, the initial parking space recognition result is further verified, and the accurate target parking space recognition result is finally determined.

[0085] In one feasible implementation, step S40 may include steps S41 to S45: Step S41: Obtain the preset confidence threshold; It should be noted that the preset reliability threshold is a pre-set numerical standard used to judge the reliability of the initial parking space recognition result. In this embodiment, the threshold is set to 0.7. Its function is to distinguish the reliability of the initial parking space recognition result. The result is considered to be highly reliable if it is higher than or equal to the threshold, while the result is lower than the threshold and needs to be further verified, so as to provide a basis for selecting different processing methods in the future.

[0086] Step S42: When the confidence level of the recognition result is greater than or equal to the preset confidence threshold, the initial parking space recognition result is taken as the target parking space recognition result. Understandably, the confidence level of the recognition result is compared with the preset confidence threshold. If the confidence level of the recognition result is ≥0.7, it means that the initial parking space recognition result is highly reliable, and the initial parking space recognition result is directly determined as the target parking space recognition result.

[0087] Step S43: When the confidence level of the recognition result is less than the preset confidence threshold, obtain the neighboring frame image of the image corresponding to the initial parking space recognition result; It should be noted that the neighboring frame image is an image that is temporally adjacent to the image corresponding to the initial parking space recognition result. In this embodiment, it specifically refers to the two frames preceding the image corresponding to the initial parking space recognition result. These images contain the state information of the parking space area within a similar time period and can be used to supplement dynamic details that may be missing during the initial recognition, such as information in scenarios like partial vehicle occlusion or temporary changes in lighting.

[0088] Understandably, when the confidence level of the recognition result is less than 0.7, the first two frames of the image corresponding to the initial parking space recognition result, i.e., the neighboring frame images, are retrieved from the image storage record to prepare for subsequent supplementary semantic features.

[0089] Step S44: Supplement the initial parking space recognition result with complete semantic features based on the neighboring frame images to obtain a supplementary recognition result; It should be noted that the complete semantic feature supplementation is an operation that uses parking space area information in neighboring frame images to make up for the lack of semantic features in the initial parking space recognition result caused by lightweight inference at the edge. For example, if only part of the parking space outline is captured during the initial recognition due to temporary occlusion by the vehicle, the complete semantic features of the parking space are supplemented by the outline information of the parking space when it is not occluded in the neighboring frame images, so that the recognition result is more comprehensive and accurate.

[0090] Understandably, by analyzing the status information of the parking spaces corresponding to the initial parking space recognition results in neighboring frame images, extracting the missing semantic features and supplementing them into the initial parking space recognition results, a supplementary recognition result is finally obtained.

[0091] Step S45: Upload the supplementary identification results to the cloud so that the cloud can review the supplementary identification results and provide feedback on the target parking space identification results.

[0092] It should be noted that the cloud server is a server with strong computing power and a complete image recognition model. The complete model it carries is based on the YOLOv8 architecture and can perform high-precision verification by combining temporal comparison and complete semantic analysis. During the verification, the supplementary recognition results are combined with the image information of the adjacent frames to determine the real status of the parking space and ensure the accuracy of the verification results.

[0093] Understandably, the supplementary recognition results are uploaded to the cloud via the cellular network. The cloud then calls the full version of the image recognition model, combines the supplementary recognition results with information from neighboring frames for verification, and after determining the actual status of the parking space, feeds back the target parking space recognition result to the edge.

[0094] This embodiment provides a lightweight parking space recognition method. By employing techniques such as multi-level compression processing of parking lot images, edge-end lightweight inference of depth-separable convolutional layers and pyramid compression layers, and collaborative optimization of high-precision cloud models, it solves the technical problems of high hardware deployment costs, large traffic consumption, insufficient recognition accuracy, and poor system linkage in intelligent parking space recognition. It achieves the beneficial effects of reducing deployment costs, reducing operating traffic, improving recognition accuracy, and enhancing system linkage.

[0095] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 Following step S40, the lightweight parking space identification method further includes steps S50-S70: Step S50: Determine whether there is an abnormal scene in the parking lot based on the target parking space recognition result. The abnormal scene includes vehicles parking beyond the boundary, parking spaces being blocked by debris, or the camera acquisition device being offset. It should be noted that abnormal scenarios refer to scenarios in the parking lot that deviate from the normal usage status of parking spaces or affect the accuracy of parking space recognition. Specifically, these include three types: vehicles parked beyond the boundaries, parking spaces being obstructed by debris, and camera acquisition devices being offset. These scenarios will interfere with the normal judgment of parking space status and need to be further investigated and confirmed through the target parking space recognition results.

