A license plate recognition result stable output method and system based on continuous frame confirmation

Abstract

The application relates to the field of license plate recognition, and particularly discloses a license plate recognition result stable output method and system based on continuous frame confirmation, which comprises the following steps: acquiring a current frame recognition result and a historical recognition result; directly outputting the current frame recognition result when an instant output condition is met; starting a continuous frame confirmation process and determining an output result based on time sequence consistency judgment when the instant output condition is not met; and through joint judgment of the time sequence consistency of single frame recognition results and continuous multiple frame recognition results, the output jump problem caused by single frame recognition fluctuation is effectively inhibited while the output response timeliness is maintained, and the output stability of the license plate recognition system in a video stream scene is improved.

Description

Technical Field

[0001] This invention belongs to the field of license plate recognition, specifically a method and system for stable output of license plate recognition results based on continuous frame confirmation. Background Technology

[0002] License plate recognition, as one of the core technologies of intelligent transportation systems, has expanded its applications from initial static image recognition to continuous recognition in video stream scenarios. In applications such as parking lot entrances and exits, road checkpoints, and community access control, cameras continuously capture video streams. The recognition system performs license plate detection and character recognition on each frame of the image, and uses the recognition results for subsequent traffic control, toll collection, or security linkage. With the development of deep learning technology, neural network-based license plate detection and optical character recognition models have achieved high accuracy rates on standard test sets.

[0003] In video streaming scenarios, existing license plate recognition systems typically employ a frame-by-frame independent recognition approach. This means that the detection and recognition process is performed separately for each frame, and the recognition result for each frame is directly output. Some systems add a confidence threshold judgment, only outputting the result if the confidence level of the recognition result is higher than a certain set threshold; otherwise, the result for that frame is discarded or a null value is output. The rationale behind this approach is to ensure the accuracy of the output results and prevent low-quality recognition results from affecting the correctness of subsequent business logic.

[0004] However, the frame-by-frame independent recognition method ignores the inherent temporal correlation between recognition results of adjacent frames in a video stream. In real-world road environments, the recognition results of the same vehicle should ideally remain consistent across multiple consecutive frames. However, due to factors such as changes in vehicle pose, lighting fluctuations, motion blur, partial occlusion, and fluctuations in the recognition model's output, discrepancies frequently occur between recognition results of adjacent frames. For example, a frame might show a complete license plate number, an adjacent frame might fail to recognize it due to brief occlusion, and then the next frame might restore the complete recognition. For such brief fluctuations, existing frame-by-frame output methods might cause the recognition result to jump from a complete license plate number to a null value and then back to a complete license plate number, or output slightly different license plate numbers between adjacent frames. This frequent change in output results can lead to logical confusion in backend business systems that rely on the recognition results, reducing the overall reliability of the system. The root cause is that existing solutions lack verification of the continuity of recognition results in the temporal dimension and fail to comprehensively judge the quality of single-frame recognition combined with the temporal consistency of historical recognition results. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for stable output of license plate recognition results based on continuous frame confirmation. This solves the problem in the prior art where the lack of continuous verification of the recognition results in the time dimension due to independent frame-by-frame recognition leads to frequent jumps in the output results caused by recognition differences between adjacent frames.

[0006] This invention provides a method for stable output of license plate recognition results based on continuous frame confirmation, comprising the following steps:

[0007] The current frame image in the video stream is acquired, license plate recognition is performed on the current frame image to obtain the current frame recognition result, and historical recognition results are acquired, including stable recognition results output before the current moment and recognition results of the most recent several frames;

[0008] The current frame recognition result is compared with the historical recognition result. When the current frame recognition result is the same as the stable recognition result in the historical recognition result, or when the confidence level of the current frame recognition result is greater than or equal to the first preset threshold and the difference between the current frame recognition result and the stable recognition result is within the preset difference tolerance range, the current frame recognition result is output as a stable output result.

