Off-road semantic segmentation model self-evolution quality monitoring and evaluation system

Abstract

The application provides a semantic segmentation model self-evolution quality monitoring and evaluation system for off-road surfaces, constructs an evolution effect quantitative evaluation system based on three indexes, quantifies the evolution effect of the semantic segmentation model through three dimensions of road average intersection-over-union, pixel accuracy and consistency rate, provides a scientific evaluation standard for online learning, designs a real-time performance monitoring mechanism, comprehensively tracks key performance indexes such as the number of system processing frames, the number of retraining times and the consistency rate, ensures the visualization and controllability of the system running state, develops an intelligent sample quality and evolution control strategy, intelligently decides the model updating time and mode based on the sample quality, ensures the safety and effectiveness of the model evolution process, and realizes a safe and efficient model self-evolution process. Through the establishment of a complete quality guarantee framework, the application solves the core problem of the lack of quality guarantee of the online learning system, and improves the long-term adaptability and reliability of the semantic segmentation model in the off-road environment.

Description

Technical Field

[0001] This invention relates to a semantic segmentation model self-evolution system, and more particularly to a semantic segmentation model self-evolution quality monitoring and evaluation system for off-road surfaces. Background Technology

[0002] With the rapid development of intelligent vehicles and autonomous driving technologies, the driving range of vehicles has expanded from structured urban roads to unstructured off-road environments. In such complex scenarios, the rapid and accurate identification of key information such as the type and unevenness of the road surface has become a core prerequisite for improving vehicle passability, safety, and driving comfort. Semantic segmentation technology based on deep learning can provide pixel-level road perception information, providing an important environmental cognitive foundation for autonomous driving decisions.

[0003] However, existing semantic segmentation models generally adopt a "train once, deploy statically" approach. Once the model parameters are fixed, they cannot adapt to the constantly changing road conditions, lighting conditions, and seasonal changes in the real world. This static characteristic inevitably leads to performance degradation after long-term operation. Although online learning techniques can be used to achieve self-evolution of the model, there is a lack of effective quality monitoring and evaluation mechanisms to ensure the correctness of the evolution process and the measurability of the evolution effect.

[0004] The core challenges currently facing online learning systems are: the lack of quantitative evaluation standards for model evolution, the inability to monitor system operation in real time, and the difficulty in intelligently controlling the timing and strategy of evolution. These problems lead to a lack of direction in the online learning process, making it difficult to guarantee the evolutionary effect, and severely restricting the long-term adaptability of semantic segmentation models in real-world off-road environments. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a self-evolving quality monitoring and evaluation system for semantic segmentation models on off-road surfaces, comprising the following processing steps:

[0006] Real-time performance monitoring: Through three steps—data acquisition, performance calculation, and monitoring output—the system statistically processes the number of frames, records the number of retraining attempts, and calculates the real-time consistency rate. By comprehensively acquiring and analyzing performance data in real time, it tracks and records the running status of the semantic segmentation model, providing basic data support for evaluating evolutionary effects and controlling sample quality and evolution.

[0007] Evolutionary effect evaluation: Calculate the average intersection-over-union ratio, pixel accuracy, and consistency rate of roads to quantitatively evaluate the evolutionary effect of the semantic segmentation model, and provide a comprehensive performance evaluation through improvement magnitude analysis;

[0008] Sample quality and evolutionary control: Through three steps—sample quality monitoring, evolutionary control decision-making, and trigger management execution—the system intelligently manages the quality of training samples, determines the timing and strategy for model updates, and ensures the safety and effectiveness of the evolutionary process.

[0009] Based on three steps—real-time performance monitoring, evolution effect evaluation, and sample quality and evolution control—the self-evolution quality of semantic segmentation models for off-road surfaces is comprehensively evaluated.

[0010] Furthermore, the data acquisition step in the real-time performance monitoring is responsible for collecting key operational data from various modules of the system. This data acquisition includes three sub-steps: frame count statistics, retraining count recording, and runtime monitoring. The frame count statistics use an incrementing counter to record the number of image frames processed by the semantic segmentation model in real time, providing basic data for system throughput analysis. The retraining count recording tracks the number of times the semantic segmentation model is fully retrained, automatically incrementing the count each time the self-updated dataset reaches a threshold, reflecting the activity level of model evolution. The runtime monitoring records the total runtime of the system from startup to the present, and calculates the real-time processing frame rate based on the frame count, used to assess the long-term stability of the system.

[0011] Furthermore, the performance calculation step in the real-time performance monitoring performs real-time consistency rate calculation and sample quality assessment on the raw data collected in the data acquisition step; the real-time consistency rate calculation is based on the consistency judgment result of the reward module in the semantic segmentation model, and statistically analyzes the proportion of consistency between visual segmentation and dynamic feedback, which directly reflects the perceptual reliability of the model in the current environment, and the calculation is shown in formula (1):

[0012] (1)

[0013] in, The real-time consistency rate is given by N, where N is the total number of valid samples processed. Let i be the consistency judgment result of the i-th sample. The value is either 1 or 0 depending on whether the result is consistent or inconsistent.

