Unmanned vehicle braking control method based on image recognition

Abstract

The invention relates to the technical field of intelligent driving control, and discloses an unmanned vehicle braking control method based on image recognition, which comprises the following steps of: for depth uncertainty caused by visual degradation and local camouflage, proposing to calculate a time sequence brightness difference according to pixels; and determining a pixel reliability threshold value based on the median and the median absolute deviation. Carrying out statistics on region-level visual reliability in a target region, and carrying out confidence correction and adaptive time sequence smoothing on the pixel-level depth variance by using the region-level visual reliability; and a visual credible risk distance is calculated according to the corrected uncertainty, a safety braking distance is determined according to the visual credible risk distance, the vehicle speed and the response time, and dynamic smooth braking force output is generated according to the risk difference, so that the robustness and traceability of a visual driving braking decision are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent driving control, in particular to an unmanned vehicle braking control method based on image recognition. BACKGROUND

[0002] Current unmanned vehicle braking control based on vehicle-mounted vision relies on image recognition and distance estimation of the front target, but there are many short-term and local visual degradation problems in actual road scenes, which makes the reliability of braking decision based on vision insufficient; these problems include local occlusion of the lens (such as mud drops, rain, snow or splashes), local pixel saturation caused by strong light transient glare or mirror reflection, and local recognition misdirection caused by small-range countermeasures patches or reflective materials; in addition, single-frame or shallow feature-based detection methods often cannot distinguish between real targets and disguised areas when encountering the above-mentioned abnormalities, thereby producing the risk of false triggering or missing triggering braking; The common practice is to use radar or lidar as a redundant sensor to supplement the visual deficiency, however these sensors also have limitations in close-range, low-reflectivity or multipath interference environments, and the confidence fusion of different sensor data in the prior art usually uses empirical weights or black-box learning fusion strategies, lacking a provable statistical confidence definition method; the current judgment of braking triggering is mostly based on fixed thresholds or simple safety distance formulas, when the visual distance estimation uncertainty is large, the fixed threshold strategy cannot effectively avoid false braking, nor can it fully utilize visual information to improve responsiveness when reliable vision is available; In summary, the prior art has not yet provided a complete solution that can robustly statistically determine pixel-level reliability, quantify regional visual reliability, and systematically integrate the reliability into depth uncertainty correction and timing smoothing without relying on a large number of training parameters or empirical weights, and finally generate braking trigger and smooth braking force output. SUMMARY

[0003] The present application provides an unmanned vehicle braking control method based on image recognition, which promotes the solution to the problems mentioned in the background art.

[0004] The present application provides the following technical solution: an unmanned vehicle braking control method based on image recognition, comprising: Collecting vehicle-mounted camera images at consecutive time points, calculating the timing brightness difference of consecutive image frames pixel by pixel, and determining the pixel reliability threshold based on the median and the median absolute deviation; Generating a pixel trust mask according to the pixel reliability threshold and calculating the regional reliability in the target detection area; Selecting a trusted pixel set according to the pixel trust mask, extracting its parallax information, and calculating the target depth in combination with the camera focal length and baseline parameters; performing variance calculation on the target depth data and correcting depth uncertainty; calculating risk distance based on the corrected depth uncertainty, and determining safe braking distance combined with vehicle speed and system reaction time; generating braking trigger signal according to the risk distance and the safe braking distance; calculating braking force output according to the braking trigger signal and risk distance difference, and performing dynamic smoothing adjustment.

[0005] Optionally, the method further comprises: collecting vehicle-mounted camera images at continuous time points, calculating time-series brightness difference of each pixel in continuous image frames, and determining pixel reliability threshold based on median and median absolute deviation, comprising: obtaining brightness values of each pixel in a target region at any two continuous time points, and recording brightness of the pixel in the current frame and the previous frame respectively; calculating absolute difference of the brightness values of the pixel in the current frame and the previous frame as time-series brightness difference of the pixel; calculating median and median absolute deviation of time-series brightness difference of all pixels in a specified image effective sampling region; setting the pixel reliability threshold as the median plus three times of a constant used for converting the median absolute deviation into a standard deviation order multiplied by the median absolute deviation.

[0006] Optionally, the method further comprises: generating pixel reliable mask according to the pixel reliability threshold, and calculating region reliability in the target detection region, comprising: outputting binary reliable identification for each pixel based on comparison between the time-series brightness difference of the pixel and the pixel reliability threshold, marking the pixel as reliable pixel when the time-series brightness difference of the pixel is less than or equal to the threshold, otherwise marking the pixel as distorted pixel to be removed; calculating arithmetic mean value of reliable identification of all pixels in the region surrounded by the region bounding box obtained by target detection, and taking the obtained mean value as region-level visual reliability of the target region in the current frame; taking the binary mask and the region-level visual reliability as input data for subsequent depth estimation and uncertainty correction.