[0009] When the current frame recognition result is empty or its confidence level is lower than the second preset threshold, the first timing window is started. Within the first timing window, the stable recognition result is maintained as the current stable output result, and the recognition results of subsequent frames within the first timing window are obtained. If at least two consecutive frames have the same and non-empty recognition results within the first timing window, the stable recognition result is replaced by the same recognition result as the new stable output result. If no at least two consecutive frames have the same and non-empty recognition results when the first timing window expires, the current stable output result is cleared.

[0010] Preferably, the duration of the first timing window is determined based on the video stream frame rate, and the duration ranges from 400 milliseconds to 1200 milliseconds.

[0011] Preferably, the preset difference tolerance range is: the current frame recognition result and the stable recognition result differ in at most one character position, while the remaining character positions are completely the same, and the confidence level of the current frame recognition result is not less than 90% of the confidence level of the stable recognition result.

[0012] Preferably, when the current frame recognition result is not empty, its confidence level is not lower than the second preset threshold, and the current frame recognition result is different from the stable recognition result, the recognition results of the subsequent N frames are obtained, and it is determined whether there are K consecutive frames with the same recognition result between the current frame recognition result and the recognition results of the subsequent N frames, where N and K are both integers greater than or equal to 2 and K≤N;

[0013] If it exists, the same recognition result for K consecutive frames will be output as a new stable output result;

[0014] If it does not exist, the stable identification result remains unchanged.

[0015] Preferably, N is between 3 and 8, and K is 2 or 3.

[0016] Preferably, comparing the current frame recognition result with the historical recognition result includes:

[0017] The license plate number string in the current frame recognition result is compared with the license plate number string in the stable recognition result one by one according to the character position, and the proportion of the number of successfully matched characters to the total number of characters in the string is used as the matching rate.

[0018] When the matching rate is greater than or equal to the fourth preset threshold, it is determined that the current frame recognition result is the same as the stable recognition result. The value range of the fourth preset threshold is 80% to 100%.

[0019] Preferably, the method further includes a step of smoothing the confidence level of the stable output result:

[0020] Calculate the smoothed confidence level using the following formula:

[0021]

[0022] in, The confidence level of the recognition result for the current frame. This represents the confidence level of the currently output stable identification results. For smoothing coefficients, The value range is from 0.2 to 0.5;

[0023] When the smoothed confidence level is lower than a third preset threshold, a reconfirmation process is triggered. The value of the third preset threshold ranges from 0.5 to 0.7.

[0024] The present invention also provides a stable output system for license plate recognition results based on continuous frame confirmation, comprising:

[0025] The image acquisition module is used to acquire the current frame image in the video stream;

[0026] The license plate recognition module is used to perform license plate recognition on the current frame image and obtain the current frame recognition result;

[0027] A caching module is used to store historical recognition results, which include stable recognition results output before the current moment and recognition results of the most recent several frames;

[0028] The state determination module is used to compare the current frame recognition result with the stable recognition results in the historical recognition results, and determine whether the current frame recognition result meets the instant output condition, wherein the instant output condition is:

[0029] The current frame identification result is the same as the stable identification result, or the confidence level of the current frame identification result is greater than or equal to the first preset threshold and the difference between the current frame identification result and the stable identification result is within a preset difference tolerance range;

[0030] The timing stabilization module is used to start a first timing window when the current frame recognition result is empty or its confidence is lower than a second preset threshold. Within the first timing window, the stable recognition result is maintained as the current stable output result, and the recognition results of subsequent frames within the first timing window are obtained. If at least two consecutive frames have the same and non-empty recognition results within the first timing window, the stable recognition result is replaced by the same recognition result as the new stable output result. If no consecutive at least two frames have the same and non-empty recognition results when the first timing window expires, the current stable output result is cleared.

[0031] The output module is used to output the stable output result.