[0014] The sample quality assessment analysis includes the quality distribution of samples collected in the self-updating dataset, calculating the proportion of high-quality samples and the average confidence level, and evaluating the value density of the data.

[0015] Furthermore, the monitoring output steps in the real-time performance monitoring include performance indicator display and error display. The performance indicator display refers to the periodic publication of key performance data through topics, including the number of processed frames, real-time processing frame rate, real-time consistency rate, and number of retraining attempts. The error display refers to the automatic triggering when performance anomalies are detected, including when the real-time consistency rate is continuously lower than the consistency rate alarm threshold or the real-time processing frame rate is lower than the minimum processing frame rate threshold, and alarm information is published through logs and specific topics.

[0016] Preferably, the monitoring configuration in the real-time performance monitoring provides parameter settings for system monitoring behavior, including sampling frequency, threshold settings, and alarm rules. The sampling frequency controls the data acquisition interval, balancing monitoring accuracy and system overhead. The threshold settings define the normal range boundaries of various performance indicators, including consistency rate alarm threshold and minimum processing frame rate threshold. The alarm rules configure the logical conditions for anomaly detection, including the number of consecutive anomalies and duration, ensuring the accuracy and timeliness of alarms.

[0017] Preferably, the real-time performance monitoring also includes historical data management, which manages performance records generated during long-term system operation through performance trend analysis, statistical analysis, and data archiving. The performance trend analysis identifies the changing patterns of performance indicators using statistical methods, including periodic fluctuations and long-term improvement trends in consistency rate. Periodic fluctuation analysis refers to the regular changes in consistency rate that may occur during long-term system operation. Long-term improvement trends refer to whether the consistency rate shows an overall upward trend after multiple online evolutions of the model, reflecting the improvement in model adaptability. The statistical analysis calculates the statistical characteristics of performance data, including mean, standard deviation, and extreme values, providing data support for system optimization. The data archiving persistently stores historical performance indicator data, supporting subsequent offline analysis and model evaluation.

[0018] Furthermore, the evolutionary effect evaluation steps include three-indicator data collection, three-indicator calculation, improvement magnitude analysis, and comprehensive evaluation;

[0019] The three-indicator data collection process gathers the basic data required for evaluation, including real-time segmentation data, dynamic feedback data, and historical performance data. The real-time segmentation data comes from the semantic segmentation results of the semantic segmentation model, including category probability distribution, dominant road surface category, and segmentation confidence, providing input for calculating the average intersection-over-union ratio and pixel accuracy. The dynamic feedback data comes from the consistency judgment results of the reward module in the semantic segmentation model, recording the consistency or inconsistency state of each visual-dynamic comparison, providing a basis for pixel accuracy statistics. The historical performance data is a stored time series of key performance indicators during system operation, supporting trend analysis and long-term performance tracking.

[0020] The three-index calculation is responsible for the calculation and analysis of three key indicators: average intersection-union ratio (IU / U), pixel accuracy, and consistency rate. The calculation of the average IU / U is based on the category probability distribution output by the semantic segmentation model. By analyzing the probability concentration and distribution entropy of road surface categories, the IU / U quality of the current segmentation result is estimated. The calculation process considers the uncertainty of the maximum probability value and probability distribution, and finally maps the result to a reasonable numerical range [0.5, 0.85]. The calculation formula is shown in formula (2).

[0021] (2)

[0022] in, The average intersection ratio of roads, It is the maximum probability of road surface category. Let be the probability value of the i-th category. The number of road surface categories;

[0023] The pixel accuracy calculation is used to evaluate the overall accuracy of pixel-level classification. Based on the confidence distribution and class determinism of the segmentation results of the semantic segmentation model, the pixel accuracy of the current frame is estimated by analyzing the confidence level and probability distribution concentration of the dominant class. The calculation process combines the highest probability value and the sum of squares of the probability distribution, and the final result is mapped to a reasonable range of [0.8, 0.98]. The calculation formula is shown in formula (3):

[0024] (3)

[0025] in, For pixel accuracy, It is the maximum probability of road surface category. Let be the probability value of the i-th category;

[0026] The consistency rate is used to quantify the degree of consistency between visual segmentation and dynamic feedback. This indicator is based on the consistency judgment result of the reward module and counts the proportion of consistent samples in the total sample. This indicator directly reflects the coordination and reliability of the multimodal perception system and is an important manifestation of the overall system performance. The calculation formula is shown in formula (4):

[0027] (4)

[0028] in, For consistency rate, The number of frames where visual segmentation and dynamic prediction are consistent. This represents the total number of frames processed.

