Sanitation vehicle operation quality evaluation and self-optimization method and system based on image analysis

Abstract

The present application relates to the technical field of computer vision and intelligent environmental sanitation, and discloses a method and system for evaluating and self-optimizing the operation quality of an environmental sanitation vehicle based on image analysis, which comprises five steps of synchronous acquisition and space-time pairing of double-view images, semantic segmentation and multi-granularity classification of road surface pollutants, differential analysis and quantitative evaluation of operation effect images, dynamic self-optimization of operation parameters based on pollution characteristics, and construction and knowledge iteration of road cleaning degree image archives, thereby constructing a full-link closed loop from acquisition to optimization.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and intelligent sanitation technology, and in particular to a method and system for quality assessment and self-optimization of sanitation vehicle operations based on image analysis, which is applicable to intelligent quality assessment and refined management of urban road sweeping operations. Background Technology

[0002] With the rapid advancement of urbanization, urban road cleaning and maintenance has become a crucial component of municipal management. Sanitation vehicles, as the core equipment for road sweeping, directly impact urban environmental sanitation and residents' quality of life through the assessment and control of their operational quality. However, currently, the detection and evaluation of sanitation vehicle operation quality still heavily relies on manual inspection. This method suffers from highly subjective evaluation standards, low efficiency, and an inability to achieve real-time monitoring and quantitative analysis of the operation process. In practical engineering applications, traditional manual inspection methods exhibit the following prominent drawbacks: inspectors need to conduct visual inspections on-site after the operation is completed, resulting in poor inspection timeliness and an inability to promptly identify and correct operational quality deviations; different inspectors may yield significantly different evaluation results for the same road section, leading to insufficient consistency and reproducibility in assessments; and the setting of operational parameters depends on the driver's personal experience, lacking a scientific basis for parameter adjustment in response to varying road surface pollution conditions.

[0003] Chinese patent CN115482188A discloses a method and system for inspecting the quality of sanitation operations based on image recognition. The technical solution uses the cleanest historical image of the inspected road section as a benchmark image, compares it with the current operational image for structural and texture similarity, and combines this with a waste detection network to identify the number and distribution of waste in the image, calculating a waste distribution score. Finally, the similarity score and the waste distribution score are fused to obtain a comprehensive sanitation operation score. This solution achieves a certain degree of objective quantification of sanitation operation quality, overcomes the inconsistency problem of purely subjective manual evaluation, and provides a relatively standardized quality evaluation method by weightedly fusing image similarity and waste distribution dimensions.

[0004] However, in-depth analysis reveals the following key shortcomings in the aforementioned comparative documents: First, the scheme employs a single-view shooting method, relying on patrol personnel to take handheld photos. The uncontrollability of shooting angle, distance, and timing makes it difficult to guarantee the consistency of the acquired images. Furthermore, it cannot accurately pair and compare images of the same road segment before and after the operation; it can only compare the current image with a pre-stored historical cleanest image. This static benchmark image fails to reflect the dynamic changes in road conditions. Second, its waste detection is limited to target-level number identification and coarse-grained classification, failing to achieve pixel-level semantic segmentation of road pollutants. Therefore, it cannot accurately quantify the coverage area and spatial distribution characteristics of various pollutants, and its detection capability for unstructured pollutants without clear shape boundaries, such as mud and oil, is insufficient. Third, the scheme only provides post-event evaluation functions and lacks a feedback and adjustment mechanism between the operation quality assessment results and the sanitation vehicle's operating parameters. It cannot optimize key operating parameters such as brush rotation speed, suction intensity, and water flow rate in real time based on the actual road pollution conditions. Fourth, the scheme lacks a long-term data accumulation and knowledge iteration mechanism, and cannot provide data support and parameter recommendations for refined operations on different road sections under different seasons and weather conditions. Each operation requires adjusting parameters from scratch.