[0096] Additionally, parking beyond the designated parking space refers to a situation where a vehicle is not completely parked within the designated parking space area, with part of the vehicle body extending beyond the parking space line. The existence of this scenario can be determined by the relative positional relationship between the vehicle outline and the parking space line in the target parking space recognition results.

[0097] Furthermore, a parking space being obscured by debris refers to a situation where there is a non-vehicle object in the parking space and the area of ​​the object exceeds 0.5 square meters. By analyzing the contour features of the parking space area in the target parking space recognition results, if a non-vehicle contour is detected and the area meets the above criteria, it can be determined as this scenario.

[0098] Additionally, camera acquisition device offset refers to a deviation in the position or angle of the camera used to acquire images of the parking area, resulting in inconsistencies between the acquired image features and the features calibrated when the device was started. By comparing the feature point positions of the image corresponding to the target parking space recognition result with the calibrated image, it can be determined whether this scenario exists.

[0099] Understandably, the system analyzes the parking space status and image features corresponding to the target parking space recognition results to determine whether the vehicle in the parking space exceeds the parking space line to confirm whether there is a vehicle parked beyond the boundary, detects whether there are non-vehicle objects with an area greater than 0.5 square meters in the parking space to confirm whether the parking space is blocked by debris, and compares the feature point positions of the current image and the calibration image to confirm whether there is a camera acquisition device offset, thereby determining whether there is an abnormal scene in the parking lot.

[0100] Step S60: When an abnormal scene exists in the parking lot, determine the warning level of the abnormal scene and obtain the warning level identification result; It should be noted that the warning level is a classification based on the degree of impact of abnormal scenarios on parking lot management and parking space recognition. In this embodiment, it is divided into two categories: Level 1 warning and Level 2 warning. Different levels correspond to different processing priorities in order to efficiently allocate management resources.

[0101] In addition, the warning level identification result refers to the specific warning level determined according to the type of abnormal scenario. This result clarifies the processing priority of abnormal scenarios and provides a basis for subsequent targeted abnormal resolution measures.

[0102] Understandably, once an abnormal scene is identified in the parking lot, the warning level is determined according to the type of abnormal scene: if the abnormal scene is a vehicle parked beyond the boundary or the parking space is blocked by debris, it is determined to be a level one warning; if the abnormal scene is the camera acquisition device being offset, it is determined to be a level two warning, and the warning level identification result is finally obtained.

[0103] Step S70: Determine the target anomaly based on the warning level identification result, and resolve the anomaly based on the target anomaly to identify the parking space.

[0104] It should be noted that target anomalies are specific abnormal situations identified based on the warning level. Level 1 and Level 2 warnings each contain different types. Level 1 warnings correspond to target anomalies such as vehicles parking outside designated areas or parking spaces being obstructed by debris. Level 2 warnings correspond to target anomalies such as camera acquisition device misalignment. These anomalies can affect the accuracy of parking space identification or the order of parking lot management.

[0105] Furthermore, anomaly resolution refers to targeted handling measures taken based on the warning level identification results. Different warning levels correspond to different resolution methods, with the aim of eliminating the interference of abnormal scenarios on parking space recognition and ensuring the continuous and accurate operation of parking space recognition.

[0106] Understandably, the specific type of target anomaly is first determined based on the warning level identification results. If it is a Level 1 warning, corresponding to a vehicle parked beyond the boundary or a parking space being obstructed by debris, the anomaly information, on-site images, and parking space number are pushed to the mobile application of the management personnel. If it is a Level 2 warning, corresponding to a camera acquisition device offset, an SMS notification is automatically triggered and a maintenance work order is generated, marking the equipment location and fault type. After the management personnel or maintenance personnel have dealt with the target anomaly, the normal parking space identification process is restored.

[0107] This embodiment provides a lightweight parking space identification method. By adding abnormal scene detection, early warning level classification, and targeted anomaly resolution techniques to the target parking space identification results, it solves the technical problems of insufficient abnormal scene identification, untimely response, and low processing efficiency in parking lot management. It achieves the beneficial effects of improving the initiative of parking lot management, reducing the workload of manual inspection, and improving the management response speed.

