A vehicle visual recognition method and system for unmanned parking lot

Abstract

The present application relates to the technical field of computer vision and image processing, in particular to a vehicle visual recognition method and system for unmanned parking lot. The technical problem of unstable quality of HDR image generated by the existing method is solved. The method comprises the following steps: obtaining a normal image of a vehicle collected by a high dynamic range camera at a target layout position; determining a first characteristic value and a second characteristic value of the vehicle based on the normal image; determining a shooting parameter of the high dynamic range camera based on the first characteristic value and the second characteristic value; instructing the high dynamic range camera to collect a long-exposure image and a short-exposure image of the vehicle based on the shooting parameter; and performing synthesis processing on the long-exposure image and the short-exposure image to determine a high dynamic range synthesis image for license plate recognition. The present application is used in the vehicle visual recognition scene of parking lot.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, specifically to a vehicle visual recognition method and system for unmanned parking lots. Background Technology

[0002] With the rapid development of smart cities and IoT technologies, unmanned parking systems are becoming increasingly common. Their efficient operation heavily relies on the automatic and accurate recognition of license plate information for vehicles entering and exiting. In such scenarios, visual recognition systems face extremely complex and variable lighting environments, such as backlighting in entrance and exit lanes, variations in natural light intensity at different times of day, and interference from artificial light sources like vehicle headlights. These factors can easily lead to localized overexposure or underexposure in the acquired license plate images, becoming a key bottleneck restricting recognition accuracy and system reliability. To address this challenge, the industry has widely adopted high dynamic range imaging technology. This technology aims to expand the dynamic range of images by synthesizing images with different exposures, thereby preserving details in both bright and dark areas of the scene. However, existing application solutions mostly focus on the implementation of High Dynamic Range (HDR) imaging technology itself, using preset and fixed exposure strategies to drive HDR image acquisition. They lack a perception and feedback mechanism for real-time scene lighting characteristics, resulting in unstable HDR image quality under extreme or rapidly changing lighting conditions. The display effect of the license plate area is inconsistent, making it difficult to maintain the performance of subsequent recognition algorithms at an ideal level. The environmental adaptability and robustness of the overall system need to be improved. Summary of the Invention

[0003] To address the technical problem of existing methods lacking a perception and feedback mechanism for real-time scene lighting characteristics, resulting in unstable HDR image quality under extreme or rapidly changing lighting conditions, the present invention aims to provide a vehicle visual recognition method and system for unmanned parking lots. The specific technical solution adopted is as follows:

[0004] In a first aspect, the present invention provides a vehicle visual recognition method for unmanned parking lots. The method includes: acquiring a normal image of a vehicle captured by a high dynamic range (HMR) camera at a target location; the normal image is an image captured by the HMR camera after disabling dynamic range synthesis; determining a first feature value and a second feature value of the vehicle based on the normal image; the first feature value is used to characterize the grayscale information of the license plate area; the second feature value is used to characterize the pixel distribution range in the normal image whose grayscale is similar to that of the license plate area; determining the shooting parameters of the HMR camera based on the first feature value and the second feature value; instructing the HMR camera to acquire long-exposure images and short-exposure images of the vehicle based on the shooting parameters; and performing synthesis processing on the long-exposure images and short-exposure images to determine a high dynamic range synthesized image for license plate recognition.

[0005] In conjunction with the first aspect mentioned above, in one possible implementation, the method further includes: acquiring multiple historical vehicle images collected at multiple candidate deployment locations at the entrance and exit of the parking lot during a preset historical monitoring period; calculating an exposure evaluation index for each candidate deployment location based on the grayscale performance of the license plate area at multiple candidate deployment locations in the historical vehicle images; the exposure evaluation index is used to characterize the risk of overexposure in the vehicle license plate area; and determining the candidate deployment location with the smallest exposure evaluation index as the target deployment location.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, the method specifically includes: determining the ratio of the number of target images to the total number of images among multiple historical vehicle general images; the target image being a historical vehicle image whose average gray level of the license plate area exceeds a preset first gray level threshold among multiple historical vehicle general images; determining the mean of the sum of the average gray levels of the license plate area among multiple historical vehicle general images; and determining an exposure evaluation index based on the ratio and the mean of the sum.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the method further includes: processing multiple historical vehicle images based on a preset deep learning detection model to determine the historical license plate region in each historical vehicle image.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the method further includes: determining historical grayscale statistical features of multiple historical vehicle general images based on the pixel grayscale values ​​of the historical license plate region of each historical vehicle general image; the historical grayscale statistical features include: the maximum average grayscale of the historical license plate region, the maximum grayscale dynamic range of the historical vehicle general image, and the average grayscale information of the historical vehicle general image.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the method specifically includes: processing a normal image based on a preset deep learning detection model to determine the license plate region in the normal image; determining a first feature value based on the gray values ​​of all pixels within the license plate region; and determining a second feature value based on the total number of pixels in the normal image whose gray values ​​differ from the first feature value by a preset gray tolerance threshold.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the method specifically includes: determining a short exposure duration compression coefficient based on a first feature value, a second feature value, and the maximum average gray level of the historical license plate area; and adjusting a preset short exposure reference duration based on the short exposure duration compression coefficient to determine the target short exposure duration.

[0011] In conjunction with the first aspect mentioned above, in one possible implementation, the shooting parameters further include: a target exposure ratio. The method further includes: determining a third feature value of a normal image; the third feature value is the difference between the maximum and minimum pixel grayscale values ​​in the normal image; determining an exposure ratio adjustment coefficient based on the third feature value, the maximum grayscale dynamic range of historical vehicle normal images, and the target short exposure duration; and adjusting a preset exposure ratio reference value based on the exposure ratio adjustment coefficient to determine the target exposure ratio.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, the method further includes: determining a fourth feature value of a normal image; the fourth feature value is the average grayscale value of all pixels in the normal image; determining a supplementary lighting necessity assessment value based on the first feature value, the fourth feature value, and the average grayscale information of historical vehicle normal images; if the supplementary lighting necessity assessment value is greater than a preset supplementary lighting threshold, determining to supplement the license plate area of ​​the vehicle.