[0032] Preferably, the time-series stabilization module is further configured to: when the current frame recognition result is not empty, its confidence level is not lower than the second preset threshold, and the current frame recognition result is different from the stable recognition result, obtain the recognition results of the subsequent N frames of images, and determine whether there are K consecutive frames of recognition results that are the same in the current frame recognition result and the recognition results of the subsequent N frames of images, where N and K are both integers greater than or equal to 2 and K≤N;

[0033] If it exists, the identical recognition results of K consecutive frames are sent to the output module as new stable output results;

[0034] If it does not exist, the stable identification result remains unchanged.

[0035] Preferably, the status determination module includes:

[0036] The string comparison unit is used to compare the license plate number string in the current frame recognition result with the license plate number string in the stable recognition result one by one according to the character position.

[0037] The matching rate calculation unit is used to calculate the proportion of successfully matched characters to the total number of characters in the string, and thus obtain the matching rate.

[0038] The determination unit is used to determine that the current frame recognition result is the same as the stable recognition result when the matching rate is greater than or equal to a fourth preset threshold. The value range of the fourth preset threshold is 80% to 100%.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] By combining real-time output conditions with continuous frame confirmation, the recognition result is directly output when it has high confidence and is consistent with historical results or has only a tolerable deviation. When the result is abnormal or has high uncertainty, it is temporarily not output and is transferred to time-series verification. This forms a complete processing link from single-frame judgment to multi-frame confirmation, realizes the self-consistency of recognition results in the time dimension, effectively suppresses the interference of single-frame false detection, missed detection and recognition jump on the output stability, and improves the continuity and predictability of system output.

[0041] This invention maintains the confirmed results under abnormal conditions and continuously monitors the consistency of subsequent frames within a window period, enabling the system to maintain effective output even when the license plate is briefly obscured or the recognition confidence level drops temporarily. At the same time, through a continuous frame confirmation mechanism, it switches to the new result in a timely manner when the recognition result changes in a real way. This establishes an adaptive balance between maintaining stable output and responding to real changes, and achieves long-term stable operation of the system under video stream input conditions, avoiding frequent output changes or information interruptions caused by single-frame quality fluctuations.

[0042] This invention, through the combined effect of confidence smoothing and reconfirmation triggering conditions, enables the system to determine whether reconfirmation is needed based on the long-term trend of recognition confidence rather than a single value. Combined with the simultaneous constraint on the number of character differences and the confidence ratio within the difference tolerance range, it achieves fine-grained control over character-level recognition deviation. Ultimately, this allows the system to maintain steady-state output accuracy while possessing adaptive capabilities to dynamic scenarios, achieving robustness and reliability in practical engineering terms. Attached Figure Description

[0043] Figure 1 A flowchart illustrating a method for stable output of license plate recognition results based on continuous frame confirmation, provided in an embodiment of the present invention;

[0044] Figure 2 This is a system structure block diagram according to an embodiment of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Figure 1 The diagram shows a flowchart of a method for stable output of license plate recognition results based on continuous frame confirmation provided by an embodiment of the present invention. Figure 2 A system structure block diagram according to an embodiment of the present invention is shown. Figure 2 As shown, the system comprises six main components: an image acquisition module, a license plate recognition module, a buffer module, a status judgment module, a timing stabilization module, and an output module.

[0047] The image acquisition module is used to acquire the current frame image in the video stream. Its input end is connected to the data output interface of the camera, and its output end is connected to the license plate recognition module.

[0048] After receiving the current frame image, the license plate recognition module performs license plate detection and optical character recognition, and outputs the recognition result of the current frame and its confidence level.

[0049] The caching module uses a circular buffer to store historical recognition results, which include stable recognition results that have been output before the current moment as well as recognition results from the most recent few frames.

[0050] The status judgment module is connected to the license plate recognition module and the cache module respectively. It is used to read the recognition result of the current frame and compare it with the stable recognition result to determine whether the conditions for immediate output are met.

[0051] The timing stabilization module is connected to the status judgment module. Based on the control signal output by the status judgment module, it performs timing of the first timing window, acquisition of subsequent frame recognition results and consistency judgment of consecutive frames, as well as switching, holding or clearing operations of stable output results.