[0108] For example, to help understand the implementation process of the lightweight parking space recognition method obtained by combining this embodiment with the above embodiment one, please refer to... Figure 3 , Figure 3 A simplified flowchart of a lightweight parking space identification method is provided, specifically: The left side describes the cloud server's workflow, while the right side describes the edge processing workflow. The cloud server first receives data from the edge via a cloud receiving module. This data is then transmitted encrypted over a cellular network to a high-precision feature optimization module. This module supplements semantic features through historical frame time-series comparison. Next, the data enters the cloud-based high-precision model, a full version of YOLOv8+ semantic distillation. Then, it enters the cloud-based decision module, which uses statistical model samples, such as occlusion / lighting abrupt changes, for model training and updates. This module uses mathematical expectation samples to detect the model, performs knowledge distillation to generate lightweight parameters, and pushes these parameters to the edge via an OTA model push module. A global sample library periodically replenishes the training with multi-scene samples, and the results are transmitted back to the edge, feeding back to the edge display platform. The edge processing module first performs 540+640 pixel normalization and ROI cropping, then enters the lightweight YOLOv8 backbone network. This network replaces the 3x3 ordinary convolution with a core improved depthwise separable convolution. After CSP Bottleneck lightweighting, the number of parameters is reduced by 65% ​​and the inference speed is improved by 40%. Next, it enters the lightweight neck network, which retains key semantics by halving the number of channels in the feature pyramid compression layer. Then, it enters the multi-scale feature fusion, retaining only three scales. Next, it enters the edge inference decision module, which encapsulates abnormal results when the confidence level is less than 0.7, including the image features of the current frame and the previous two frames. Then, it enters the cellular network, which transmits data through AG / 50. When the confidence level is greater than 0.7, it outputs the parking space status locally and pushes the encrypted lightweight parameters to the edge model management module. This diagram combines the technical solutions mentioned in the disclosure document, including hardware integration and deployment solutions, traffic-optimized data acquisition and transmission solutions, lightweight high-precision deep learning model solutions, baseline-free dynamic map stitching and management platform solutions, and abnormal state intelligent early warning solutions. Through the collaborative work of the cloud and edge terminals, it achieves high precision and high efficiency in intelligent identification of parking space vacancies.

[0109] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the lightweight parking space identification method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0110] This application also provides a lightweight parking space recognition device, please refer to... Figure 4The lightweight parking space recognition device includes: The data acquisition module 10 is used to stitch together the images captured by the camera to obtain a parking lot map, and to determine the parking lot image data based on the parking lot map. Image preprocessing module 20 is used to perform multi-level image compression processing on the parking lot image data to obtain processed image data; The edge inference module 30 is used to input the processed image data into a preset image recognition model for edge-end lightweight inference to obtain an initial parking space recognition result and a recognition result confidence level. The preset image recognition model includes a depth-separable convolutional layer and a pyramid compression layer. The parking space recognition module 40 is used to recognize parking spaces based on the initial parking space recognition result and the confidence level of the recognition result, and obtain the target parking space recognition result.

[0111] The lightweight parking space identification device provided in this application, employing the lightweight parking space identification method described in the above embodiments, can solve the technical problem of accurately identifying the occupancy status of parking spaces in parking lots with complex parking space distribution, frequent environmental changes, and limited communication and deployment conditions. Compared with the prior art, the beneficial effects of the lightweight parking space identification device provided in this application are the same as those of the lightweight parking space identification method provided in the above embodiments, and other technical features in the lightweight parking space identification device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0112] In one embodiment, the parking space recognition module 40 is further configured to obtain a preset confidence threshold. When the confidence level of the recognition result is less than the preset confidence threshold, obtain the neighboring frame image of the image corresponding to the initial parking space recognition result; The initial parking space recognition result is supplemented with complete semantic features based on the neighboring frame images to obtain a supplementary recognition result; The supplementary identification results are uploaded to the cloud so that the cloud can review them and provide feedback on the target parking space identification results.

[0113] In one embodiment, the parking space recognition module 40 is further configured to determine whether there is an abnormal scene in the parking lot based on the target parking space recognition result, wherein the abnormal scene includes vehicles parking beyond the boundary, parking spaces being blocked by debris, or the camera acquisition device being offset. When an abnormal scenario exists in the parking lot, the warning level of the abnormal scenario is determined, and the warning level identification result is obtained; Based on the warning level identification result, the target anomaly is determined, and the anomaly is resolved in order to identify the parking space.