[0013] Secondly, the present invention provides a vehicle vision recognition system for unmanned parking lots. The system includes: an image acquisition module for acquiring ordinary images of a vehicle captured by a high dynamic range (HDR) camera at a target location; the ordinary image is an image acquired after the HDR camera's dynamic range synthesis function is disabled; a feature value determination module for determining a first feature value and a second feature value of the vehicle based on the ordinary image; the first feature value characterizes the grayscale information of the license plate area; the second feature value characterizes the distribution range of pixels in the ordinary image whose grayscale is similar to that of the license plate area; a parameter decision module for determining the shooting parameters of the HDR camera based on the first and second feature values; a shooting control module for instructing the HDR camera to acquire long-exposure and short-exposure images of the vehicle based on the shooting parameters; and an image synthesis module for synthesizing the long-exposure and short-exposure images to determine a high dynamic range synthesized image for license plate recognition.

[0014] The present invention has the following beneficial effects:

[0015] This invention first uses a high dynamic range (HDR) camera to acquire a single frame of ordinary image with the synthesis function disabled, and analyzes the grayscale distribution characteristics of the license plate and similar grayscale regions. Then, based on this analysis, it dynamically determines the specific parameters for the camera to capture multiple frames of synthesized images. Finally, it controls the camera to acquire and synthesize HDR images according to these parameters for recognition. This transforms the use of HDR cameras from a traditional, fixed-parameter image acquisition mode to an intelligent closed-loop control system based on real-time scene perception. By first determining the current lighting and license plate status reflected in the ordinary image, and then generating the optimal shooting command accordingly, it fundamentally overcomes the problem of overexposure or underexposure of the license plate area caused by drastic changes in lighting. This results in a stable output of license plate images with clear details and appropriate contrast, greatly improving the accuracy of license plate recognition and the overall robustness of the system under complex lighting conditions in all weather conditions. This solves the technical problem of existing methods lacking a perception and feedback mechanism for real-time scene lighting characteristics, leading to unstable HDR image quality under extreme or rapidly changing lighting conditions. Attached Figure Description

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a vehicle visual recognition method for unmanned parking lots, provided as an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of a vehicle vision recognition system architecture for an unmanned parking lot, provided as an embodiment of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a vehicle visual recognition method and system for unmanned parking lots proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] The following description, in conjunction with the accompanying drawings, details a specific scheme for a vehicle visual recognition method and system for unmanned parking lots provided by the present invention.

[0022] Please see Figure 1 The diagram illustrates a vehicle visual recognition method and system flowchart for unmanned parking lots according to an embodiment of the present invention. The method includes the following steps S101-S105, which will be described in detail below.

[0023] S101. Acquire ordinary images of the vehicle captured by the high dynamic range camera at the target deployment location.

[0024] Among them, the ordinary image is the image captured by the high dynamic range camera after the dynamic range synthesis function is turned off.

[0025] In one possible implementation, when a vehicle enters the recognition area at the parking lot entrance / exit, an image is captured by a high dynamic range (HDL) camera deployed at the target location. During the acquisition process, the HDL camera is controlled to operate in a single-exposure conventional shooting mode, i.e., its dynamic range synthesis function is disabled, to obtain a single frame of ordinary image. The target deployment location is determined based on the analysis and optimization of historical ordinary vehicle images captured at multiple candidate deployment locations during historical monitoring periods. Historical data shows that this location has the lowest risk of causing overexposure of the license plate area.

[0026] S102. Based on the ordinary image, determine the first feature value and the second feature value of the vehicle.

[0027] The first feature value is used to characterize the grayscale information of the license plate area; the second feature value is used to characterize the distribution range of pixels in the ordinary image whose grayscale is similar to that of the license plate area.

[0028] In one possible implementation, the ordinary image is parsed to extract two key quantitative indicators used to quantitatively characterize the illumination state of the license plate area in the current scene. First, the ordinary image is processed using the same pre-set deep learning detection model as used for processing historical images (exemplarily, the pre-set deep learning detection model is the YOLO model) to identify and select the license plate area in the image, thus obtaining the license plate area in the ordinary image. Based on this license plate area, a first feature value is determined by statistically analyzing the grayscale values ​​of all pixels within the box, representing the grayscale level of the license plate area itself. Then, using this first feature value as a benchmark, statistics are performed across the entire ordinary image to determine a second feature value representing the distribution range of pixels with similar grayscale values ​​to the license plate area in the image.

[0029] S103. Based on the first feature value and the second feature value, determine the shooting parameters of the high dynamic range camera.

[0030] In one possible implementation, the first and second feature values, along with a preset benchmark value obtained from historical data analysis, are input into the parameter decision logic to calculate the shooting parameters used to control the high dynamic range camera to perform multi-frame exposure acquisition. The parameter decision logic aims to dynamically adapt the optimal exposure strategy based on the current license plate grayscale and its distribution characteristics in the scene. Specifically, firstly, based on the first and second feature values ​​and a historical statistical value reflecting extreme grayscale conditions, a compression coefficient for adjusting the short exposure duration is derived; then, this coefficient is used to scale a preset short exposure reference duration to obtain the target short exposure duration used when finally performing short exposure acquisition.

[0031] S104 indicates that the high dynamic range camera acquires long-exposure and short-exposure images of the vehicle based on the shooting parameters.

[0032] In one possible implementation, the shooting parameters determined in step S103, including the target short exposure duration and the target exposure ratio, are converted into control commands that drive the high dynamic range camera hardware to perform specific exposure operations. Commands are sent to the camera deployed at the target location via a designated communication interface, instructing its image signal processor to set the integration time of the short exposure frame based on the target short exposure duration and the integration time of the long exposure frame based on the product of the target short exposure duration and the target exposure ratio. Subsequently, within a single shooting cycle, the high dynamic range camera sensor completes two exposures sequentially according to the two different integration times, capturing a short exposure image reflecting the highlight details of the scene and a long exposure image reflecting the shadow details of the scene, respectively.