[0052] The output module connects to the timing stabilization module and outputs the final stable output result to an external display device or business system.

[0053] The above modules can be implemented using FPGA logic circuits, embedded processors running software programs, or a combination of hardware and software. The specific implementation does not constitute a limitation on the present invention. The working process of each module is described in detail below with reference to various embodiments.

[0054] Embodiment 1: This embodiment takes the parking lot exit scenario as an example to illustrate the immediate output process under normal illumination conditions, as well as the abnormality retention and recovery process when the license plate is temporarily occluded. The camera collects the video stream of the exit passage at a frame rate of 30 frames per second, and the system sequentially performs license plate detection and character recognition on each frame of image.

[0055] After the image acquisition module acquires the current frame image, the license plate recognition module recognizes the current frame image and outputs the current frame recognition result. The recognition result at least includes the license plate number string and its corresponding confidence level. The caching module pre-stores the stable recognition results that have been output before the current moment. Assume that the currently output stable recognition result is "Jing A12345" with a confidence level of 0.92.

[0056] The state judgment module reads the current frame recognition result from the license plate recognition module, and simultaneously reads the stable recognition result from the caching module to compare the two.

[0057] If the current frame recognition result is "Jing A12345", which is completely identical to the stable recognition result, the state judgment module determines that the "identical result" branch in the immediate output condition is satisfied, and sends an immediate output instruction to the timing stabilization module. The timing stabilization module outputs the recognition result via the output module, and updates the stable recognition result in the caching module at the same time.

[0058] If the current frame recognition result is "Jing A12346", which differs from the stable recognition result "Jing A12345" only in the 7th character, and the confidence level of the current frame recognition result is 0.88. The state judgment module first calculates the confidence ratio in the following manner: , the ratio is not less than 0.90 (i.e., 90%). Since there is only one character position difference between the two results and the confidence ratio meets the requirement, the state judgment module determines that the "preset difference tolerance range" branch in the immediate output condition is satisfied, and outputs "Jing A12346" as the stable output result via the output module.

[0059] The tolerance range is set as follows: In license plate recognition under long-distance or low-illumination conditions, trailing characters (especially numbers and letters, such as "5" and "6", "8" and "B") are prone to confusion. A difference of at most one character is regarded as a reasonable fluctuation range of the same license plate in engineering, and allowing such differences to be output directly can avoid unnecessary confirmation delay. A difference of more than one character indicates that the license plates may be different, and a confirmation process must be entered.

[0060] If the recognition result of the current frame is empty, that is, the license plate recognition module does not detect any license plate area in the current frame, or the confidence of the recognition result of the current frame is lower than the second preset threshold, the state judgment module determines that the recognition result of the current frame does not meet the immediate output condition, and sends an abnormal signal to the timing stabilization module. In this embodiment, the second preset threshold is set to 0.50.

[0061] After receiving the abnormal signal, the timing stabilization module starts the first timing window. The duration of the first timing window is determined according to the frame rate of the video stream. In this embodiment, the frame rate of the video stream is 30fps, and the duration of the first timing window is set to 800 milliseconds, which corresponds to approximately 24 frames of images. During this window, the timing stabilization module keeps "Jing A12345" as the current stable output result unchanged, and the output module continues to output this result to the outside. At the same time, the timing stabilization module continuously acquires the recognition results of subsequent frames within the window.

[0062] If "Jing B67890" is continuously recognized in the 3rd frame and the 4th frame within the timing window, and the recognition results of both frames are not empty, the timing stabilization module determines that there is a situation where the recognition results of at least two consecutive frames are the same and not empty, replaces "Jing A12345" with "Jing B67890" as the new stable output result, outputs it via the output module, and updates the stable recognition result in the cache module.

[0063] If when the first timing window expires (that is, after 800 milliseconds), the timing stabilization module still does not detect the situation where the recognition results of at least two consecutive frames are the same and not empty, it is determined that the current license plate has left the camera area or the recognition conditions continue to deteriorate, the timing stabilization module clears the current stable output result, and the output module stops outputting any license plate information.