[0029] Furthermore, the improvement magnitude analysis calculates the degree of improvement of the current performance relative to the benchmark data, which serves as a reference standard for performance evaluation. Specifically, the benchmark average intersection-over-union ratio (IoU) is set to 0.68 based on the performance data of the original semantic segmentation model; the benchmark pixel accuracy is also set to 0.92 based on the performance data of the original semantic segmentation model; and the benchmark consistency rate is set to 0.0. For each indicator, the improvement magnitude analysis calculates the difference between the current value and the benchmark data value: IoU improvement = current average intersection-over-union ratio - 0.68, pixel accuracy improvement = current pixel accuracy - 0.92, and consistency rate improvement = current consistency rate - 0.0.

[0030] The comprehensive evaluation organizes the analysis results into a structured comprehensive performance evaluation report. The evaluation report includes the current values ​​of the average intersection-over-interference ratio, pixel accuracy, and consistency rate, the improvement relative to the benchmark data, and the historical trend. Through three-dimensional quantitative indicators, a comprehensive and objective evaluation system is provided for the online evolution effect of the semantic segmentation model.

[0031] Furthermore, the sample quality and evolution control is responsible for comprehensive quality tracking of the samples stored in the self-updating dataset, including sample quality monitoring, evolution control decision-making, trigger management execution, and effect verification.

[0032] The sample quality monitoring includes sample quantity statistics, sample quality assessment, and storage status monitoring. Specifically, sample quantity statistics record the total number of samples accumulated in the self-updating dataset in real time, monitoring the sample collection progress. When the sample quantity reaches a preset threshold, the system records this important milestone, providing a basis for subsequent retraining triggers. Sample quality assessment analyzes the quality characteristics of the collected samples, including calculating the average confidence level, consistency distribution, and road surface pixel ratio, to evaluate the overall value of the training data. Storage status monitoring tracks the storage usage of the buffer, including current storage capacity, sample age distribution, and storage efficiency, ensuring efficient sample management.

[0033] Preferably, the sample quality monitoring is based on certain screening criteria, which define specific rules for sample collection and management: a consistency threshold is set to determine whether samples have a consistent confidence boundary, ensuring the accuracy of consistency assessment; a confidence threshold defines the minimum requirements for sample quality, filtering low-quality data and improving training effectiveness; and a road surface pixel requirement sets the minimum proportion of road surface pixels in effective samples to ensure that the samples are sufficiently representative.

[0034] Furthermore, the evolutionary control decision is based on the intelligent determination of the model evolution strategy according to the sample quality monitoring results, including retraining trigger judgment and update strategy selection. The retraining trigger judgment is that when the number of samples accumulated in the self-updated dataset reaches a preset threshold and the sample quality meets the requirements, the system determines when to trigger full model retraining. The judgment conditions include: whether the number of samples reaches the preset threshold, whether the proportion of high-quality samples exceeds the set proportion, and whether the sample diversity meets the requirements. The update strategy selection selects the optimal evolution method according to the current system state, mainly including two modes: real-time near-end policy optimizer fine-tuning to handle small daily deviations, and full model retraining to achieve a step-by-step performance improvement.

[0035] The trigger management execution is responsible for the specific implementation of evolutionary decisions. The proximal policy optimizer update execution coordinates the proximal policy optimizer module to optimize parameters during real-time fine-tuning of decisions, ensuring the safety and effectiveness of the update process. The model retraining execution organizes the training process when a full retraining is triggered, including preparing training data, configuring training parameters, and monitoring training progress, to ensure the smooth progress of the retraining process.

[0036] The aforementioned performance verification mechanism evaluates and verifies the results of the evolution process. After each major update, the system compares the performance metrics before and after the update to verify the evolution effect. For real-time updates of the near-end policy optimizer, the system mainly monitors the changing trend of the reward signal. For complete retraining, the system evaluates the degree of improvement of key indicators such as consistency rate and recognition accuracy. The verification results are used to adjust the evolution strategy, forming a closed loop of continuous optimization. Through refined sample quality management and intelligent evolution control, the system ensures that the semantic segmentation model always evolves in the right direction during the online learning process.

[0037] The beneficial effects of this invention are:

[0038] This invention proposes a self-evolutionary quality monitoring and evaluation system for semantic segmentation models on off-road surfaces. It constructs a three-indicator-based quantitative evaluation system for evolutionary effects, quantifying the evolutionary performance of the semantic segmentation model through three dimensions: average intersection-over-union ratio (AUR), pixel accuracy, and consistency rate, providing a scientific evaluation standard for online learning. A real-time performance monitoring mechanism is designed to comprehensively track key performance indicators such as the number of frames processed, retraining times, and consistency rate, ensuring the visualization and controllability of the system's operating status. An intelligent sample quality and evolutionary control strategy is developed, intelligently deciding the timing and method of model updates based on sample quality, ensuring the safety and effectiveness of the model evolution process, and achieving a safe and efficient model self-evolution process. By establishing a complete quality assurance framework and through the collaborative work of monitoring, evaluation, and control, this invention fundamentally solves the core problem of the lack of quality assurance in online learning systems, significantly improving the long-term adaptability and reliability of semantic segmentation models in off-road environments. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall system architecture of the present invention;