[0033] S105. Perform composite processing on the long exposure image and the short exposure image to determine the high dynamic range composite image for license plate recognition.

[0034] One possible implementation involves fusing and enhancing long-exposure and short-exposure images to generate a composite image with significantly expanded dynamic range. First, the two source images are spatially aligned to eliminate pixel position discrepancies caused by slight vehicle movement or shooting intervals. Then, a fusion weight is assigned to each pixel position in the aligned images, determining the proportion of information extracted from each image. Based on the fusion weight, the pixel values ​​at corresponding positions in the two images are fused, thus preserving highlight details from the short-exposure image and shadow details from the long-exposure image in the composite result. Finally, the fused intermediate image undergoes post-processing, including color interpolation, noise reduction, and sharpening, to generate a high dynamic range composite image with a standard color space that can be directly used by subsequent license plate recognition algorithms.

[0035] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment utilizes a high dynamic range camera to acquire a single frame of ordinary image when the synthesis function is turned off and analyzes the distribution characteristics of its license plate grayscale and similar grayscale regions. Based on this analysis result, the specific parameters for the camera to capture multiple frames of synthesized images are dynamically determined. Finally, the camera is controlled to acquire and synthesize high dynamic range images according to these parameters for recognition. The use of high dynamic range cameras is transformed from the traditional, fixed-parameter image acquisition mode to an intelligent closed-loop control system based on real-time scene perception. By first determining the current illumination and license plate status reflected in the ordinary image, and then generating the optimal shooting command accordingly, the problem of overexposure or underexposure of the license plate area caused by drastic changes in illumination can be fundamentally overcome. This results in a stable output of license plate images with clear details and appropriate contrast, greatly improving the accuracy of license plate recognition and the overall robustness of the system under complex lighting conditions in all weather conditions. This solves the technical problem that existing methods lack a perception and feedback mechanism for real-time scene illumination characteristics, leading to unstable HDR image quality under extreme or rapidly changing lighting conditions.

[0036] In one possible implementation, before acquiring the ordinary image of the vehicle captured by the high dynamic range camera at the target deployment location, it is necessary to determine the target deployment location. This process can be specifically implemented through the following steps S201-S203, which will be described in detail below.

[0037] S201. Obtain multiple historical vehicle images collected at multiple candidate deployment locations at the entrance and exit of the parking lot within a preset historical monitoring period.

[0038] In one possible implementation, to determine the final target deployment locations for real-time identification, a historical image database must first be collected and established. This process is performed on multiple candidate deployment locations at all available camera installation locations at the parking lot entrance and exit. At each candidate location, a high dynamic range (HMR) camera is deployed and set to a state where dynamic range synthesis is disabled to simulate the single-exposure working mode of a regular camera. Within a preset historical monitoring period (exemplarily, one consecutive week), images of passing vehicles are automatically acquired at each candidate location at fixed time intervals (exemplarily, 10 seconds), thereby obtaining multiple sets of historical vehicle images corresponding to each location.

[0039] S202. Based on the grayscale performance of the license plate area at multiple candidate locations in historical vehicle images, calculate the exposure evaluation index for each candidate location.

[0040] Among them, the exposure assessment index is used to characterize the risk of overexposure in the vehicle license plate area.

[0041] In one possible implementation, firstly, based on a preset deep learning detection model (exemplarily, the preset deep learning detection model is the YOLO model), historical license plate regions are located from each historical image, and the average grayscale of the license plate region in each image is calculated. Then, two key historical data dimensions are statistically analyzed for each candidate location: the first dimension is the ratio of the number of images where the average grayscale of the license plate region exceeds a preset overexposure grayscale threshold within a preset historical monitoring period to the total number of historical images collected at that location; this ratio reflects the frequency of overexposure of the license plate at that location. The second dimension is the cumulative sum of the average grayscale of the license plate region in all historical images within the same monitoring period; this sum reflects the overall brightness or cumulative grayscale of the license plate image at that location. Finally, by combining the above two data dimensions—the overexposure frequency ratio and the cumulative grayscale sum—a single quantitative value is generated to characterize the risk of overexposure of the license plate at that location, i.e., the exposure evaluation index.

[0042] S203. The candidate deployment location with the lowest exposure evaluation index is determined as the target deployment location.

[0043] One possible implementation involves comparing the exposure assessment metrics of all candidate locations. These metrics directly and quantitatively characterize the risk of overexposure of the license plate area in historical monitoring data for that location; a higher metric value indicates a worse historical performance and higher risk, while a lower metric value indicates a better historical performance and lower risk. The candidate location with the lowest exposure assessment metric value is selected from these comparisons and ultimately designated as the target location.

[0044] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment comprehensively evaluates and selects the optimal location by analyzing the grayscale performance of the license plate area in images collected from multiple candidate locations within a historical time period. The beneficial effect of this solution is that it transforms the camera installation location from a subjective selection based on experience to an objective optimization decision driven by long-term historical lighting data. By selecting the location with the lowest risk of license plate overexposure according to historical data, it minimizes the extreme lighting interference (such as direct sunlight at fixed times) inherently introduced by improper installation angles from the physical source, providing a more stable data acquisition foundation for subsequent image processing algorithms.

[0045] In one possible implementation, the process of calculating the exposure evaluation index for each candidate license plate location based on the grayscale performance of the license plate area at multiple candidate locations in a historical vehicle image can be specifically implemented through the following steps S301-S304, which will be explained in detail below.

[0046] S301. Determine the ratio of the number of target images to the total number of images in multiple historical vehicle ordinary images.

[0047] The target image is a historical vehicle image in which the average gray level of the license plate area exceeds a preset first gray level threshold among multiple historical vehicle ordinary images.