[0064] The selection basis for the 800-millisecond duration of the first timing window in this embodiment is as follows: In the parking lot exit scenario, the time for a vehicle to pass through the camera area is usually 2 to 5 seconds, and the effective residence time of the license plate in the picture is sufficient to support at least one round of consecutive frame confirmation. The 800-millisecond window can filter out single-frame recognition noise and missed detections caused by instantaneous occlusion, and will not retain old license plate information after the vehicle has left due to an excessively long window. This duration can be adjusted in the range of 400 milliseconds to 1200 milliseconds according to the actual scenario. A shorter window can be used when the frame rate is high to speed up the response, and a longer window is required when the frame rate is low to ensure a sufficient number of confirmation frames.

[0065] The calculation method of the confidence is as follows: The license plate recognition model outputs a confidence for each character of the license plate number , where , is the total number of characters in the license plate number. The overall confidence of the license plate is obtained by taking the arithmetic average of the confidence of each character:

[0066]

[0067] For the case where the recognition result is empty, the system directly records the overall confidence as zero.

[0068] Second Embodiment: In this embodiment, taking the urban road checkpoint scenario as an example, it is described how the system realizes stable output through consecutive frame confirmation when the recognition result of the current frame is not empty, the confidence is not lower than the second preset threshold, but different from the current stable recognition result.

[0069] The camera collects lane video streams at a frame rate of 25 frames per second. Assume that the current output stable recognition result in the cache module is "Hu A88888", with a confidence of 0.95.

[0070] After the image acquisition module obtains the current frame, the license plate recognition module outputs the recognition result "Hu A88889" with a confidence of 0.85. After reading the result, the state judgment module sequentially performs the following judgments:

[0071] First step: judge whether the recognition result of the current frame is empty: it is not empty, proceed;

[0072] Second step: judge whether the confidence 0.85 is lower than the second preset threshold 0.50: it is not lower, proceed;

[0073] Third step: judge whether the recognition result of the current frame is the same as the stable recognition result "Hu A88888": they are different, proceed;

[0074] Fourth step: judge whether the preset difference tolerance range is satisfied: there is 1 character difference at the 8th character between "Hu A88889" and "Hu A88888", but the current frame confidence 0.85 divided by the stable confidence 0.95 is equal to 0.895, which is lower than 0.90, does not meet the requirement that the confidence is not less than 90% of the stable confidence, so the difference tolerance range is not satisfied.

[0075] Since none of the above four conditions meet the requirements for immediate output, the state judgment module sends a confirmation request signal to the timing stabilization module. After receiving the signal, the timing stabilization module starts the consecutive frame confirmation process in the normal switching scenario: obtain the recognition results of the subsequent N frames of images, and judge whether there are K consecutive frames with the same recognition result among the recognition results of the current frame and the subsequent N frames of images.

[0076] In this embodiment, N=5 and K=2 are preset inside the timing stabilization module. The system sequentially obtains the recognition results of the subsequent 5 frames of images, which are recorded as shown in the following table:

[0077]

[0078] The time sequence stabilization module checks one by one in the receiving order whether there are two consecutive frames with the same recognition result. The judgment process is as follows:

[0079] Compare the current frame with the first subsequent frame: "Hu A88889" is identical to "Hu A88889", forming two consecutive identical frames, and both are non-empty.

[0080] After the time sequence stabilization module confirms the consistency of these two consecutive frames, it takes "Hu A88889" as the new stable output result, outputs it via the output module, and updates the stable recognition result in the cache module at the same time. At this point, the subsequent frames are no longer checked, and the confirmation process ends in advance.

[0081] If the recognition results of the next 5 frames alternate between "Hu A88889" and "Hu A88888", and two consecutive frames cannot be completely identical, the time sequence stabilization module keeps the original stable recognition result "Hu A88888" unchanged, the output module continues to output this result, and the stable recognition result in the cache module remains unchanged.