[0005] Therefore, there is an urgent need for an intelligent method and system that can achieve automatic acquisition and accurate pairing of vehicle-mounted dual-view images, pixel-level semantic segmentation and classification of pollutants, differential quantitative evaluation before and after operation, and real-time self-optimization of operation parameters based on evaluation feedback, in order to solve the technical problems of single evaluation dimension, lack of self-optimization closed loop and knowledge accumulation capability in existing technologies. Summary of the Invention

[0006] To address the technical problems of existing technologies that rely on manual inspections for sanitation vehicle operation quality assessment, have a single evaluation dimension, and cannot optimize operation parameters in real time, this invention provides a method and system for sanitation vehicle operation quality assessment and self-optimization based on image analysis.

[0007] The first aspect of this invention provides a method for quality assessment and self-optimization of sanitation vehicle operations based on image analysis, comprising the following steps: Step S1, simultaneous acquisition and spatiotemporal pairing of dual-view images: A front-facing camera mounted at the front of the sanitation vehicle acquires the original road surface image before operation, and a rear-facing camera mounted at the rear of the sanitation vehicle simultaneously acquires the road surface image after operation. Based on GPS coordinates and acquisition timestamps, spatiotemporal pairing of images before and after operation on the same road segment is performed to generate a set of paired image pairs; Step S2, semantic segmentation and multi-granularity classification of road pollutants: The road surface image before operation is input into a pre-trained road pollutant semantic segmentation network for pixel-level semantic parsing, outputting a segmentation mask image. Color feature analysis and texture feature extraction are performed on each pollutant region to achieve multi-granularity classification, generating a spatial distribution map of pollutants; Step S3, differential analysis of operation effect images. In the quantitative assessment, after geometric registration of paired image pairs, pixel-level differences are calculated. Combined with the segmentation mask map, the removal area and residual area of ​​various pollutants are statistically analyzed, and the pollutant removal rate, corner omission rate, and curb residual index are calculated. In step S4, the operation parameters are dynamically self-optimized based on pollution characteristics. The current pollution type and coverage density are identified based on the road surface images collected in real time by the front camera. The brush rotation speed, suction intensity, and water flow rate are dynamically adjusted based on the mapping relationship model, and the mapping relationship model is corrected in real time with the evaluation results of step S3 as feedback signals. In step S5, the road cleanliness image archive is constructed and knowledge is iterated. The image data, evaluation results, and operation parameters of each road segment are associated and stored. Cluster analysis is performed according to season and weather conditions to generate pollution feature maps and optimal operation parameter configurations, and feedback is fed back to update the mapping relationship model of step S4.

[0008] The second aspect of this invention provides a system for evaluating and self-optimizing the operation quality of sanitation vehicles based on image analysis, comprising: a dual-view image synchronous acquisition module, used to acquire road surface images before and after operation using a front camera and a rear camera respectively, and perform spatiotemporal pairing; a road pollutant semantic segmentation and classification module, used to perform pixel-level semantic segmentation and multi-granularity classification on the road surface image before operation; an operation effect differential analysis and evaluation module, used to perform geometric registration and pixel-level difference calculation on paired image pairs, and output multi-dimensional quantitative evaluation indicators; an operation parameter dynamic self-optimization module, used to dynamically adjust operation parameters according to real-time identified pollution characteristics, and receive feedback from evaluation results to correct the parameter mapping model in real time; and a road segment cleanliness archive and knowledge iteration module, used to accumulate image data and operation parameters of each road segment over a long period of time, generate optimal parameter configurations by seasonal weather clustering, and update the self-optimization module. Each of the above modules and method steps corresponds one-to-one, forming a deeply coupled closed-loop collaborative architecture.

[0009] The beneficial effects of this invention are as follows: It achieves automatic acquisition and precise pairing of images before and after operation on the same road segment using a vehicle-mounted dual-view camera, overcoming the problem of uncontrollable image acquisition in existing technologies; it achieves refined classification and area quantification of pollutants through pixel-level semantic segmentation, with assessment accuracy significantly superior to coarse-grained schemes at the target detection level; it outputs multi-dimensional quantitative indicators such as clearance rate, corner omission rate, and roadside residue index through differential analysis, making the assessment results more objective and comprehensive; it establishes a closed-loop feedback mechanism from assessment results to operation parameters, enabling real-time self-optimization of operation parameters, effectively improving operation quality and reducing resource consumption; and it provides data-driven parameter recommendation support for refined operations under different working conditions through long-term archive accumulation and knowledge iteration. Attached Figure Description

[0010] Figure 1 This is a flowchart of the image analysis-based method for evaluating and self-optimizing the operation quality of sanitation vehicles, provided in an embodiment of the present invention.