[0048] In one possible implementation, for each of multiple historical vehicle images, the average grayscale of the license plate region is calculated based on its historical license plate area. A preset first grayscale threshold is set to determine whether the grayscale is excessive. Historical vehicle images with an average grayscale exceeding this threshold are identified as target images. Subsequently, the total number of all historical vehicle images at that location is counted, along with the number of those identified as target images, and the ratio between the two is calculated.

[0049] It is understood that the first grayscale threshold can be set according to the grayscale range of the image. This invention does not limit this. For example, the grayscale range of the image is 0-255, and the first grayscale threshold is 220.

[0050] S302. Determine the mean of the sum of the average gray levels of the license plate area in multiple historical vehicle images.

[0051] In one possible implementation, for each of the multiple historical vehicle general images, the accurate average gray value of the license plate area is calculated based on its historical license plate area. Then, the average gray values ​​of the license plate areas calculated from all the historical vehicle general images are summed to determine the mean of the sum.

[0052] S303. Determine the exposure evaluation index based on the ratio and the mean of the sums.

[0053] For example, the exposure evaluation index of deployment location a in time period j. Satisfy the following formula 1:

[0054] Formula 1

[0055] in, Indicates the time period Inside, deployment location The number of historical vehicle images collected where the average gray level of the license plate area exceeds a preset first gray level threshold. Indicates the time period Inside, deployment location Total number of historical vehicle images collected; Indicates the time period Inside, deployment location The first collection The average grayscale value of all pixels within the license plate area in a typical historical vehicle image.

[0056] The physical meaning is the deployment location. During the period The frequency of overexposed license plates in images; the higher the value, the higher the probability of overexposure at that location. It is the average of the cumulative sums, and its physical meaning is position. During the period The average of the sum of gray values ​​of all license plate areas reflects the overall gray level of the license plate image at that location. The larger the value, the brighter the location is or the more severe the overexposure. The two are multiplicative, which means that the final risk index is positively correlated with the overexposure frequency and the overall gray level. An increase in either one will lead to an increase in the final risk assessment value. The physical meaning is that it comprehensively quantifies the candidate deployment positions. During the period The risk of the license plate area being overexposed is a comprehensive indicator used to compare the advantages and disadvantages of different positions.

[0057] The technical solution provided in the above embodiments can bring at least the following beneficial effects: This embodiment provides a scientific, quantitative, and comprehensive evaluation model. It not only considers the frequency of overexposure but also takes into account the severity of overexposure or overbrightness, avoiding misjudgments that may be caused by a single indicator. This comprehensive evaluation makes site selection decisions more refined and reliable, ensuring that the selected location has optimal overall lighting performance in long-term operation.

[0058] In one possible implementation, it is also necessary to determine the historical license plate area in the general image of each historical vehicle. This process can be specifically implemented through the following S401, which will be described in detail below.

[0059] S401. Based on a preset deep learning detection model, process multiple historical vehicle images to determine the historical license plate region in each historical vehicle image.

[0060] In one possible implementation, a preset deep learning detection model (exemplarily, a YOLO model) is invoked and run. Historical vehicle images are input one by one into this trained model. The model performs forward inference on each image and outputs predicted bounding boxes for the vehicle targets and, more precisely, the license plate targets. From the model output, bounding boxes corresponding to the license plate category and with a confidence level higher than a preset threshold (exemplarily, 0.75) are selected as the automated localization result of the license plate location in that historical image, i.e., the historical license plate region box.

[0061] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment realizes the automated, batch, and high-precision extraction of key areas (license plates) from massive historical image data. This replaces inefficient and inconsistent manual annotation or traditional image processing methods, providing an accurate and consistent data foundation for subsequently reliably calculating the grayscale performance of the license plate area at each location and then making scientific deployment decisions, making large-scale, data-driven parking lot camera planning and deployment possible.

[0062] In one possible implementation, it is also necessary to determine the historical grayscale statistical features of multiple historical vehicle general images. This process can be specifically implemented through the following S501, which will be described in detail below.

[0063] S501. Based on the pixel grayscale values ​​of the historical license plate region of each historical vehicle ordinary image, determine the historical grayscale statistical features of multiple historical vehicle ordinary images.

[0064] In one possible implementation, after determining the historical license plate areas, a systematic statistical analysis is performed based on the original pixel grayscale values ​​contained within these areas to extract a set of core statistics characterizing the historical lighting conditions at that location, namely, historical grayscale statistical features. This set of features includes at least the following three statistical dimensions: First, the maximum average grayscale value of pixels within the license plate area bounding box in all historical vehicle images is calculated as the maximum average grayscale of the historical license plate area, recording the most extreme bright license plate conditions that have occurred historically; Second, the difference between the maximum and minimum pixel grayscale values ​​in the overall image of each historical vehicle image is calculated, and the maximum value among these differences is selected as the maximum grayscale dynamic range of the historical vehicle image, representing the maximum span between the brightest and darkest parts of a scene that has occurred historically; Third, the average grayscale of all pixels in the overall image of each historical vehicle image is calculated, and these average values ​​are then statistically analyzed (e.g., the mean is calculated) to obtain the average grayscale information of the historical vehicle image, reflecting the overall lighting level of the historical imaging at that location.

[0065] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment extracts scattered, raw historical image data into a series of representative reference benchmark values. These statistical characteristics constitute the system's memory of normal or historical typical lighting conditions, providing a comparable objective benchmark for real-time judgment of whether the current lighting is too bright, too dark, or has an excessive dynamic range, making the system's parameter adjustment decisions based on evidence.

[0066] In one possible implementation, the process of determining the shooting parameters of the high dynamic range camera based on the first feature value and the second feature value can be specifically implemented through the following S601-S603, which will be described in detail below.

[0067] S601. Process ordinary images based on a preset deep learning detection model to determine the license plate region in the ordinary image.