[0007] Optionally, the method further comprises: selecting reliable pixel set according to the pixel reliable mask, extracting disparity information of the reliable pixel set, and calculating target depth combined with focal length of the camera and baseline parameters, comprising: confirming the pixels marked as reliable as reliable pixel set; extracting disparity measurement value of each pixel in the reliable pixel set; performing geometric inverse calculation on the focal length parameter of the camera, the baseline length and the disparity measurement value of the pixel, and obtaining spatial depth estimation value of the pixel according to the operation sequence of focal length multiplied by baseline divided by disparity; The camera focal length and binocular baseline distance are known parameters obtained by device calibration, and the known parameters are geometrically calculated according to the measured parallax to output the pixel-level target depth.

[0008] Optionally, the variance calculation on the target depth data and the correction of the depth uncertainty include: The average depth of the target region is calculated in the trusted pixel set, and the square of the difference between the depth of each pixel and the average depth is averaged to obtain the original depth variance; The region-level visual reliability of the target region is obtained; The original depth variance is confidence-corrected by using the region-level visual reliability as a correction factor, specifically by first taking the square root of the original depth variance to obtain the depth standard deviation, and then dividing the depth standard deviation by the sum of the region-level visual reliability and a small positive number to obtain the weighted depth uncertainty; The small positive number is used to prevent division by zero.

[0009] Optionally, the variance calculation on the target depth data and the correction of the depth uncertainty further include: The weighted depth uncertainty is subjected to a temporal smoothing correction, and the weighted depth uncertainty of the current frame is linearly weighted with the smoothed depth uncertainty of the previous frame by an adaptive smoothing weight to synthesize the temporal smoothed depth uncertainty of the current frame; The adaptive smoothing weight is obtained by transforming the region-level visual reliability by a logical function, and the input of the transformation is the difference between the region-level visual reliability and a preset reliability balance threshold multiplied by a smoothing sensitivity coefficient, and the output value of the transformation is between zero and one, which is used as the mixing weight of the new observation and the historical smoothing value of the current frame; The smoothing sensitivity coefficient and the reliability balance threshold are preset parameters according to the requirements.

[0010] Optionally, the risk distance is calculated based on the corrected depth uncertainty, and the safe braking distance is determined in combination with the vehicle speed and the system reaction time, including: The risk distance is calculated by subtracting the confidence control coefficient multiplied by the temporal smoothed depth uncertainty from the target average depth; The current position and speed of the vehicle and the reaction time parameter of the vehicle control system are obtained; According to the vehicle speed and the reaction time and the maximum allowed braking acceleration specified by the system, the safe braking distance is determined by multiplying the speed and the reaction time first, and then adding the speed squared divided by twice the maximum braking acceleration; The risk distance and the safe braking distance are used as comparison quantities for subsequent braking trigger judgment.

[0011] Optionally, the braking trigger signal is generated according to the risk distance and the safe braking distance, including: Obtain the visually reliable risk distance and the corresponding safe braking distance; When the visually credible risk distance is less than or equal to the safe braking distance, a braking trigger signal is generated and output; otherwise, no braking trigger signal is generated. The braking trigger signal serves as a Boolean trigger for subsequent braking force calculation and execution.

[0012] Optionally, the step of calculating the braking force output based on the braking trigger signal and the risk distance difference, and performing dynamic smooth adjustment, includes: When the braking trigger signal is valid, the difference between the safe braking distance and the visually credible risk distance is calculated as the risk distance difference; Based on the risk distance difference, it is mapped to the actual braking force output using a smooth approximation function. The form of the mapping function is that the braking force output is zero when the risk distance difference is zero, and it monotonically increases with the increase of the risk distance difference until the limit value of the vehicle's maximum braking force. When the brake trigger signal is invalid, the braking force output is zero, and the monitoring status is maintained.