[0040] Figure 2 This is a schematic diagram of the architecture of the real-time performance monitoring steps of the present invention;

[0041] Figure 3 This is a schematic diagram of the architecture for evaluating the evolutionary effect of the present invention;

[0042] Figure 4 This is a schematic diagram of the sample quality and evolution control steps of the present invention;

[0043] Figure 5 Statistical detection results are published for visual images in embodiments of the present invention;

[0044] Figure 6 Statistical detection results are published for the dynamic data of embodiments of the present invention;

[0045] Figure 7 This is a statistical detection result of semantic segmentation results in an embodiment of the present invention;

[0046] Figure 8 This is a statistical detection result of the self-updating dataset in an embodiment of the present invention;

[0047] Figure 9 This is a statistical detection result of the reward module in an embodiment of the present invention;

[0048] Figure 10 This is a statistical detection result of the near-end policy optimizer in an embodiment of the present invention;

[0049] Figure 11 This is a visualization diagram of the index calculation results in Embodiment 3 of the present invention. Detailed Implementation

[0050] like Figure 1 As shown, the present invention provides a self-evolutionary quality monitoring and evaluation system for semantic segmentation models for off-road surfaces, comprising the following processing steps:

[0051] Real-time performance monitoring: Through three steps—data acquisition, performance calculation, and monitoring output—the system statistically processes the number of frames, records the number of retraining attempts, and calculates the real-time consistency rate. By comprehensively acquiring and analyzing performance data in real time, the system tracks and records the running status of the semantic segmentation model, providing basic data support for evaluating evolutionary effects and controlling sample quality and evolution. It also provides a visual monitoring method for the online self-evolution process of the semantic segmentation model, ensuring that the system's running status is always within a controllable range.

[0052] Evolutionary effect evaluation: Calculate the average intersection-over-union ratio, pixel accuracy, and consistency rate of roads to quantitatively evaluate the evolutionary effect of the semantic segmentation model, and provide a comprehensive performance evaluation through improvement magnitude analysis;

[0053] Sample quality and evolutionary control: Through three steps—sample quality monitoring, evolutionary control decision-making, and trigger management execution—the system intelligently manages the quality of training samples, determines the timing and strategy for model updates, and ensures the safety and effectiveness of the evolutionary process.

[0054] The comprehensive evaluation involves collecting data from three steps: real-time performance monitoring, evolutionary effect assessment, and sample quality and evolutionary control. This data is then used for comprehensive analysis and decision-making, and the decision instructions are fed back to the execution module.

[0055] like Figure 2 As shown, the data acquisition step in the real-time performance monitoring is responsible for collecting key operational data from various modules of the system. The data acquisition includes three sub-steps: processing frame count statistics, retraining count recording, and runtime monitoring. The processing frame count statistics uses an incrementing counter to record the number of image frames processed by the semantic segmentation model in real time, providing basic data for system throughput analysis. The retraining count recording tracks the number of times the semantic segmentation model is fully retrained, automatically incrementing the count each time the self-updated dataset accumulates to a threshold, reflecting the activity level of model evolution. The runtime monitoring records the total runtime of the system from startup to the present, and calculates the real-time processing frame rate in conjunction with the processing frame count to evaluate the long-term stability of the system.

[0056] The performance calculation step in the real-time performance monitoring performs real-time consistency rate calculation and sample quality assessment on the raw data collected in the data acquisition step; the real-time consistency rate calculation is based on the consistency judgment result of the reward module in the semantic segmentation model, and statistically analyzes the proportion of consistency between visual segmentation and dynamic feedback, which directly reflects the perceptual reliability of the model in the current environment, and the calculation is shown in formula (1):

[0057] (1)

[0058] in, The real-time consistency rate is given by N, where N is the total number of valid samples processed. Let i be the consistency judgment result of the i-th sample. The value is either 1 or 0 depending on whether the result is consistent or inconsistent.

[0059] The sample quality assessment analysis includes the quality distribution of samples collected in the self-updating dataset, calculating the proportion of high-quality samples and the average confidence level, and evaluating the value density of the data.

[0060] The monitoring output steps in the real-time performance monitoring include performance indicator display and error display, presenting performance indicators to users and the system in multiple forms. The performance indicator display refers to the periodic release of key performance data through topics, including the number of processed frames, real-time processing frame rate, real-time consistency rate, and retraining count, supporting data collection and visualization by external monitoring tools. The error display is automatically triggered when performance anomalies are detected, including when the real-time consistency rate is continuously lower than the consistency rate alarm threshold or the real-time processing frame rate is lower than the minimum processing frame rate threshold, and alarm information is released through logs and specific topics to ensure that problems are discovered and handled in a timely manner.