[0068] One possible implementation involves invoking the same pre-trained deep learning detection model used in the historical data processing stage. A regular image is input into this trained model, which performs forward inference and outputs the category and bounding box predictions for each target contained within. From the model's output, bounding boxes that are classified as license plates and whose prediction confidence is higher than a preset threshold are selected. These bounding boxes represent the high-precision localization of the current vehicle's license plate in the image, i.e., the license plate region box in the regular image.

[0069] S602. Determine the first feature value based on the grayscale values ​​of all pixels within the license plate area.

[0070] In one possible implementation, the grayscale value of each pixel covered within the license plate region is read, and statistical calculations are performed based on this raw grayscale data to obtain a representative value that can characterize the overall brightness of the region, namely the first feature value. The first feature value is determined to be the average grayscale value of all pixels within the license plate region, which is calculated by summing the grayscale values ​​of all pixels within the region and then dividing by the total number of pixels.

[0071] S603. Based on the total number of pixels in a normal image whose grayscale value is less than the difference between the first feature value and the grayscale value is less than a preset grayscale tolerance threshold, determine the second feature value.

[0072] In one possible implementation, the entire pixel array of a normal image is scanned and compared based on a first feature value. A preset grayscale tolerance threshold (exemplarily 20) is set to define the grayscale range close to the first feature value. Each pixel in the normal image is traversed, and the absolute value of the difference between its grayscale value and the first feature value is calculated. The total number of pixels whose absolute difference is less than the preset grayscale tolerance threshold is counted, and this total number is determined as the second feature value.

[0073] It is understandable that the grayscale tolerance threshold is used to define the tolerance range for grayscale similarity, and can be optimized and corrected based on actual results; this invention does not limit this. The magnitude of the second feature value (i.e., the total number of pixels that meet the conditions) intuitively reflects whether pixels with similar grayscale to the license plate area are widely distributed or concentrated in the local area of ​​the license plate in the entire image. If this value is very small, it indicates that only a very small area such as the license plate is at this grayscale level; if this value is very large, it indicates that the grayscale of most areas in the image is similar to that of the license plate.

[0074] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment provides an effective means to accurately quantify and describe the grayscale state of a license plate in the current scene and its relationship with the surrounding environment. The first feature value directly reflects the brightness of the license plate itself, while the second feature value cleverly characterizes the universality or isolation of the license plate grayscale in the entire image. The combination of the two can clearly distinguish between different situations such as strong reflective backlighting where the license plate is bright and the surrounding area is dark, and typical backlighting where the entire image is dark, providing a key and clear input basis for subsequent differentiated parameter adjustment strategies.

[0075] In one possible implementation, the shooting parameters include: target short exposure time; the process of determining the shooting parameters of the high dynamic range camera based on the first feature value and the second feature value can be specifically implemented through the following S701-S702, which will be described in detail below.

[0076] S701. Based on the first feature value, the second feature value, and the maximum average gray value of the historical license plate area, determine the short exposure time compression coefficient.

[0077] For example, the short exposure compression factor of a normal image of the k-th vehicle. The following formula 2 is satisfied:

[0078] Formula 2

[0079] in, Indicates the first The first feature value of a normal image of a vehicle (average gray value of the license plate area); This represents the maximum average grayscale of the historical license plate area in all historical vehicle images obtained from historical data analysis. Indicates the first The second feature value of a typical image of a vehicle; It is a preset reference area, such as the total number of pixels in the image, which is the benchmark value used for normalized comparison; For parameter tuning coefficients, if When the value is 0, it is set to the minimum value other than 0. Dimensions and same.

[0080] This value represents the proportion of the current license plate's grayscale relative to the brightest historical value. The larger the value, the brighter the current license plate is, and the closer it is to or beyond the historical limit. It is the reciprocal of the total area of ​​pixel regions in the current image that are similar in grayscale to the license plate, representing the isolation of the license plate's grayscale. The smaller the value (e.g., in cases of strong backlighting where the license plate appears as an isolated bright spot), the larger this value becomes; multiplying the two values ​​together results in... Increase the value in two cases: when the license plate itself is very bright (first option is larger); when the license plate is an isolated bright area in the image (second option is larger). The larger the value, the more drastically the shorter the exposure time needs to be to suppress the highlights; The physical meaning is to quantify the urgency or necessity of compressing the baseline short exposure time to adapt to the current scene (especially the grayscale state of license plates).

[0081] S702. Adjust the preset short exposure reference duration based on the short exposure duration compression coefficient to determine the target short exposure duration.

[0082] In one possible implementation, a preset short exposure reference duration is calculated using a short exposure compression factor to obtain a calibrated new duration value, i.e., the target short exposure duration. This calculation aims to transform the coefficient representing the necessity of compression into a precise time parameter that can directly control the hardware exposure behavior.

[0083] For example, the target short exposure time of a normal image of the k-th vehicle. The following formula 3 is satisfied:

[0084] Formula 3

[0085] in, The preset short exposure duration reference size is, for example, 1ms; Let be the compression factor for the short exposure time of the ordinary image of the k-th vehicle.

[0086] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment realizes adaptive and precise control of the short exposure time, a key parameter in HDR imaging. For different lighting conditions identified, the most suitable short exposure time can be automatically calculated and set, thereby ensuring that in the synthesized HDR image, the highlight details of the license plate can be effectively preserved without overflow, or that the license plate information in the dark areas can be fully captured while noise is controllable, significantly improving the success rate of obtaining a usable license plate image in a single shot.

[0087] In one possible implementation, the shooting parameters also include: target exposure ratio; after determining the target short exposure duration, it is also necessary to adjust the target exposure ratio, which can be specifically implemented through the following S801-S803, which will be explained in detail below.

[0088] S801. Determine the third feature value of a normal image.

[0089] The third feature value is the difference between the maximum and minimum pixel gray values ​​in a normal image.

[0090] In one possible implementation, all pixels in the ordinary image are first traversed to find the maximum and minimum pixel gray values; then, the difference between the maximum and minimum pixel gray values ​​is calculated, and this difference is the third feature value.