[0013] The present invention has the following beneficial effects: 1. By using pixel temporal difference as the basic measure, significant changes in pixels within a short temporal scale can be directly reflected, thereby quantifying sudden contamination (such as mud droplets, water droplets, or short-term strong light spots) at the pixel level; by using the median and MAD as statistics to replace the mean and variance, the impact of single-point extreme values ​​on the threshold can be significantly reduced, making the threshold stable even in noisy environments; the threshold is an intra-frame adaptive calculation result, which can dynamically adjust the pixel screening criteria under different scenarios, avoiding the risk of fixed thresholds failing under environmental changes; 2. By generating a pixel-based reliable mask based on a pixel reliability threshold and statistically analyzing regional-level visual reliability in the target detection area, confidence transfer and aggregation from the pixel level to the regional level are achieved. After the pixel-level mask removes local anomalies, the regional-level reliability serves as a quantitative representation of the overall target reliability, reflecting the overall distribution of pixel reliability at the target level. This regional-level metric can be used to determine whether the target is occluded, camouflaged, or locally damaged, thus providing explicit confidence input for downstream depth estimation, tracking, and decision-making. Compared to relying solely on pixel counts or single matching scores, the averaged statistics of regional-level reliability reduce the impact of noise points and reflect spatial coherence, which helps identify false detections caused by local adversarial patches or small-scale reflections. Furthermore, introducing regional reliability as an input parameter into subsequent algorithms allows the system to automatically reduce its dependence on visual ranging results when the overall visual confidence decreases, thereby implementing a vision-control safety degradation strategy and reducing the risk of false triggering. 3. By performing stereo matching only on pixels deemed trustworthy, the false matching rate and depth bias introduced by pseudo-disparity can be significantly reduced, thereby improving the effective pixel ratio and statistical stability of single-frame depth estimation. Utilizing the intra-camera participation baseline relationship and adopting an analytical inverse calculation method, following the principles of geometric optics, the output depth estimate has a traceable physical meaning, facilitating quantitative analysis of error sources. This approach avoids dependence on semantic dimensions or learning models, thus maintaining consistent inference logic even in unseen objects or abnormal scenes. Finally, depth estimation based on trustworthy pixels provides a more reliable input for subsequent uncertainty quantification and braking risk assessment, improving the overall accuracy and interpretability of braking determination. 4. By statistically analyzing the sample variance of the reliable pixel depth, the dispersion of depth measurement can be directly obtained. This quantitative result characterizes the instantaneous reliability of depth estimation. By using the regional visual reliability as a correction term to amplify or attenuate the depth variance, a confidence adjustment mechanism corresponding to visual quality is formed, thereby avoiding excessive trust in the depth result under low visual confidence conditions. This mechanism achieves adaptive correction to visual noise, automatically increasing the safety margin when visual degradation occurs, while maintaining a small estimation uncertainty when visual confidence is high. This adaptive uncertainty correction helps reduce erroneous braking decisions caused by visual errors, while providing the controller with a quantified risk boundary so that subsequent physical-based braking calculations can adopt an appropriate safety margin. 5. By linearly mixing the uncertainty of the current frame with the smoothing value of the previous frame using adaptive weights, unnecessary short-term jitter can be avoided from being transmitted to the control end when the visual state changes abruptly, and new reliable observations can be quickly reflected when the visual state stabilizes and recovers. The adaptive weights are obtained by logistic transformation based on the regional reliability, so that the weights can be continuously and differentiably adjusted with the overall visual confidence. This continuous weight change suppresses the frequent switching caused by the threshold effect. In terms of systems engineering, this temporal smoothing mechanism reduces the braking jitter or passenger discomfort caused by input fluctuations in the controller, and achieves automatic switching between smoothness and sensitivity based on real-time observation without relying on historical experience parameters, thereby improving the temporal robustness of the entire perception-decision channel. 6. By subtracting the uncertainty multiplied by the confidence control coefficient from the average depth as the risk distance, a conservative distance estimate considering the confidence interval can be numerically expressed. This design directly links statistical confidence with geometric distance, making risk assessment based on both ranging and measurement confidence. Using the confidence control coefficient as a parameterization method, the required confidence coverage can be adjusted according to the system's security strategy, thereby achieving an adjustable risk boundary from aggressive to conservative. This method overcomes the failure problem of fixed distance thresholds in visually uncertain environments, making the triggering logic have controllable tolerance to ranging noise and reducing the probability of misjudgment and omission. 7. By multiplying the vehicle's current speed by the system response delay and adding a braking distance term that depends on the vehicle's maximum braking acceleration, the resulting safe braking distance reflects the physical space required for the vehicle to decelerate and come to a stop from detection. This calculation based on vehicle operating conditions and control delay avoids blind decision-making based on purely visual benchmarks. The parameters used in this calculation process are all data that can be obtained from vehicle sensors or calibration, making the safe distance verifiable and engineering feasible, and providing quantifiable hard constraints for braking triggering, thus enhancing the provable safety boundary of the overall system. 8. By mapping the risk distance difference to a smooth braking force output with the vehicle's maximum braking force as the upper limit, the needs of ride comfort and safety response can be balanced. The mapping function with monotonically increasing and smooth characteristics continuously transforms the decision quantity into the execution quantity, which can avoid large abrupt changes in braking force caused by binary triggering, thereby reducing occupant discomfort and transient load on vehicle dynamics. The mapping function can be adjusted by parameters to control the slope and saturation value of the braking response, so that the output is gentle deceleration under slight risks and gradually approaches the system limit under greater risks to ensure braking effect. Limiting the output upper limit to the vehicle's maximum braking force ensures that the command is within the physically feasible range, avoids issuing unrealizable control commands, and thus reduces the coordination cost of abnormal states between the controller and the braking execution unit. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the basic process of the present invention. Detailed Implementation

[0015] 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.