[0114] In one embodiment, the edge inference module 30 is further configured to input the processed image data into a depth-separable convolutional layer for convolution to obtain convolutional image data; The convolutional image data is input into the pyramid compression layer for edge-end lightweight inference to obtain the initial recognition result; The initial identification results are subjected to category loss calculation and localization loss calculation to obtain the initial parking space identification result and the confidence level of the identification result.

[0115] In one embodiment, the data acquisition module 10 is further configured to extract stable feature points of the camera-captured image according to a scale-invariant feature transformation strategy. The adjacent images captured by the camera are input into the fast nearest neighbor search packet processor to obtain overlapping feature points; Mismatched points among the overlapping feature points are deleted according to the random sampling consistency strategy to obtain corrected feature points; The parking lot map is determined based on the stable feature points and the corrected feature points.

[0116] In one embodiment, the data acquisition module 10 is further configured to acquire a first acquisition frequency and a second acquisition frequency, wherein the first acquisition frequency is less than the second acquisition frequency; When the parking space status of the parking lot map is not changed, the parking area image is collected according to the first collection frequency to obtain parking lot image data. When a vehicle is detected entering or leaving a parking space area on the parking lot map, parking area images are acquired according to a second acquisition frequency to obtain parking lot image data.

[0117] In one embodiment, the image preprocessing module 20 is further configured to perform region cropping processing on the parking lot image data to obtain an effective image region; The effective image region is subjected to image compression processing to obtain a compressed image region; The compressed image region is encoded to obtain processed image data.

[0118] This application provides a lightweight parking space recognition device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the lightweight parking space recognition method in Embodiment 1 above.

[0119] The following is for reference. Figure 5This document illustrates a structural schematic diagram of a lightweight parking space recognition device suitable for implementing embodiments of this application. The lightweight parking space recognition device in this application embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The lightweight parking space recognition device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0120] like Figure 5 As shown, the lightweight parking space recognition device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM (Read Only Memory) 1002 or a program loaded from storage device 1003 into RAM (Random Access Memory) 1004. RAM 1004 also stores various programs and data required for the operation of the lightweight parking space recognition device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the lightweight parking space recognition device to communicate wirelessly or wiredly with other devices to exchange data. Although a lightweight parking space recognition device with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0121] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0122] The lightweight parking space recognition device provided in this application, employing the lightweight parking space recognition method described in the above embodiments, solves the technical problem of accurately identifying the occupancy status of parking spaces in parking lots with complex parking space distribution, frequent environmental changes, and limited communication and deployment conditions. Compared with the prior art, the beneficial effects of the lightweight parking space recognition device provided in this application are the same as those of the lightweight parking space recognition method provided in the above embodiments, and other technical features of this lightweight parking space recognition device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0123] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0124] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0125] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the lightweight parking space identification method in the above embodiments.

[0126] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), Erasable Programmable Read Only Memory (EPROM), optical fiber, CD-ROM (CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0127] The aforementioned computer-readable storage medium may be included in a lightweight parking space recognition device; or it may exist independently and not assembled into a lightweight parking space recognition device.

[0128] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the lightweight parking space recognition device, the lightweight parking space recognition device performs the following actions: stitching together images captured by a camera to obtain a parking map, and determining parking image data based on the parking map; performing multi-level image compression processing on the parking image data to obtain processed image data; inputting the processed image data into a preset image recognition model for edge-end lightweight inference to obtain an initial parking space recognition result and a recognition result confidence level, wherein the preset image recognition model includes depthwise separable convolutional layers and pyramid compression layers; and performing parking space recognition based on the initial parking space recognition result and the recognition result confidence level to obtain a target parking space recognition result.

[0129] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0130] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0131] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0132] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned lightweight parking space identification method. This solves the technical problem of accurately identifying the occupancy status of parking spaces in parking lots with complex parking space distribution, frequent environmental changes, and limited communication and deployment conditions. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the lightweight parking space identification method provided in the above embodiments, and will not be repeated here.

[0133] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the lightweight parking space recognition method described above.

[0134] The computer program product provided in this application can solve the technical problem of accurately identifying the occupancy status of parking spaces in parking lots with complex parking space distribution, frequent environmental changes, and limited communication and deployment conditions. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the lightweight parking space identification method provided in the above embodiments, and will not be repeated here.