[0091] S802. Based on the third feature value, the maximum grayscale dynamic range of historical vehicle ordinary images, and the target short exposure duration, determine the exposure ratio adjustment coefficient.

[0092] In one possible implementation, the third feature value, the maximum grayscale dynamic range of the historical vehicle ordinary image (and the short exposure duration of the step target) are input together into a parameter calculation function for deciding the exposure ratio to determine the exposure ratio adjustment coefficient.

[0093] For example, the exposure ratio adjustment factor of the normal image of the k-th vehicle. Satisfy the following formula 4:

[0094] Formula 4

[0095] Indicates the current number In a typical image of a vehicle, the difference between the maximum and minimum pixel grayscale values ​​(i.e., the third feature value) is used to characterize the overall grayscale dynamic range of the current scene. This represents the general images of all historical vehicles obtained from historical data analysis. The maximum value (i.e., the maximum grayscale dynamic range of the historical image). Indicates the first The vehicle's target exposure time is short; This represents a preset short exposure reference duration (e.g., 1 millisecond) with time dimensions, which is the benchmark value used for normalization comparison; P represents the truncation mapping function, used to map the input value to [-0.8, 0.8]. The calculation method is as follows: first, normalize the maximum and minimum values, mapping to [0, 1], then multiply by 1.6, mapping to [0, 1.6], and finally subtract 0.8, mapping to [-0.8, 0.8], to determine the final exposure ratio adjustment coefficient, where the maximum and minimum values ​​are determined based on historical monitoring data statistics; For parameter tuning coefficients, if When the value is 0, it is set to the minimum value other than 0. Dimensions and same.

[0096] It is the scene dynamic range ratio, which represents the proportion of the current scene grayscale span relative to the historical maximum span. The larger the value, the larger the scene light ratio. It is the exposure time compression ratio. It is a fixed reference value. When When the exposure time is reduced (to suppress highlights), the ratio increases, which directly indicates the degree to which the current exposure time has been compressed relative to the reference value. The degree to which the exposure ratio needs to be increased is positively correlated with and When the scene has a very high light ratio ( (Large) and has adopted an extremely short exposure to deal with the highlights ( When (large), The value will increase significantly, indicating that the exposure ratio needs to be greatly increased to fully capture shadow details.

[0097] S803. Adjust the preset exposure ratio reference value based on the exposure ratio adjustment coefficient to determine the target exposure ratio.

[0098] In one possible implementation, the adjustment operation essentially involves scaling a preset reference value based on an adjustment coefficient that reflects the dynamic range requirements of the current scene and the short exposure strategy. Specifically, a calibrated target exposure ratio is generated by performing a specific calculation between the exposure ratio adjustment coefficient and the preset exposure ratio reference value.

[0099] For example, the target exposure ratio of a normal image of the k-th vehicle. The following formula 5 is satisfied:

[0100] Formula 5

[0101] in, For example, the exposure ratio reference value is set to 120:1; is the exposure ratio adjustment factor for the normal image of the k-th vehicle.

[0102] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment adds a layer of protection, capable of identifying extreme situations where even optimizing HDR shooting parameters cannot obtain sufficient image signals (overall environment is too dark). Through intelligent triggering of supplemental lighting, the system can actively improve lighting conditions, making up for the limitations of relying solely on electronic adjustments, ensuring that even in environments such as nighttime or extremely dark conditions, identifiable license plate images can still be obtained through auxiliary lighting, thereby achieving reliable operation around the clock.

[0103] In one possible implementation, it is also necessary to determine whether the license plate area of ​​the vehicle needs to be illuminated. This process can be implemented through the following steps S901-S903, which will be explained in detail below.

[0104] S901. Determine the fourth feature value of a normal image.

[0105] The fourth feature value is the average gray level of all pixels in a normal image.

[0106] In one possible implementation, each pixel in the ordinary image is traversed, its grayscale value is obtained, and statistical calculations are performed based on the grayscale values ​​of all pixels to obtain their arithmetic mean, which is the fourth feature value.

[0107] S902. Based on the first feature value, the fourth feature value, and the average grayscale information of historical vehicle ordinary images, determine the evaluation value of the necessity of supplementary lighting.

[0108] In one possible implementation, to determine whether the current ambient light is low enough to require the activation of an auxiliary light source, the first feature value (the grayscale of the license plate itself), the fourth feature value (the average grayscale of the entire scene), and the average grayscale information of historical ordinary vehicle images (historical overall average grayscale level) are input into an evaluation function. The logic of this evaluation function aims to comprehensively compare the grayscale relationships between the local area of ​​the license plate, the current scene as a whole, and the historical average level to quantify the degree to which the current scene deviates from normal lighting conditions, thereby calculating a quantitative score for decision-making, namely, the evaluation value of the necessity of supplementary lighting.

[0109] For example, the evaluation value of the supplementary lighting necessity for a normal image of the k-th vehicle. Satisfy the following formula 6:

[0110] Formula 6

[0111] in, It is the first eigenvalue; It is the fourth eigenvalue; Indicates the first The average grayscale value of all pixels in a typical historical image of a vehicle; This represents the total number of ordinary images of historical vehicles; For parameter tuning coefficients, if When the value is 0, it is set to the minimum value other than 0. Dimensions and same; Used to perform minimum normalization of maximum value The value is mapped to the interval [0, 1], thus solving the problem that the direct product result may be negative, making the evaluation value meaningless. =0 indicates that no additional lighting is needed (e.g., the scene is very bright). =1 indicates that the necessity of supplemental lighting is at its theoretical maximum (both the license plate and the overall environment are extremely dark). The increase from 0 to 1 intuitively and continuously reflects the growing necessity for supplemental lighting.