[0082] In the above comparison process, the string comparison unit inside the state judgment unit compares the license plate number string in the current frame recognition result with the license plate number string in the stable recognition result character by character according to the character position, and the matching rate calculation unit counts the proportion of the number of successfully matched characters in the total number of characters of the string as the matching rate. For example, "Hu A88889" and "Hu A88888" have a total of 8 characters, except that the 8th character "9" is different from "8", the other 7 characters are the same, and the matching rate is:

[0083]

[0084] In this embodiment, the fourth preset threshold is set to 85%, and 87.5% is greater than 85%, so the state judgment module determines that "Hu A88889" and "Hu A88888" are the same result. The fourth preset threshold can be adaptively adjusted in the range of 80% to 100% according to the character accuracy distribution of the recognition model. A model with higher character accuracy can adopt a higher threshold (such as 90%), and a model with lower character accuracy can appropriately reduce the threshold (such as 80%) to achieve the optimal comparison effect.

[0085] The values of N and K affect the reliability and response delay of consecutive frame confirmation. If the value of N is too small (such as 2 or 3), the confirmation samples are insufficient, and occasional identical false detections may be misjudged as valid switches; if the value of N is too large (such as 10 or more), the confirmation period is too long, resulting in response lag. In this embodiment, the configuration of N=5 and K=2 can complete a round of confirmation in about 200 milliseconds at a frame rate of 25fps, which has been experimentally verified to achieve a good balance between reliability and real-time performance. The preferred range of N is 3 to 8, and the preferred range of K is 2 or 3.

[0086] The difference between this embodiment and the first embodiment lies in the cause for triggering continuous frame confirmation. In the first embodiment, the triggering cause is that the current frame is empty or of low confidence, which belongs to an abnormal scenario. At this time, the system starts the first timing window and maintains the old result. In this embodiment, the triggering cause is that the current frame has a recognition result but is inconsistent with the new result, which belongs to a normal switching scenario. At this time, the system uses the presence of K consecutive identical frames in the subsequent N frames as the switching basis. The processing logic of both scenarios is uniformly managed by the timing stabilization module, but different timing and counting strategies are adopted to meet the engineering requirements of different scenarios.

[0087] Embodiment 3: This embodiment takes the expressway toll station scenario as an example to describe the specific implementation of confidence smoothing processing and reconfirmation triggering, as well as the behavior of the system when the confidence of multiple consecutive frames continues to be low.

[0088] In this scenario, while the output module outputs a stable recognition result, the system also performs smoothing processing on the confidence of the stably output result, so as to avoid the system frequently entering the confirmation or clearing process due to severe fluctuation of the single-frame confidence. The smoothing processing is executed by the confidence smoothing unit inside the state judgment module, and the smoothed confidence is calculated according to the following formula:

[0089]

[0090] wherein, is the confidence of the current frame recognition result currently output by the license plate recognition module, is the confidence of the currently output stable recognition result in the cache module, is the smoothing coefficient.

[0091] In this embodiment, the smoothing coefficient takes a value of 0.3. Assume that the currently output stable recognition result in the cache module is "Su C12345", and its confidence .

[0092] Case 1: The current frame recognition result is the same as the stable recognition result, both being "Su C12345", but the current frame confidence is significantly lower than the stable confidence. The confidence smoothing unit calculates the smoothed confidence:

[0093]

[0094] The smoothed confidence 0.855 is greater than the third preset threshold 0.60 set in this embodiment, so the state judgment module does not trigger the reconfirmation process, the system normally outputs "Su C12345" as the stable output result, and at the same time the cache module updates the smoothed confidence 0.855 as the new confidence of this result.

[0095] Case 2: The recognition result of the current frame is different from the stable recognition result, the recognition result of the current frame is "Su C12346", and the confidence level . The confidence smoothing unit calculates as follows:

[0096]

[0097] The smoothed confidence level of 0.75 is still greater than the third preset threshold of 0.60, the system does not trigger reconfirmation, and continues to maintain "Su C12345" as the stable output result. The consideration for such processing is that the stable recognition result has a high historical confidence level of 0.90, and the single-frame low-confidence result is not sufficient to negate the confirmed stable output.