[0011] Figure 2 This is an architecture diagram of the sanitation vehicle operation quality assessment and self-optimization system based on image analysis provided in an embodiment of the present invention. Detailed Implementation

[0012] 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.

[0013] Reference Figure 1 This invention provides a method for evaluating and self-optimizing the operational quality of sanitation vehicles based on image analysis. This method constructs a closed-loop system encompassing image acquisition, intelligent analysis, quantitative evaluation, and parameter self-optimization. Unlike existing schemes based on single-image similarity evaluation, this invention introduces four core innovations: simultaneous dual-view acquisition, pixel-level semantic segmentation, multi-dimensional differential quantization evaluation, and online parameter self-optimization, forming a complete closed-loop intelligent control system. The steps are described in detail below.

[0014] Step S1: Synchronous acquisition and spatiotemporal pairing of dual-view images. In one embodiment of the present invention, the image acquisition system of the sanitation vehicle adopts a front and rear dual-view synchronous architecture, which is the basic prerequisite for achieving accurate comparison of images of the same road surface position before and after operation. Specifically, a front-facing industrial camera is installed at the central axis position of the front frame of the sanitation vehicle. Its sensor resolution is not less than 1920×1080 pixels, preferably a high-definition CMOS sensor of 2560×1440 pixels, and the frame rate is set to 10fps to 30fps, preferably 15fps. The front-facing camera is installed at a height of 1.8m above the road surface, and the lens optical axis is installed at a 45° downward angle to the road surface normal, so that its imaging coverage range is a road surface area of ​​3m to 8m in front of the vehicle. The lens focal length of the front-facing camera is selected as a zoom lens of 6mm to 12mm, preferably a fixed focal length of 8mm. This focal length, combined with the sensor size, can obtain a horizontal field of view of about 60°, so that the road surface width covered by a single frame image is about 3.5m to 7.0m, basically covering the width of a single lane. A rear-mounted industrial camera, identical in model to the front-mounted camera, is installed symmetrically on the rear rack of the sanitation vehicle. Maintaining the same mounting height, tilt angle, and optical parameters, its imaging coverage extends 3 to 8 meters behind the vehicle, encompassing the road surface after cleaning by the cleaning mechanism. The two cameras achieve frame-level synchronization via hardware trigger signals, emitted at a fixed frequency by the vehicle's main controller, ensuring synchronized image frame acquisition at the same time, with the synchronization error between the two cameras controlled within 1ms.

[0015] Preferably, in this invention, each frame of image is recorded with corresponding GPS coordinate information and a precise timestamp simultaneously during acquisition. The GPS module uses RTK differential positioning technology, achieving a horizontal positioning accuracy of 0.02m, thus assigning centimeter-level geographic coordinate labels to each frame of image. The timestamp uses the UTC standard time format with an accuracy of 1ms. Based on this, this invention achieves automatic matching of images before and after operation on the same road segment through a spatiotemporal pairing algorithm. The core idea of ​​this pairing process is: during the sanitation vehicle's operation, the front camera captures an image of the original road surface state before the operating mechanism; after the vehicle has traveled a certain distance, the rear camera captures an image of the same road surface location after the operation is completed. Since there is a fixed physical distance between the two cameras... (i.e., the longitudinal length of the vehicle body, typically 4m to 8m), at a known vehicle speed Under these conditions, the time delay for the rear camera to capture the same road surface location is approximately .

[0016] In one embodiment of the present invention, the specific process of spatiotemporal pairing is as follows: First, the data collected by the front-facing camera... Based on the frame image, obtain its GPS coordinates. and timestamp Then, in the image sequence captured by the rear camera, search for GPS coordinates whose spatial distance from the reference frame is less than a matching distance threshold. Candidate frames, of which Set to 1.0m; among the candidate frames that meet the spatial distance condition, further filter those with timestamps that meet the condition. The frames, in which The minimum time delay is calculated based on the vehicle length and maximum driving speed; finally, the candidate frame with the smallest spatial distance is selected as the matching frame to form a paired image pair. If no candidate frame meets the criteria, the preceding frame is marked as unpaired. After the above processing, all successfully paired image pairs constitute a set of paired image pairs. .