[0112] It is the reciprocal of the average grayscale value of the license plate. The smaller the value (the darker the license plate), the larger this value is; It is the sum of the differences between the historical average gray level and the current overall average gray level, representing the overall degree of deviation of the current image's overall gray level from the historical average level. The smaller the value (the darker the overall image), the larger the difference between each item in this term, and the larger the sum. The logic behind multiplying the two terms is that the necessity of supplemental lighting depends on two factors simultaneously: whether the license plate itself is dark enough (the first factor), and whether the entire environment is abnormally dark, and darker than historical norms (the second factor). When the license plate is dark (the first factor is large) and the overall environment is also abnormally dark (the second factor is large), The value will increase significantly, indicating that lighting conditions need to be actively improved by supplementing light; The physical meaning is to comprehensively assess whether the current ambient light is dark enough to affect the acquisition of license plate information, to the point that it is necessary to activate the active lighting device.

[0113] S903. If the assessment value of the necessity of supplementary lighting is greater than the preset supplementary lighting threshold, determine to supplement the license plate area of ​​the vehicle.

[0114] In one possible implementation, the assessment value of the necessity of supplemental lighting is compared with a preset supplemental lighting threshold (exemplarily, the preset supplemental lighting threshold is 0.7). The supplemental lighting threshold is a calibrated threshold value used to determine whether the lack of ambient light has reached a level requiring manual intervention. If the comparison result shows that the assessment value is greater than the preset supplemental lighting threshold, a specific control command is generated and output to initiate supplemental lighting operation on the license plate area of ​​the vehicle.

[0115] Please see Figure 2 This illustration shows a system architecture diagram 200 of a vehicle vision recognition system for unmanned parking lots according to an embodiment of the present invention. The system includes: an image acquisition module 201, used to acquire ordinary images of vehicles captured by a high dynamic range camera at a target deployment location; the ordinary images are images acquired after the dynamic range synthesis function of the high dynamic range camera is turned off; a feature value determination module 202, used to determine a first feature value and a second feature value of the vehicle based on the ordinary image; the first feature value is used to characterize the grayscale information of the license plate area; the second feature value is used to characterize the distribution range of pixels in the ordinary image whose grayscale is similar to that of the license plate area; a parameter decision module 203, used to determine the shooting parameters of the high dynamic range camera based on the first feature value and the second feature value; a shooting control module 204, used to instruct the high dynamic range camera to acquire long-exposure images and short-exposure images of the vehicle based on the shooting parameters; and an image synthesis module 205, used to perform synthesis processing on the long-exposure images and short-exposure images to determine a high dynamic range synthesized image for license plate recognition.

[0116] The technical solutions provided in the above embodiments offer at least the following beneficial effects: This embodiment provides a direct and clear implementation architecture at the hardware entity or software system level. This modular system design not only makes the deployment and implementation of the technical solutions clearer and more convenient, but also facilitates industrial applications.

[0117] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0118] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A vehicle visual recognition method for unmanned parking lots, characterized in that, The method includes: Acquire ordinary images of the vehicle captured by a high dynamic range camera at the target deployment location; the ordinary images are images captured by the high dynamic range camera after the dynamic range synthesis function is turned off; The ordinary image is processed based on a preset deep learning detection model to determine the license plate region in the ordinary image; Based on the grayscale values ​​of all pixels within the license plate area, a first feature value is determined; the first feature value is used to characterize the grayscale information of the license plate area. Based on the total number of pixels in the ordinary image whose grayscale value differs from the first feature value by less than a preset grayscale tolerance threshold, a second feature value is determined; the second feature value is used to characterize the distribution range of pixels in the ordinary image whose grayscale value is similar to that of the license plate area. Based on the first feature value, the second feature value, and the maximum average gray level of the historical license plate area, the short exposure time compression coefficient is determined; the short exposure time compression coefficient of the ordinary image of the k-th vehicle is... Satisfy the following formula: Indicates the first The first feature value of a typical image of a vehicle; This represents the maximum average grayscale of the historical license plate area in all historical vehicle images obtained from historical data analysis. Indicates the first The second feature value of a typical image of a vehicle; It is a preset reference area; These are the parameter tuning coefficients; This is the normalization function; The preset short exposure reference duration is adjusted based on the short exposure duration compression coefficient to determine the target short exposure duration; Short exposure time for the target of the ordinary image of the k-th vehicle Satisfy the following formula: This is a preset reference size for short exposure duration; Let be the compression factor for the short exposure time of the ordinary image of the k-th vehicle; Determine a third feature value for the ordinary image; the third feature value is the difference between the maximum and minimum pixel grayscale values ​​in the ordinary image. Based on the third feature value, the maximum grayscale dynamic range of historical vehicle ordinary images, and the target short exposure duration, the exposure ratio adjustment coefficient is determined; Exposure ratio adjustment factor for the normal image of the kth vehicle Satisfy the following formula Indicates the current number In a typical image of a vehicle, the difference between the maximum and minimum pixel grayscale values; This represents the general images of all historical vehicles obtained from historical data analysis. The maximum value; Indicates the first The vehicle's target exposure time is short; This represents a preset short exposure reference duration with time dimensions; P represents the truncation mapping function, used to map the input value to [-0.8, 0.8]. The calculation method is to first normalize the maximum and minimum values, mapping to [0, 1], then multiply by 1.6, mapping to [0, 1.6], and finally subtract 0.8, mapping to [-0.8, 0.8], to determine the final exposure ratio adjustment coefficient, where the maximum and minimum values ​​are determined based on historical monitoring data statistics. These are the parameter tuning coefficients; The preset exposure ratio reference value is adjusted based on the exposure ratio adjustment coefficient to determine the target exposure ratio; The target exposure ratio of the normal image of the k-th vehicle Satisfy the following formula: The preset exposure ratio reference value; is the exposure ratio adjustment factor for the normal image of the k-th vehicle; The integration time of the short exposure frame is set according to the target short exposure duration, and the integration time of the long exposure frame is set according to the product of the target short exposure duration and the target exposure ratio. The high dynamic range camera is instructed to acquire long-exposure images and short-exposure images of the vehicle based on the integration time of the short-exposure frame and the integration time of the long-exposure frame; The long-exposure image and the short-exposure image are combined to determine a high dynamic range composite image for license plate recognition.