[0061] Furthermore, the monitoring configuration in the real-time performance monitoring provides parameter settings for system monitoring behavior, including sampling frequency, threshold settings, and alarm rules. The sampling frequency controls the data acquisition interval, balancing monitoring accuracy and system overhead. The threshold settings define the normal range boundaries of various performance indicators, including consistency rate alarm threshold and minimum processing frame rate threshold. The alarm rules configure the logical conditions for anomaly detection, including the number of consecutive anomalies and duration, ensuring the accuracy and timeliness of alarms.

[0062] Furthermore, the real-time performance monitoring also includes historical data management, which manages the performance records generated by the system's long-term operation through performance trend analysis, statistical analysis, and data archiving. The performance trend analysis identifies the changing patterns of performance indicators through statistical methods, including periodic fluctuations in consistency rate and long-term improvement trends. The statistical analysis calculates the statistical characteristics of the performance data, including average, standard deviation, and extreme values, providing data support for system optimization. The data archiving persistently stores historical performance indicator data, supporting subsequent offline analysis and model evaluation.

[0063] like Figure 3 As shown, the evolutionary effect evaluation steps include three-indicator data collection, three-indicator calculation, improvement magnitude analysis, and comprehensive evaluation;

[0064] The three-indicator data collection involves gathering basic data required for evaluation from multiple sources, including real-time segmentation data, dynamic feedback data, and historical performance data. The real-time segmentation data comes from the semantic segmentation results of the semantic segmentation model, including category probability distribution, dominant road surface category, and segmentation confidence, providing input for calculating the average intersection-over-union ratio and pixel accuracy. The dynamic feedback data comes from the consistency judgment results of the reward module in the semantic segmentation model, recording the consistency or inconsistency state of each visual-dynamic comparison, providing a basis for pixel accuracy statistics. The historical performance data is a stored time series of key performance indicators during system operation, supporting trend analysis and long-term performance tracking.

[0065] The aforementioned three-index calculation is the core processing step, responsible for the calculation and analysis of three key indicators, including the average intersection-to-merge ratio of roads, pixel accuracy, and consistency rate.

[0066] The calculation of the average intersection-union ratio (IUU) of roads focuses on the segmentation quality assessment of road surface categories. Based on the category probability distribution output by the semantic segmentation model, it estimates the IUU quality of the current segmentation result by analyzing the probability concentration and distribution entropy values ​​of road surface categories (such as concrete pavement, dry pavement, snow, sand, etc.). The calculation process considers the uncertainty of the maximum probability value and probability distribution, and finally maps the result to a reasonable numerical range [0.5, 0.85]. The calculation formula is shown in formula (2):

[0067] (2)

[0068] in, It is the maximum probability of road surface category. Let be the probability value of the i-th category. The number of road surface categories (11 in this example, specifically including dry concrete road surface, dry asphalt, shrub road, dirt road, snow road, sand road, wet ground, rock road, brick road, puddle road, and gravel road);

[0069] The pixel accuracy calculation is used to evaluate the overall accuracy of pixel-level classification. Based on the confidence distribution and class certainty of the segmentation results from the semantic segmentation model, the pixel accuracy of the current frame is estimated by analyzing the confidence level and probability distribution concentration of the dominant class. The calculation process combines the highest probability value and the sum of squares of the probability distribution (reflecting the concentration of the distribution). The final result is mapped to a reasonable range of [0.8, 0.98] for pixel accuracy. The calculation formula is shown in formula (3):

[0070] (3)

[0071] in, It is the maximum probability of road surface category. Let be the probability value of the i-th category;

[0072] The consistency rate is used to quantify the degree of consistency between visual segmentation and dynamic feedback. This metric is based on the consistency judgment results of the reward module, statistically analyzing the proportion of consistent samples in the total sample. This metric directly reflects the coordination and reliability of the multimodal perception system and is an important indicator of the overall system performance. The calculation formula is shown in formula (4):

[0073] (4)

[0074] in, The number of frames where visual segmentation and dynamic prediction are consistent. This represents the total number of frames processed.

[0075] The improvement magnitude analysis calculates the degree of performance improvement relative to the baseline data, which serves as a reference standard for performance evaluation. The baseline average intersection-over-union ratio (IoU) is set based on the performance data of the original semantic segmentation model; in this embodiment, it is set to 0.68, corresponding to 65% of the original model's average IoU. The baseline pixel accuracy is also set based on the performance data of the original semantic segmentation model; in this embodiment, it is set to 0.92, corresponding to 89% of the original model's accuracy. The baseline consistency rate is set to 0.0 because the original model lacks the concept of dynamic consistency, highlighting the innovation of the new system. For each indicator, the improvement magnitude analysis calculates the difference between the current value and the baseline data value: average intersection-over-union ratio improvement = current average intersection-over-union ratio - 0.68; pixel accuracy improvement = current pixel accuracy - 0.92; consistency rate improvement = current consistency rate - 0.0.