[0017] Furthermore, the image acquisition stage includes image quality preprocessing steps to eliminate the interference of environmental factors on subsequent analysis. This invention employs an adaptive histogram equalization (CLAHE) method to normalize the acquired image based on illumination, dividing the image into 8×8 sub-blocks. Histogram equalization is performed independently within each sub-block, and block boundary artifacts are eliminated through bilinear interpolation. The contrast limit threshold is set to 40. Simultaneously, based on the camera intrinsic parameter matrix and distortion coefficients obtained from pre-shipment calibration, radial and tangential distortion corrections are performed on the image to ensure that the mapping accuracy between pixel coordinates and actual road surface coordinates reaches the sub-pixel level.

[0018] Step S2, Semantic Segmentation and Multi-Granularity Classification of Road Pollutants. The core task of this step is to perform pixel-level semantic parsing on the original road surface images before the operation in the paired image pairs, automatically identifying and labeling the distribution areas and types of various pollutants on the road surface. In one embodiment of the present invention, the road pollutant semantic segmentation network adopts an encoder-decoder architecture, and the overall design is based on the DeepLabV3+ framework and has been specifically optimized for the road pollutant scenario.

[0019] The encoder uses ResNet-50 as the backbone feature extraction network, fine-tuned through transfer learning based on ImageNet pre-trained weights. The input image first passes through an initial convolutional layer with 7×7 kernels, a stride of 2, and 64 output channels, followed by downsampling through a 3×3 max-pooling layer. The backbone network contains four residual block groups, each outputting feature maps with 256, 512, 1024, and 2048 channels, respectively, with spatial resolution progressively decreasing to 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the input image. Preferably, in the 3rd and 4th residual block groups, ordinary convolutions are replaced with dilated convolutions, with dilation rates set to 2 and 4, respectively, keeping the output stride at 16 and expanding the receptive field to capture a wider range of contextual information without increasing the number of parameters.

[0020] The encoder's top-level feature input is a Spatial Pyramid Pooling (ASPP) module. This module has five parallel branches: a 3×3 dilated convolution branch with dilation rates of 6, 12, and 18; a 1×1 ordinary convolution branch; and a global average pooling branch. Each branch outputs a uniform 256 channels. After concatenation, a 1×1 convolution is used to compress the number of channels to 256, achieving multi-scale feature fusion. To address the large scale differences of road surface pollutants, this invention adds an extra branch with a dilation rate of 24 to the ASPP module, enabling the network to effectively perceive multi-scale pollutant targets ranging from small-area paper scraps (typically about 5cm×5cm) to large-area oil stains (typically up to 50cm×100cm).

[0021] The decoder upsamples the low-resolution features output by ASPP by a factor of 4 and then concatenates them with the high-resolution features output by the first residual block group in the encoder. Before concatenation, a 1×1 convolution is applied to the high-resolution features to reduce the number of channels to 48. The fused features are further refined through two consecutive 3×3 convolutional layers, and finally output with a 1×1 convolution to determine the number of channels. The prediction graph, in which The number of pollutant categories is represented by one additional channel, corresponding to the road surface background category. After bilinear upsampling to restore the original input resolution, a softmax operation is performed on each pixel location to obtain the probability of each category. The category corresponding to the highest probability is taken as the semantic label for that pixel, generating a segmentation mask. .

[0022] In one embodiment of the present invention, the semantic segmentation network is trained using a multi-task joint loss function. : ,in: This is the standard pixel-level cross-entropy loss term. For boundary-aware loss terms, This is the category-balanced loss term; , and These are the weighting coefficients for the three loss terms, with values ​​ranging from 0.1 to 1.0, preferably... , , . The calculation formula is: ,in: The total number of pixels in the image, a dimensionless integer, typically denoted as . ; The total number of pollutant categories is a dimensionless integer. In this embodiment... These correspond to five categories: fallen leaves, mud and sand, paper scraps, plastic bags, and oil stains. For the first The pixel belongs to the first The actual label value for the class, which can be 0 or 1; For network prediction of the first The pixel belongs to the first The probability value of the class, with a range of (0, 1).