[0098] Case 3: The confidence levels of consecutive multiple frames are continuously low. Assume that the current frame confidence levels of subsequent frames are 0.40, 0.38, 0.42, 0.35, and 0.39 respectively, and the state judgment module calculates the smoothed confidence level frame by frame. The frame-by-frame recursion process is as follows:

[0099] After the first frame:

[0100]

[0101] After the second frame:

[0102]

[0103] After the third frame:

[0104]

[0105] At this time, 0.5733 is already lower than the third preset threshold of 0.60, and the state judgment module sends a reconfirmation trigger signal to the temporal stability module. After receiving the signal, the temporal stability module starts a consecutive frame confirmation process to re-evaluate whether the current license plate is still valid. If subsequent frames restore high-confidence recognition results (for example, the recognition results of 2 consecutive frames are the same and the confidence levels are both higher than 0.80), the system updates the stable output result after completing the confirmation. If subsequent frames continue to have low confidence levels and no qualified recognition result is obtained after the first timing window expires, the temporal stability module clears the stable output result.

[0106] In this embodiment, the smoothing coefficient is selected according to the following principles: A larger value means that the confidence level of the current frame has a greater impact on the smoothing result, the system responds faster but has weaker anti-noise capability; A smaller value means that the historical confidence level has a greater impact, the system is more stable but the response is more lagged. Verified by experiments, A value between 0.2 and 0.5 achieves a good balance between response speed and noise immunity. The third preset threshold is between 0.5 and 0.7. If the value is below 0.5, the reconfirmation trigger is too sluggish and may output low-confidence results for a long time; if the value is above 0.7, the trigger is too sensitive and normal fluctuations in ambient light may cause unnecessary reconfirmation processes.

[0107] The three embodiments described above address three aspects: "real-time output and exception retention," "continuous frame confirmation in normal switching scenarios," and "confidence smoothing and re-confirmation triggering," respectively, combining... Figure 1 and Figure 2 The system structure shown illustrates the technical solution of this invention in detail. The modules work collaboratively in different embodiments to achieve stable output of license plate recognition results based on continuous frame confirmation. Those skilled in the art should understand that the parameter values ​​in each embodiment are illustrative and can be adaptively adjusted according to specific scenarios and recognition model performance in practical applications, all of which fall within the protection scope of this invention.

Claims

1. A method for stable output of license plate recognition results based on continuous frame confirmation, characterized in that, Includes the following steps: The current frame image in the video stream is acquired, license plate recognition is performed on the current frame image to obtain the current frame recognition result, and historical recognition results are acquired, including stable recognition results output before the current moment and recognition results of the most recent several frames; The current frame recognition result is compared with the historical recognition result. When the current frame recognition result is the same as the stable recognition result in the historical recognition result, or when the confidence level of the current frame recognition result is greater than or equal to the first preset threshold and the difference between the current frame recognition result and the stable recognition result is within the preset difference tolerance range, the current frame recognition result is output as a stable output result. When the current frame recognition result is empty or its confidence level is lower than the second preset threshold, the first timing window is started. Within the first timing window, the stable recognition result is maintained as the current stable output result, and the recognition results of subsequent frames within the first timing window are obtained. If at least two consecutive frames have the same and non-empty recognition results within the first timing window, the stable recognition result is replaced by the same recognition result as the new stable output result. If no at least two consecutive frames have the same and non-empty recognition results when the first timing window expires, the current stable output result is cleared.

2. The method according to claim 1, characterized in that, The duration of the first timing window is determined based on the video stream frame rate, and the duration ranges from 400 milliseconds to 1200 milliseconds.

3. The method according to claim 1, characterized in that, The preset difference tolerance range is as follows: the current frame recognition result and the stable recognition result differ in at most one character position, while the remaining character positions are completely the same, and the confidence level of the current frame recognition result is not less than 90% of the confidence level of the stable recognition result.