[0135] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A lightweight parking space identification method, characterized in that, The method includes: Images captured by the camera are stitched together to obtain a parking lot map, and parking lot image data is determined based on the parking lot map; The parking lot image data is subjected to multi-level image compression processing to obtain processed image data; The processed image data is input into a preset image recognition model for edge-end lightweight inference to obtain initial parking space recognition results and recognition result confidence. The preset image recognition model includes depth-separable convolutional layers and pyramid compression layers. Based on the initial parking space identification result and the confidence level of the identification result, parking space identification is performed to obtain the target parking space identification result.

2. The method as described in claim 1, characterized in that, The step of identifying parking spaces based on the initial parking space identification result and the confidence level of the identification result to obtain the target parking space identification result includes: Obtain the preset reliability threshold; When the confidence level of the recognition result is less than the preset confidence threshold, obtain the neighboring frame image of the image corresponding to the initial parking space recognition result; The initial parking space recognition result is supplemented with complete semantic features based on the neighboring frame images to obtain a supplementary recognition result; The supplementary identification results are uploaded to the cloud so that the cloud can review them and provide feedback on the target parking space identification results.

3. The method as described in claim 1, characterized in that, After the step of identifying parking spaces based on the initial parking space identification result and the confidence level of the identification result to obtain the target parking space identification result, the method further includes: Based on the target parking space recognition results, it is determined whether there are any abnormal scenarios in the parking lot. The abnormal scenarios include vehicles parking beyond the boundaries, parking spaces being obstructed by debris, or the camera acquisition device being offset. When an abnormal scenario exists in the parking lot, the warning level of the abnormal scenario is determined, and the warning level identification result is obtained; Based on the warning level identification result, the target anomaly is determined, and the anomaly is resolved in order to identify the parking space.

4. The method as described in claim 1, characterized in that, The step of inputting the processed image data into a preset image recognition model for edge-end lightweight inference to obtain an initial parking space recognition result and a recognition result confidence level, wherein the preset image recognition model includes depthwise separable convolutional layers and pyramid compression layers, includes: The processed image data is input into a depth-separable convolutional layer for convolution to obtain convolutional image data. The convolutional image data is input into the pyramid compression layer for edge-end lightweight inference to obtain the initial recognition result; The initial identification results are subjected to category loss calculation and localization loss calculation to obtain the initial parking space identification result and the confidence level of the identification result.

5. The method as described in claim 1, characterized in that, The steps for stitching together images captured by cameras to obtain a parking lot map include: Stable feature points are extracted from images captured by the camera based on a scale-invariant feature transformation strategy; The adjacent images captured by the camera are input into the fast nearest neighbor search packet processor to obtain overlapping feature points; Mismatched points among the overlapping feature points are deleted according to the random sampling consistency strategy to obtain corrected feature points; The parking lot map is determined based on the stable feature points and the corrected feature points.

6. The method as described in claim 1, characterized in that, The steps for determining parking lot image data based on the parking lot map include: Obtain a first sampling frequency and a second sampling frequency, wherein the first sampling frequency is less than the second sampling frequency; When the parking space status of the parking lot map is not changed, the parking area image is collected according to the first collection frequency to obtain parking lot image data. When a vehicle is detected entering or leaving a parking space area on the parking lot map, parking area images are acquired according to a second acquisition frequency to obtain parking lot image data.

7. The method as described in claim 1, characterized in that, The step of performing multi-level image compression processing on the parking lot image data to obtain processed image data includes: The parking lot image data is cropped to obtain an effective image region; The effective image region is subjected to image compression processing to obtain a compressed image region; The compressed image region is encoded to obtain processed image data.

8. A lightweight parking space recognition device, characterized in that, The device includes: The data acquisition module is used to stitch together images captured by the camera to obtain a parking lot map, and to determine parking lot image data based on the parking lot map; The image preprocessing module is used to perform multi-level image compression processing on the parking lot image data to obtain processed image data; The edge inference module is used to input the processed image data into a preset image recognition model for lightweight edge inference to obtain the initial parking space recognition result and the confidence level of the recognition result. The preset image recognition model includes a depth-separable convolutional layer and a pyramid compression layer. The parking space recognition module is used to recognize parking spaces based on the initial parking space recognition result and the confidence level of the recognition result, and obtain the target parking space recognition result.

9. A lightweight parking space recognition device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the lightweight parking space identification method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the lightweight parking space identification method as described in any one of claims 1 to 7.

Images