2. The vehicle visual recognition method for unmanned parking lots according to claim 1, characterized in that, The method further includes: Acquire multiple historical vehicle images collected at multiple candidate deployment locations at the entrance and exit of the parking lot within a preset historical monitoring period; Based on the grayscale performance of the license plate area at multiple candidate locations in the historical vehicle general image, an exposure evaluation index is calculated for each candidate location; the exposure evaluation index is used to characterize the risk of overexposure in the vehicle license plate area. The candidate placement location with the lowest exposure evaluation index is determined as the target placement location.

3. The vehicle visual recognition method for unmanned parking lots according to claim 2, characterized in that, The method involves calculating the exposure evaluation index for each candidate license plate location based on the grayscale performance of the license plate area at multiple candidate locations in the historical vehicle image, including: Determine the ratio of the number of target images to the total number of images among the plurality of historical vehicle ordinary images; the target image is a historical vehicle image whose average gray level of the license plate area exceeds a preset first gray level threshold among the plurality of historical vehicle ordinary images. Determine the mean of the sum of the average gray levels of the license plate regions in the multiple historical vehicle images; The exposure evaluation index is determined based on the ratio and the mean of the sums.

4. The vehicle visual recognition method for unmanned parking lots according to claim 3, characterized in that, The method further includes: The multiple historical vehicle images are processed using a preset deep learning detection model to determine the historical license plate region in each historical vehicle image.

5. The vehicle visual recognition method for unmanned parking lots according to claim 4, characterized in that, The method further includes: Based on the pixel grayscale values ​​of the historical license plate region in each historical vehicle ordinary image, the historical grayscale statistical features of the multiple historical vehicle ordinary images are determined; the historical grayscale statistical features include: the maximum average grayscale of the historical license plate region, the maximum grayscale dynamic range of the historical vehicle ordinary image, and the average grayscale information of the historical vehicle ordinary image.

6. The vehicle visual recognition method for unmanned parking lots according to claim 1, characterized in that, The method further includes: Determine the fourth feature value of the ordinary image; the fourth feature value is the average gray level of all pixels in the ordinary image; Based on the first feature value, the fourth feature value, and the average grayscale information of the historical vehicle ordinary image, a supplementary lighting necessity assessment value is determined. If the assessment value of the necessity of supplemental lighting is greater than the preset supplemental lighting threshold, it is determined that supplemental lighting should be applied to the license plate area of ​​the vehicle.

7. A vehicle vision recognition system for unmanned parking lots, characterized in that, include: The image acquisition module is used to acquire ordinary images of the vehicle captured by the high dynamic range camera at the target deployment location; The ordinary image is the image captured by the high dynamic range camera after the dynamic range synthesis function is turned off; The feature value determination module is used to process the ordinary image based on a preset deep learning detection model to determine the license plate region in the ordinary image; The first feature value is determined based on the grayscale values ​​of all pixels within the license plate area; The first feature value is used to characterize the grayscale information of the license plate area; The second feature value is determined based on the total number of pixels in the ordinary image whose grayscale value is less than the difference between the first feature value and the grayscale value is less than a preset grayscale tolerance threshold. The second feature value is used to characterize the distribution range of pixels in the ordinary image whose gray level is similar to that of the license plate area; The parameter decision module is used to determine the short exposure compression coefficient based on the first feature value, the second feature value, and the maximum average gray level of the historical license plate area; the short exposure compression coefficient of the ordinary image of the k-th vehicle. Satisfy the following formula: Indicates the first The first feature value of a typical image of a vehicle; This represents the maximum average grayscale of the historical license plate area in all historical vehicle images obtained from historical data analysis. Indicates the first The second feature value of a typical image of a vehicle; It is a preset reference area; These are the parameter tuning coefficients; This is the normalization function; The preset short exposure reference duration is adjusted based on the short exposure duration compression coefficient to determine the target short exposure duration; Short exposure time for the target of the ordinary image of the k-th vehicle Satisfy the following formula: This is a preset reference size for short exposure duration; Let be the compression factor for the short exposure time of the ordinary image of the k-th vehicle; Determine a third feature value for the ordinary image; the third feature value is the difference between the maximum and minimum pixel grayscale values ​​in the ordinary image. Based on the third feature value, the maximum grayscale dynamic range of historical vehicle ordinary images, and the target short exposure duration, the exposure ratio adjustment coefficient is determined; Exposure ratio adjustment factor for the normal image of the kth vehicle Satisfy the following formula Indicates the current number In a typical image of a vehicle, the difference between the maximum and minimum pixel grayscale values; This represents the general images of all historical vehicles obtained from historical data analysis. The maximum value; Indicates the first The vehicle's target exposure time is short; It represents a preset short exposure reference duration with a time dimension; P represents the truncation mapping function, which maps the input value to [-0.8, 0.8]. The calculation method is as follows: first, the maximum and minimum values ​​are normalized and mapped to [0, 1], then multiplied by 1.6 and mapped to [0, 1.6], and finally 0.8 is subtracted and mapped to [-0.8, 0.8] to determine the final exposure ratio adjustment coefficient. The maximum and minimum values ​​are determined based on the statistical values ​​of historical monitoring data. These are the parameter tuning coefficients; The preset exposure ratio reference value is adjusted based on the exposure ratio adjustment coefficient to determine the target exposure ratio; The target exposure ratio of the normal image of the k-th vehicle Satisfy the following formula: The preset exposure ratio reference value; is the exposure ratio adjustment factor for the normal image of the k-th vehicle; The integration time of the short exposure frame is set according to the target short exposure duration, and the integration time of the long exposure frame is set according to the product of the target short exposure duration and the target exposure ratio. The shooting control module is used to instruct the high dynamic range camera to acquire long-exposure images and short-exposure images of the vehicle based on the integration time of the short-exposure frame and the integration time of the long-exposure frame; An image synthesis module is used to synthesize the long-exposure image and the short-exposure image to determine a high dynamic range synthesized image for license plate recognition.

Images