[0016] Example 1, refer to Figure 1 A braking control method for unmanned vehicles based on image recognition, comprising: The vehicle-mounted camera acquires images at consecutive time points, calculates the temporal brightness difference for each pixel in the consecutive image frames, and determines the pixel reliability threshold based on the median and the median absolute deviation. Generate a pixel reliability mask based on the pixel reliability threshold and statistically analyze the region reliability in the target detection area; Based on the pixel trust mask, a set of trust pixels is selected, its parallax information is extracted, and the target depth is calculated by combining the camera focal length and baseline parameters. Variance is calculated on the target depth data, and depth uncertainty is corrected. The risk distance is calculated based on the corrected depth uncertainty, and the safe braking distance is determined by combining vehicle speed and system reaction time. Based on the risk distance and the safe braking distance, a braking trigger signal is generated; The braking force output is calculated based on the braking trigger signal and the risk distance difference, and then dynamically and smoothly adjusted.

[0017] The acquisition of vehicle-mounted camera images at consecutive time points, calculation of temporal brightness differences for each pixel of consecutive image frames, and determination of pixel reliability thresholds based on the median and median absolute deviation include: Obtain the brightness value of each pixel in the target area at any two consecutive time points, and record the brightness of the pixel in the current frame and the previous frame respectively; The absolute difference between the brightness values ​​of the pixel in the current frame and the previous frame is used as the temporal brightness difference of the pixel. Within the specified effective sampling area of ​​the image, calculate the median and median absolute deviation of the temporal brightness difference for all pixels; The pixel reliability threshold is set to be based on the median mentioned above, plus three times the constant used to convert the median absolute deviation to the order of standard deviation, multiplied by the median absolute deviation.

[0018] The step of generating a pixel-reliable mask based on a pixel reliability threshold and statistically analyzing the region reliability in the target detection region includes: Using the pixel temporal brightness difference and the pixel reliability threshold as a comparison benchmark, a binary reliable identifier is output for each pixel. When the temporal brightness difference of the pixel is less than or equal to the threshold, it is marked as a reliable pixel; otherwise, it is marked as a distorted pixel that needs to be removed. Within the bounding box of the region obtained by target detection, the arithmetic mean of the trust labels of all pixels in the region is calculated, and the resulting average value is used as the region-level visual reliability of the target region in the current frame. The aforementioned binary mask and region-level visual reliability serve as input data for subsequent depth estimation and uncertainty correction.

[0019] The step of selecting a set of reliable pixels based on the pixel reliable mask, extracting its disparity information, and calculating the target depth by combining the camera focal length and baseline parameters includes: The set of pixels marked as trustworthy is confirmed as the set of trustworthy pixels; Extract the disparity measurement value for each pixel in the set of trustworthy pixels; The focal length parameter of the camera, the binocular baseline length, and the disparity measurement value of the pixel are geometrically inversely calculated. The spatial depth estimate of the pixel is obtained by multiplying the focal length by the baseline and then dividing by the disparity. The camera focal length and binocular baseline distance are known parameters obtained from device calibration, and geometric inverse calculation is performed based on these known parameters and the measured disparity to output pixel-level target depth.

[0020] The step of calculating the variance of the target depth data and correcting for depth uncertainty includes: The average depth of the target region is calculated in the set of reliable pixels, and the squares of the differences between the depth of each pixel and the average depth are averaged to obtain the original depth variance. Obtain the regional-level visual reliability of the target area; The confidence correction of the original depth variance is performed using regional visual reliability as a correction factor. Specifically, the depth standard deviation is obtained by first taking the square root of the original depth variance, and then the weighted depth uncertainty is obtained by dividing the depth standard deviation by the sum of regional visual reliability and a small positive number. The small positive number is used to prevent division by zero.

[0021] The step of calculating the variance of the target depth data and correcting for depth uncertainty further includes: Temporal smoothing correction is applied to the weighted depth uncertainty. The weighted depth uncertainty of the current frame is linearly weighted and synthesized with the smoothed depth uncertainty of the previous frame according to adaptive weights, and then used as the temporal smoothed depth uncertainty of the current frame. The adaptive smoothing weights are obtained by transforming the regional visual reliability through a logistic function. The input of the transformation is the difference between the regional visual reliability and the preset reliability balance threshold multiplied by the smoothing sensitivity coefficient. The transformation output value is between zero and one, and serves as the mixed weight of the new observation and the historical smoothing value in the current frame. The smoothing sensitivity coefficient and reliability balance threshold are settable parameters as required.