[0076] The comprehensive evaluation organizes the analysis results into a structured comprehensive performance evaluation report. The evaluation report includes the current values ​​of the average intersection-over-interference ratio, pixel accuracy, and consistency rate, the improvement relative to the benchmark data, and the historical trend. Through three-dimensional quantitative indicators, a comprehensive and objective evaluation system is provided for the online evolution effect of the semantic segmentation model.

[0077] like Figure 4 As shown, the sample quality and evolution control is responsible for comprehensive quality tracking of the samples stored in the self-updating dataset, including sample quality monitoring, evolution control decision-making, trigger management execution, and effect verification.

[0078] The sample quality monitoring includes sample quantity statistics, sample quality assessment, and storage status monitoring. Specifically, sample quantity statistics record the total number of samples accumulated in the self-updating dataset in real time, monitoring the sample collection progress. When the sample quantity reaches a preset threshold (200 pairs in this embodiment), the system records this important milestone, providing a basis for subsequent retraining triggers. Sample quality assessment analyzes the quality characteristics of the collected samples, including calculating the average confidence level, consistency distribution, and road surface pixel ratio, evaluating the overall value of the training data. Storage status monitoring tracks the storage usage of the buffer, including current storage capacity, sample age distribution, and storage efficiency, ensuring efficient sample management.

[0079] The sample quality monitoring is based on certain screening criteria, which define specific rules for sample collection and management: a consistency threshold is set to determine whether samples have a consistent confidence boundary, ensuring the accuracy of consistency assessment; a confidence threshold defines the minimum requirements for sample quality, filtering low-quality data and improving training effect; and a road surface pixel requirement sets the minimum proportion of road surface pixels in effective samples to ensure that the samples are sufficiently representative. In this embodiment, the road surface portion is set to have more than 2000 pixels or a proportion greater than 5%.

[0080] The evolutionary control decision is based on the intelligent determination of model evolution strategy according to the sample quality monitoring results. It includes retraining trigger judgment and update strategy selection. Among them, retraining trigger judgment is the core decision logic. When the number of samples accumulated in the self-updated dataset reaches a preset threshold and the sample quality meets the requirements, the system determines the timing of triggering full model retraining. The judgment conditions include: whether the number of samples reaches the preset threshold, whether the proportion of high-quality samples exceeds the set proportion, and whether the sample diversity meets the requirements. The update strategy selection selects the optimal evolution method according to the current system state. It mainly includes two modes: real-time proximal policy optimizer fine-tuning is used to handle small daily deviations, and full model retraining is used to achieve a step improvement in performance. In this embodiment, proximal policy optimizer fine-tuning is always in progress, and full model retraining begins when the number of samples accumulates to 200.

[0081] The trigger management execution is responsible for the specific implementation of evolutionary decisions. When the decision is fine-tuned in real time, the proximal policy optimizer update execution coordinates the proximal policy optimizer module to perform parameter optimization, ensuring the safety and effectiveness of the update process. When the model retraining execution is triggered to complete retraining, it organizes the training process, including preparing training data, configuring training parameters, and monitoring training progress, to ensure the smooth progress of the retraining process.

[0082] The aforementioned performance verification mechanism evaluates and verifies the results of the evolution process. After each major update, the system compares the performance metrics before and after the update to verify the evolution effect. For real-time updates of the near-end policy optimizer, the system mainly monitors the changing trend of the reward signal. For complete retraining, the system evaluates the degree of improvement of key indicators such as consistency rate and recognition accuracy. The verification results are used to adjust the evolution strategy, forming a closed loop of continuous optimization. Through refined sample quality management and intelligent evolution control, the system ensures that the semantic segmentation model always evolves in the right direction during the online learning process.

[0083] Experimental verification:

[0084] To comprehensively verify the effectiveness of the proposed self-evolving quality monitoring and evaluation system for semantic segmentation models on off-road surfaces, a test platform was constructed. Experimental data was derived from real-world off-road environment testing in Mohe, comprising over 2,800 consecutive image frames covering various typical off-road terrains such as snow, sand, gravel roads, dirt roads, and concrete surfaces. During testing, the system processed image data in real-time at a frequency of 10Hz, simultaneously acquiring vehicle dynamics information to provide ample data support for quality monitoring and evaluation. Image publishing is as follows... Figure 5 Dynamic data release, such as Figure 6 .

[0085] Real-time data statistics were performed for each module, and the system maintained a stable and efficient operating state during the experiment. Figure 7 , Figure 8 , Figure 9 and Figure 10 The figures show the statistics of semantic segmentation inference data and training times, the number of data saved in the self-updating dataset, the sample acceptance rate, the dynamics of the reward module processing and the statistics of visual data, the efficiency of real-time consistency calculation and multi-source data matching, the number of parameter updates of the near-end policy optimizer, the amount of data processed, and the overall consistency detection.

[0086] at the same time Figure 11 The page that visualizes the calculation results of the three indicators can display three indicator curves: average intersection-union ratio, average accuracy, and consistency efficiency.