4. The method according to claim 1, characterized in that, When the current frame recognition result is not empty, its confidence level is not lower than the second preset threshold, and the current frame recognition result is different from the stable recognition result, the recognition results of the subsequent N frames are obtained, and it is determined whether there are K consecutive frames with the same recognition result between the current frame recognition result and the recognition results of the subsequent N frames, where N and K are both integers greater than or equal to 2 and K≤N; If it exists, the same recognition result for K consecutive frames will be output as a new stable output result; If it does not exist, the stable identification result remains unchanged.

5. The method according to claim 4, characterized in that, N takes values ​​from 3 to 8, and K takes values ​​of 2 or 3.

6. The method according to claim 1, characterized in that, The step of comparing the current frame recognition result with the historical recognition result includes: The license plate number string in the current frame recognition result is compared with the license plate number string in the stable recognition result one by one according to the character position, and the proportion of the number of successfully matched characters to the total number of characters in the string is used as the matching rate. When the matching rate is greater than or equal to the fourth preset threshold, it is determined that the current frame recognition result is the same as the stable recognition result. The value range of the fourth preset threshold is 80% to 100%.

7. The method according to claim 1, characterized in that, It also includes a step of smoothing the confidence level of the stable output results: Calculate the smoothed confidence level using the following formula: in, The confidence level of the recognition result for the current frame. This represents the confidence level of the currently output stable identification results. For smoothing coefficients, The value range is from 0.2 to 0.5; When the smoothed confidence level is lower than a third preset threshold, a reconfirmation process is triggered. The value of the third preset threshold ranges from 0.5 to 0.

7.

8. A system for stable output of license plate recognition results based on continuous frame confirmation, characterized in that, include: The image acquisition module is used to acquire the current frame image in the video stream; The license plate recognition module is used to perform license plate recognition on the current frame image and obtain the current frame recognition result; A caching module is used to store historical recognition results, which include stable recognition results output before the current moment and recognition results of the most recent several frames; The state determination module is used to compare the current frame recognition result with the stable recognition results in the historical recognition results, and determine whether the current frame recognition result meets the instant output condition, wherein the instant output condition is: The current frame identification result is the same as the stable identification result, or the confidence level of the current frame identification result is greater than or equal to the first preset threshold and the difference between the current frame identification result and the stable identification result is within a preset difference tolerance range; The timing stabilization module is used to start a first timing window when the current frame recognition result is empty or its confidence is lower than a second preset threshold. Within the first timing window, the stable recognition result is maintained as the current stable output result, and the recognition results of subsequent frames within the first timing window are obtained. If at least two consecutive frames have the same and non-empty recognition results within the first timing window, the stable recognition result is replaced by the same recognition result as the new stable output result. If no consecutive at least two frames have the same and non-empty recognition results when the first timing window expires, the current stable output result is cleared. The output module is used to output the stable output result.

9. The system according to claim 8, characterized in that, The time-series stabilization module is further configured to: when the current frame recognition result is not empty, its confidence level is not lower than the second preset threshold, and the current frame recognition result is different from the stable recognition result, obtain the recognition results of the subsequent N frames of images, and determine whether there are consecutive K frames of recognition results that are the same in the current frame recognition result and the recognition results of the subsequent N frames of images, where N and K are both integers greater than or equal to 2 and K≤N; If it exists, the identical recognition results of K consecutive frames are sent to the output module as new stable output results; If it does not exist, the stable identification result remains unchanged.

10. The system according to claim 8, characterized in that, The status determination module includes: The string comparison unit is used to compare the license plate number string in the current frame recognition result with the license plate number string in the stable recognition result one by one according to the character position. The matching rate calculation unit is used to calculate the proportion of successfully matched characters to the total number of characters in the string, and thus obtain the matching rate. The determination unit is used to determine that the current frame recognition result is the same as the stable recognition result when the matching rate is greater than or equal to a fourth preset threshold. The value range of the fourth preset threshold is 80% to 100%.

Images