[0022] The calculation of the risk distance based on the corrected depth uncertainty, and the determination of the safe braking distance in combination with vehicle speed and system reaction time, includes: The visually reliable risk distance is calculated by subtracting the confidence control coefficient from the target average depth and multiplying it by the time-smoothed depth uncertainty. Obtain the vehicle's current position and speed, and the vehicle control system's reaction time parameters; Based on the vehicle speed and reaction time, as well as the maximum permissible braking acceleration specified by the system, the safe braking distance is determined by multiplying the speed and reaction time by the speed squared and then dividing by twice the maximum braking acceleration. The risk distance and the safe braking distance are used as comparison parameters for subsequent braking trigger determination.

[0023] The step of generating a braking trigger signal based on the risk distance and the safe braking distance includes: Obtain the visually reliable risk distance and the corresponding safe braking distance; When the visually credible risk distance is less than or equal to the safe braking distance, a braking trigger signal is generated and output; otherwise, no braking trigger signal is generated. The braking trigger signal serves as a Boolean trigger for subsequent braking force calculation and execution.

[0024] The calculation of braking force output based on the braking trigger signal and the risk distance difference, and the dynamic smooth adjustment thereof, includes: When the braking trigger signal is valid, the difference between the safe braking distance and the visually credible risk distance is calculated as the risk distance difference; Based on the risk distance difference, it is mapped to the actual braking force output using a smooth approximation function. The form of the mapping function is that the braking force output is zero when the risk distance difference is zero, and it monotonically increases with the increase of the risk distance difference until the limit value of the vehicle's maximum braking force. When the brake trigger signal is invalid, the braking force output is zero, and the monitoring status is maintained.