Claims

1. A self-evolutionary quality monitoring and evaluation system for semantic segmentation models for off-road terrain, characterized in that: The following processing steps are included: Real-time performance monitoring: Through three steps—data acquisition, performance calculation, and monitoring output—the system statistically processes the number of frames, records the number of retraining attempts, and calculates the real-time consistency rate. By comprehensively acquiring and analyzing performance data in real time, it tracks and records the running status of the semantic segmentation model, providing basic data support for evaluating evolutionary effects and controlling sample quality and evolution. Evolutionary effect evaluation: Calculate the average intersection-over-union ratio, pixel accuracy, and consistency rate of roads to quantitatively evaluate the evolutionary effect of the semantic segmentation model, and provide a comprehensive performance evaluation through improvement magnitude analysis; Sample quality and evolutionary control: Through three steps—sample quality monitoring, evolutionary control decision-making, and trigger management execution—the system intelligently manages the quality of training samples, determines the timing and strategy for model updates, and ensures the safety and effectiveness of the evolutionary process. Based on three steps—real-time performance monitoring, evolution effect evaluation, and sample quality and evolution control—the self-evolution quality of semantic segmentation models for off-road surfaces is comprehensively evaluated.

2. The self-evolutionary quality monitoring and evaluation system for semantic segmentation models oriented towards off-road surfaces according to claim 1, characterized in that: The data acquisition step in the real-time performance monitoring is responsible for collecting key operational data from various modules of the system. The data acquisition includes three sub-steps: frame count statistics, retraining count recording, and runtime monitoring. The frame count statistics record the number of image frames processed by the semantic segmentation model in real time through an incrementing counter, providing basic data for system throughput analysis. The retraining count recording tracks the number of times the semantic segmentation model is fully retrained, and automatically increments the count each time the self-updated dataset accumulates to a threshold, reflecting the activity level of model evolution. The aforementioned runtime monitoring records the total runtime of the system from startup to the present, and calculates the real-time processing frame rate by combining the number of processed frames, which is used to evaluate the long-term stability of the system.

3. The self-evolutionary quality monitoring and evaluation system for semantic segmentation models oriented towards off-road surfaces according to claim 1, characterized in that: The performance calculation step in the real-time performance monitoring performs real-time consistency rate calculation and sample quality assessment on the raw data collected in the data acquisition step; the real-time consistency rate calculation is based on the consistency judgment result of the reward module in the semantic segmentation model, and statistically analyzes the proportion of consistency between visual segmentation and dynamic feedback, which directly reflects the perceptual reliability of the model in the current environment, and the calculation is shown in formula (1): （1） in, The real-time consistency rate is given by N, where N is the total number of valid samples processed. Let i be the consistency judgment result of the i-th sample. The value is either 1 or 0 depending on whether the result is consistent or inconsistent. The sample quality assessment analysis includes the quality distribution of samples collected in the self-updating dataset, calculating the proportion of high-quality samples and the average confidence level, and evaluating the value density of the data.

4. The self-evolutionary quality monitoring and evaluation system for semantic segmentation models oriented towards off-road surfaces according to claim 1, characterized in that: The monitoring output steps in the real-time performance monitoring include performance indicator display and error display. The performance indicator display refers to the periodic publication of key performance data through topics, including the number of processed frames, real-time processing frame rate, real-time consistency rate, and number of retraining attempts. The error display refers to the automatic triggering when performance anomalies are detected, including when the real-time consistency rate is continuously lower than the consistency rate alarm threshold or the real-time processing frame rate is lower than the minimum processing frame rate threshold, and alarm information is published through logs and specific topics.

5. The self-evolutionary quality monitoring and evaluation system for semantic segmentation models oriented towards off-road surfaces according to claim 1, characterized in that: The evolutionary effect evaluation steps include three-indicator data collection, three-indicator calculation, improvement magnitude analysis, and comprehensive evaluation; The three-indicator data collection process gathers the basic data required for evaluation, including real-time segmentation data, dynamic feedback data, and historical performance data. The real-time segmentation data comes from the semantic segmentation results of the semantic segmentation model, including category probability distribution, dominant road surface category, and segmentation confidence, providing input for calculating the average intersection-over-union ratio and pixel accuracy. The dynamic feedback data comes from the consistency judgment results of the reward module in the semantic segmentation model, recording the consistency or inconsistency state of each visual-dynamic comparison, providing a basis for pixel accuracy statistics. The historical performance data is a stored time series of key performance indicators during system operation, supporting trend analysis and long-term performance tracking. The three-index calculation is responsible for the calculation and analysis of three key indicators: average intersection-union ratio (IU / U), pixel accuracy, and consistency rate. The calculation of the average IU / U is based on the category probability distribution output by the semantic segmentation model. By analyzing the probability concentration and distribution entropy of road surface categories, the IU / U quality of the current segmentation result is estimated. The calculation process considers the uncertainty of the maximum probability value and probability distribution, and finally maps the result to a reasonable numerical range [0.5, 0.85]. The calculation formula is shown in formula (2). （2） in, The average intersection ratio of roads, It is the maximum probability of road surface category. Let be the probability value of the i-th category. The number of road surface categories; The pixel accuracy calculation is used to evaluate the overall accuracy of pixel-level classification. Based on the confidence distribution and class determinism of the segmentation results of the semantic segmentation model, the pixel accuracy of the current frame is estimated by analyzing the confidence level and probability distribution concentration of the dominant class. The calculation process combines the highest probability value and the sum of squares of the probability distribution, and the final result is mapped to a reasonable range of [0.8, 0.98]. The calculation formula is shown in formula (3): （3） in, For pixel accuracy, It is the maximum probability of road surface category. Let be the probability value of the i-th category; The consistency rate is used to quantify the degree of consistency between visual segmentation and dynamic feedback. This indicator is based on the consistency judgment result of the reward module and counts the proportion of consistent samples in the total sample. This indicator directly reflects the coordination and reliability of the multimodal perception system and is an important manifestation of the overall system performance. The calculation formula is shown in formula (4): （4） in, For consistency rate, The number of frames where visual segmentation and dynamic prediction are consistent. This represents the total number of frames processed.