[0025] Example 2: A braking control method for unmanned vehicles based on image recognition, further comprising: The acquisition of vehicle-mounted camera images at consecutive time points, calculation of temporal brightness differences for each pixel of consecutive image frames, and determination of pixel reliability thresholds based on the median and median absolute deviation include: Acquire image frames at every two consecutive time points and Each pixel in Brightness values ​​are statistically analyzed as follows: and ; Calculate image frames The timing brightness difference is: ; Set image frame The pixel reliability threshold is: ; in: : The defined effective sampling area of ​​the image; , : Calculate the median and the median absolute deviation (MAD) statistic; The approximate conversion factor from MAD to standard deviation is a statistical constant. Threshold factor, indicating the use of statistical methods. The principle is as follows: By using pixel temporal difference as the basic measure, significant changes in pixels within a short temporal scale can be directly reflected, thereby quantifying sudden contamination (such as mud droplets, water droplets, or short-term strong light spots) at the pixel level; by using the median and MAD as statistics instead of the mean and variance, the impact of single-point extreme values ​​on the threshold can be significantly reduced, making the threshold stable even in noisy environments; this threshold is an intra-frame adaptive calculation result, which can dynamically adjust the pixel selection criteria under different scenarios, avoiding the risk of fixed thresholds becoming invalid under environmental changes; The step of generating a pixel-reliable mask based on a pixel reliability threshold and statistically analyzing the region reliability in the target detection region includes: Acquire image frames Timing brightness difference and acquire image frames. Pixel reliability threshold ; The output binary mask is: ; in: Represents an image frame Medium pixel Reliable pixels; Indicates the distorted pixels that need to be removed; The statistical reliability of the target detection area is: ; in: The total number of pixels within the bounding box of the region. By generating a pixel-reliability mask based on a pixel reliability threshold and statistically analyzing the regional visual reliability in the target detection region, confidence transfer and aggregation from the pixel level to the regional level are achieved. After the pixel-level mask removes local anomalies, the regional reliability serves as a quantitative representation of the overall confidence of the target, reflecting the overall distribution of pixel reliability at the target level. This regional metric can be used to determine whether the target is occluded, camouflaged, or locally damaged, thus providing explicit confidence input for downstream depth estimation, tracking, and decision-making. Compared to relying solely on pixel counts or single matching scores, the averaged statistics of regional reliability reduce the impact of noise points and reflect spatial coherence, which helps identify false detections caused by local adversarial patches or small-scale reflections. Furthermore, introducing regional reliability as an input parameter into subsequent algorithms allows the system to automatically reduce its dependence on visual ranging results when the overall visual confidence decreases, thereby implementing a vision-control safety degradation strategy and reducing the risk of false triggering. The step of selecting a set of reliable pixels based on the pixel reliable mask, extracting its disparity information, and calculating the target depth by combining the camera focal length and baseline parameters includes: Extracting a set of reliable pixels ; Obtain a set of trusted pixels Each pixel parallax The parallax In binocular images captured by vehicle binocular cameras, the displacement difference of the same object point in the horizontal pixel coordinates can be directly measured using existing standard stereo matching algorithms. The pixel depth is calculated as follows: ; in: Camera focal length; Binocular baseline distance, specifically the center-to-center distance between the two cameras in a binocular camera system; Pixel-level target depth. By performing stereo matching only on pixels deemed trustworthy, the false matching rate and depth bias introduced by pseudo-disparity can be significantly reduced, thereby improving the effective pixel ratio and statistical stability of single-frame depth estimation. Utilizing the intra-camera participation baseline relationship and adopting an analytical inverse calculation method, following the principles of geometric optics, the output depth estimate has a traceable physical meaning, facilitating quantitative analysis of error sources. This approach avoids dependence on semantic dimensions or learning models, thus maintaining consistent inference logic even in unseen objects or anomalous scenes. Finally, depth estimation based on trustworthy pixels provides a more reliable input for subsequent uncertainty quantification and braking risk assessment, improving the overall accuracy and interpretability of braking determination. The step of calculating the variance of the target depth data and correcting for depth uncertainty includes: The average depth of the target area is statistically analyzed. ; in: Trusted pixel set; The number of pixels in the trusted pixel set; Pixel-level target depth; Calculate the original variance: ; Obtaining regional reliability ; Introducing a region-level confidence correction term: ; in: Regional-level visual reliability; The value is 0.01 to prevent division by zero of tiny positive numbers; Weighted depth variance. By statistically analyzing the sample variance of reliable pixel depths, the dispersion of depth measurements can be directly obtained. This quantification result characterizes the instantaneous reliability of depth estimation. Regional-level visual reliability is used as a correction term to amplify or attenuate the depth variance, forming a confidence adjustment mechanism corresponding to visual quality. This avoids over-reliance on depth results under low visual confidence conditions. This mechanism achieves adaptive correction to visual noise, automatically increasing the safety margin when visual degradation occurs, while maintaining small estimation uncertainty when visual confidence is high. This adaptive uncertainty correction helps reduce erroneous braking decisions caused by visual errors, while providing the controller with a quantified risk boundary so that subsequent physics-based braking calculations can use appropriate safety margins. The step of calculating the variance of the target depth data and correcting for depth uncertainty further includes: Smoothing correction is applied to the weighted depth variance: ; in: The temporally smoothed depth variance of the current frame; : Smooth depth variance of the previous frame; Weighted depth variance; Adaptive smoothing weights, with values ​​of: ; in: The smoothing sensitivity coefficient, which can be set to 5–10, is used to control the transition speed. A larger value results in a faster transition speed; a larger value is used if a faster response is required. If a smoother performance and tolerance for momentary jitter are required, choose a smaller value. ; : Reliability balance threshold, which can be taken as 0.5, when Increase, for a given , The value will decrease, resulting in a smoother but slower braking system response; this reduces braking fluctuations caused by momentary visual misjudgments, but may delay utilizing visual advantages as soon as vision becomes available again; when Decrease The larger the sensor, the faster the system can use the current visual information and the faster the braking response. However, if there is noise or adversarial interference in the vision, the risk of misjudgment increases. when High time, This indicates that the current frame is reliable, and the smoothing weight is reduced to preserve response speed; when When low, The presence of noise indicates that smoothing is enhanced to suppress flicker. By linearly mixing the uncertainty of the current frame with the smoothing value of the previous frame using adaptive weights, unnecessary short-term jitter can be avoided from being transmitted to the control end when visual state changes abruptly, and new reliable observations can be quickly reflected when visual stability is restored. The adaptive weights are obtained by logistic transformation based on regional reliability, so that the weights can be continuously and differentiably adjusted with the overall visual confidence. This continuous weight change suppresses frequent switching caused by the threshold effect. In terms of systems engineering, this temporal smoothing mechanism reduces braking jitter or passenger discomfort caused by input fluctuations in the controller, and achieves automatic switching between smoothing and sensitivity based on real-time observations without relying on historical experience parameters, thereby improving the temporal robustness of the entire perception-decision channel. The calculation of the risk distance based on the corrected depth uncertainty, and the determination of the safe braking distance in combination with vehicle speed and system reaction time, includes: Define the visually credible risk distance: ; in: Target average depth; : Depth variance after weighted and smoothed correction; The confidence control coefficient, with a value of 2, corresponds to a 95% confidence interval. A larger value indicates stronger safety redundancy but also a more conservative approach. By subtracting the uncertainty multiplied by the confidence control coefficient from the average depth as the risk distance, a conservative distance estimate considering the confidence interval can be numerically expressed. This design directly links statistical confidence with geometric distance, making risk assessment based on both ranging and measurement confidence. Using the confidence control coefficient as a parameterization method, the required confidence coverage can be adjusted according to the system's safety strategy, thereby achieving an adjustable risk boundary from aggressive to conservative. This method overcomes the failure problem of fixed distance thresholds in visually uncertain environments, making the triggering logic have controllable tolerance to ranging noise and reducing the probability of misjudgment and omission. Get vehicle speed Response time with vehicle control system ; The safe braking threshold is calculated as follows: ; in: The maximum permissible braking acceleration specified by the control system.