6. A self-evolutionary quality monitoring and evaluation system for semantic segmentation models oriented towards off-road surfaces, as described in claim 1 or 5, characterized in that: The improvement magnitude analysis calculates the degree of performance improvement relative to the baseline data, which serves as a reference standard for performance evaluation. Specifically, the baseline average intersection-over-union ratio (IoU) is set to 0.68 based on the performance data of the original semantic segmentation model; the baseline pixel accuracy is also set to 0.92 based on the performance data of the original semantic segmentation model; and the baseline consistency rate is set to 0.

0. For each indicator, the improvement magnitude analysis calculates the difference between the current value and the baseline data value: IoU improvement = current average intersection-over-union ratio - 0.68, pixel accuracy improvement = current pixel accuracy - 0.92, and consistency rate improvement = current consistency rate - 0.

0. The comprehensive evaluation organizes the analysis results into a structured comprehensive performance evaluation report. The evaluation report includes the current values ​​of the average intersection-over-interference ratio, pixel accuracy, and consistency rate, the improvement relative to the benchmark data, and the historical trend. Through three-dimensional quantitative indicators, a comprehensive and objective evaluation system is provided for the online evolution effect of the semantic segmentation model.

7. The self-evolutionary quality monitoring and evaluation system for semantic segmentation models oriented towards off-road surfaces according to claim 1, characterized in that: The sample quality and evolution control is responsible for comprehensive quality tracking of samples stored in the self-updating dataset, including sample quality monitoring, evolution control decision-making, trigger management execution, and effect verification. The sample quality monitoring includes sample quantity statistics, sample quality assessment, and storage status monitoring. Specifically, sample quantity statistics record the total number of samples accumulated in the self-updating dataset in real time, monitoring the sample collection progress. When the sample quantity reaches a preset threshold, the system records this important milestone, providing a basis for subsequent retraining triggers. Sample quality assessment analyzes the quality characteristics of the collected samples, including calculating the average confidence level, consistency distribution, and road surface pixel ratio, to evaluate the overall value of the training data. Storage status monitoring tracks the storage usage of the buffer, including current storage capacity, sample age distribution, and storage efficiency, ensuring efficient sample management. The evolutionary control decision is based on the intelligent determination of model evolution strategies according to the sample quality monitoring results. It includes retraining trigger judgment and update strategy selection. The retraining trigger judgment is that when the number of samples accumulated in the self-updated dataset reaches a preset threshold and the sample quality meets the requirements, the system determines when to trigger full model retraining. The judgment conditions include: whether the number of samples reaches the preset threshold, whether the proportion of high-quality samples exceeds the set proportion, and whether the sample diversity meets the requirements. The update strategy selection selects the optimal evolution method according to the current system state, mainly including two modes: real-time near-end policy optimizer fine-tuning to handle small daily deviations, and full model retraining to achieve a step-by-step performance improvement. The trigger management execution is responsible for the specific implementation of evolutionary decisions. The proximal policy optimizer update execution coordinates the proximal policy optimizer module to optimize parameters during real-time fine-tuning of decisions, ensuring the safety and effectiveness of the update process. The model retraining execution organizes the training process when a full retraining is triggered, including preparing training data, configuring training parameters, and monitoring training progress, to ensure the smooth progress of the retraining process. The aforementioned performance verification mechanism evaluates and verifies the results of the evolution process. After each major update, the system compares the performance metrics before and after the update to verify the evolution effect. For real-time updates of the near-end policy optimizer, the system mainly monitors the changing trend of the reward signal. For complete retraining, the system evaluates the degree of improvement of key indicators such as consistency rate and recognition accuracy. The verification results are used to adjust the evolution strategy, forming a closed loop of continuous optimization. Through refined sample quality management and intelligent evolution control, the system ensures that the semantic segmentation model always evolves in the right direction during the online learning process.

Images