[0026] Safe braking distance. By multiplying the vehicle's current speed by the system response delay and adding a braking distance term that depends on the vehicle's maximum braking acceleration, the resulting safe braking distance reflects the physical space required for the vehicle to decelerate and come to a stop from detection. This calculation, based on vehicle operating conditions and control delay, avoids blind decision-making based on purely visual benchmarks. The parameters used in this calculation process are all data that can be obtained from vehicle sensors or calibration, making the safe distance verifiable and engineering feasible, and providing a quantifiable hard constraint for braking triggering, thus enhancing the provable safety boundary of the overall system. The step of generating a braking trigger signal based on the risk distance and the safe braking distance includes: Obtain risk distance Safety braking distance ; If it exists: ; Then output braking trigger signal ,otherwise .

[0027] The calculation of braking force output based on the braking trigger signal and the risk distance difference, and the dynamic smooth adjustment thereof, includes: When the brake trigger signal hour: Obtain risk distance Safety braking distance And calculate the risk distance difference: ; Smoothly adjust braking force output: ; in: Current braking force output; Maximum braking force of the vehicle; Braking force adjustment coefficient, which can be 2 to 3, the larger the value, the faster the braking response; like Then let The system maintains a monitoring state. By mapping the risk distance difference to a smooth braking force output with the vehicle's maximum braking force as the upper limit, the system can balance the needs of ride comfort and safety response. Employing a monotonically increasing mapping function with smoothing characteristics continuously transforms the decision quantity into the execution quantity, avoiding large abrupt changes in braking force caused by binary triggering, thereby reducing occupant discomfort and transient loads on vehicle dynamics. The mapping function, adjustable by parameters, controls the slope and saturation value of the braking response, resulting in gentle deceleration output under slight risks and gradually approaching the system limit under greater risks to ensure braking effectiveness. Limiting the output upper limit to the vehicle's maximum braking force ensures that the command is within the physically feasible range, avoiding the issuance of unrealizable control commands, and thus reducing the coordination costs of abnormal states between the controller and the braking execution unit.

[0028] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0029] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A braking control method for unmanned vehicles based on image recognition, characterized in that, include: The vehicle-mounted camera acquires images at consecutive time points, calculates the temporal brightness difference for each pixel in the consecutive image frames, and determines the pixel reliability threshold based on the median and the median absolute deviation. Generate a pixel reliability mask based on the pixel reliability threshold and statistically analyze the region reliability in the target detection area; Based on the pixel trust mask, a set of trust pixels is selected, its parallax information is extracted, and the target depth is calculated by combining the camera focal length and baseline parameters. Variance is calculated on the target depth data, and depth uncertainty is corrected. The risk distance is calculated based on the corrected depth uncertainty, and the safe braking distance is determined by combining vehicle speed and system reaction time. Based on the risk distance and the safe braking distance, a braking trigger signal is generated; The braking force output is calculated based on the braking trigger signal and the risk distance difference, and then dynamically and smoothly adjusted.

2. The unmanned vehicle braking control method based on image recognition according to claim 1, characterized in that, The acquisition of vehicle-mounted camera images at consecutive time points, calculation of temporal brightness differences for each pixel of consecutive image frames, and determination of pixel reliability thresholds based on the median and median absolute deviation include: Obtain the brightness value of each pixel in the target area at any two consecutive time points, and record the brightness of the pixel in the current frame and the previous frame respectively; The absolute difference between the brightness values ​​of the pixel in the current frame and the previous frame is used as the temporal brightness difference of the pixel. Within the specified effective sampling area of ​​the image, calculate the median and median absolute deviation of the temporal brightness difference for all pixels; The pixel reliability threshold is set to be based on the median mentioned above, plus three times the constant used to convert the median absolute deviation to the order of standard deviation, multiplied by the median absolute deviation.

3. The image recognition-based braking control method for unmanned vehicles according to claim 1, characterized in that, The step of generating a pixel-reliable mask based on a pixel reliability threshold and statistically analyzing the region reliability in the target detection region includes: Using the pixel temporal brightness difference and the pixel reliability threshold as a comparison benchmark, a binary reliable identifier is output for each pixel. When the temporal brightness difference of the pixel is less than or equal to the threshold, it is marked as a reliable pixel; otherwise, it is marked as a distorted pixel that needs to be removed. Within the bounding box of the region obtained by target detection, the arithmetic mean of the trust labels of all pixels in the region is calculated, and the resulting average value is used as the region-level visual reliability of the target region in the current frame. The aforementioned binary mask and region-level visual reliability serve as input data for subsequent depth estimation and uncertainty correction.

4. The unmanned vehicle braking control method based on image recognition according to claim 1, characterized in that, The step of selecting a set of reliable pixels based on the pixel reliable mask, extracting its disparity information, and calculating the target depth by combining the camera focal length and baseline parameters includes: The set of pixels marked as trustworthy is confirmed as the set of trustworthy pixels; Extract the disparity measurement value for each pixel in the set of trustworthy pixels; The focal length parameter of the camera, the binocular baseline length, and the disparity measurement value of the pixel are geometrically inversely calculated. The spatial depth estimate of the pixel is obtained by multiplying the focal length by the baseline and then dividing by the disparity. The camera focal length and binocular baseline distance are known parameters obtained from device calibration, and geometric inverse calculation is performed based on these known parameters and the measured disparity to output pixel-level target depth.

5. The unmanned vehicle braking control method based on image recognition according to claim 1, characterized in that, The step of calculating the variance of the target depth data and correcting for depth uncertainty includes: The average depth of the target region is calculated in the set of reliable pixels, and the squares of the differences between the depth of each pixel and the average depth are averaged to obtain the original depth variance. Obtain the regional-level visual reliability of the target area; The confidence correction of the original depth variance is performed using regional visual reliability as a correction factor. Specifically, the depth standard deviation is obtained by first taking the square root of the original depth variance, and then the weighted depth uncertainty is obtained by dividing the depth standard deviation by the sum of regional visual reliability and a small positive number. The small positive number is used to prevent division by zero.

6. The unmanned vehicle braking control method based on image recognition according to claim 5, characterized in that, The step of calculating the variance of the target depth data and correcting for depth uncertainty further includes: Temporal smoothing correction is applied to the weighted depth uncertainty. The weighted depth uncertainty of the current frame is linearly weighted and synthesized with the smoothed depth uncertainty of the previous frame according to adaptive weights, and then used as the temporal smoothed depth uncertainty of the current frame. The adaptive smoothing weights are obtained by transforming the regional visual reliability through a logistic function. The input of the transformation is the difference between the regional visual reliability and the preset reliability balance threshold multiplied by the smoothing sensitivity coefficient. The transformation output value is between zero and one, and serves as the mixed weight of the new observation and the historical smoothing value in the current frame. The smoothing sensitivity coefficient and reliability balance threshold are settable parameters as required.

7. The unmanned vehicle braking control method based on image recognition according to claim 1, characterized in that, The calculation of the risk distance based on the corrected depth uncertainty, and the determination of the safe braking distance in combination with vehicle speed and system reaction time, includes: The visually reliable risk distance is calculated by subtracting the confidence control coefficient from the target average depth and multiplying it by the time-smoothed depth uncertainty. Obtain the vehicle's current position and speed, and the vehicle control system's reaction time parameters; Based on the vehicle speed and reaction time, as well as the maximum permissible braking acceleration specified by the system, the safe braking distance is determined by multiplying the speed and reaction time by the speed squared and then dividing by twice the maximum braking acceleration. The risk distance and the safe braking distance are used as comparison parameters for subsequent braking trigger judgment.

8. The unmanned vehicle braking control method based on image recognition according to claim 1, characterized in that, The step of generating a braking trigger signal based on the risk distance and the safe braking distance includes: Obtain the visually reliable risk distance and the corresponding safe braking distance; When the visually credible risk distance is less than or equal to the safe braking distance, a braking trigger signal is generated and output; otherwise, no braking trigger signal is generated. The braking trigger signal serves as a Boolean trigger for subsequent braking force calculation and execution.

9. The unmanned vehicle braking control method based on image recognition according to claim 1, characterized in that, The calculation of braking force output based on the braking trigger signal and the risk distance difference, and the dynamic smooth adjustment thereof, includes: When the braking trigger signal is valid, the difference between the safe braking distance and the visually credible risk distance is calculated as the risk distance difference; Based on the risk distance difference, it is mapped to the actual braking force output using a smooth approximation function. The form of the mapping function is that the braking force output is zero when the risk distance difference is zero, and it monotonically increases with the increase of the risk distance difference until the limit value of the vehicle's maximum braking force. When the brake trigger signal is invalid, the braking force output is zero, and the monitoring status is maintained.

Images