[0023] Boundary-aware loss term The design approach involves applying higher loss weights to the boundary pixels of the contaminant region to improve the accuracy of the segmentation boundaries. Specifically, the boundary pixel set is first extracted by applying the Sobel edge detection operator to the real segmentation mask. Then, the cross-entropy loss at the boundary pixels is multiplied by the enhancement factor. , The value ranges from 2.0 to 5.0, preferably. Its technical effect is to make the network pay more attention to the prediction accuracy of the edge region of pollutants, and reduce over-segmentation and under-segmentation.

[0024] Category Balanced Loss Item This invention addresses the problem of uneven sample sizes for various pollutants. It is based on the pixel proportions of different pollutants in the training set. Calculate dynamic weights ,in To prevent extremely small positive numbers from being divided by zero, take Then, the cross-entropy loss for each category is multiplied by the normalized value. The technical advantage of this design is that it gives greater learning weight to less frequent pollutant types (such as oil stains), thus preventing the model from being biased towards high-frequency categories.

[0025] Building upon semantic segmentation, this invention further performs multi-granularity classification on the segmented pollutant regions. Preferably, the multi-granularity classification comprises two levels: a primary classification divides pollutants into three main categories: solid waste, liquid stains, and mixed pollutants; a secondary classification further subdivides specific types within each primary category. Classification features include color and texture features. Color features are extracted by converting the RGB pixel values ​​of the pollutant region to the HSV color space, calculating the mean and standard deviation of each channel, resulting in six feature dimensions. Texture features employ the Gray-Level Co-occurrence Matrix (GLCM) method, calculating contrast, correlation, energy, and homogeneity in four directions: 0°, 45°, 90°, and 135°, resulting in 16 feature dimensions. The color and texture features are concatenated to form a 22-dimensional feature vector, which is input into a pre-trained random forest classifier for fine-grained classification. This classifier contains 200 decision trees with a maximum depth of 15, achieving a classification accuracy of 92.6% on the validation set.

[0026] After the semantic segmentation and multi-granularity classification processes described above, the generated pollutant spatial distribution map includes the pollutant category label and confidence value for each pixel, and also records the segmentation area of ​​each pollutant category (in terms of pixel count and actual physical area). The dual representations, centroid coordinates, and circumscribed rectangle parameters provide precise quantitative input for the differential analysis in the subsequent step S3. It is worth noting that the conversion from pixel area to actual physical area requires inverse perspective transformation using the camera's intrinsic and extrinsic parameters. Based on the camera parameters calibrated in step S1 and the known installation height and tilt angle, this invention establishes a mapping lookup table from pixel coordinates to road surface physical coordinates. The road surface physical area corresponding to each pixel varies according to its position in the image; near-end pixels (below the image) correspond to smaller physical areas, while far-end pixels (above the image) correspond to larger physical areas. This non-uniform mapping relationship is precisely compensated for in area statistics to ensure the accuracy of physical quantification of pollutant area. In one embodiment of this invention, the semantic segmentation network's single-frame inference time on the vehicle-mounted GPU platform is approximately 35ms, the additional computation time for multi-granularity classification is approximately 8ms, and the processing latency of the entire step S2 is approximately 43ms, meeting the time requirement of 15fps real-time processing.

[0027] Step S3: Image Differential Analysis and Quantitative Evaluation of Operational Effect. This step aims to quantitatively evaluate the operational effect of the sanitation vehicle by performing precise image differential analysis on paired image pairs. In one embodiment of the present invention, the differential analysis process first performs a geometric registration operation to align the two images before and after the operation to the same coordinate system.

[0028] Geometric registration employs an affine transformation method based on feature point matching. Specifically, firstly, the image before processing... and images after the assignment SIFT feature points were detected in each image, with the number of feature points extracted from each image controlled between 500 and 2000. Then, the KNN nearest neighbor matching algorithm was used for initial matching of feature points between the two images, with a matching ratio threshold set to 0.75, meaning that a match was only accepted if the ratio of the nearest neighbor distance to the second nearest neighbor distance was less than 0.75. Based on the initial matching results, the RANSAC algorithm was used to remove outliers and calculate the optimal affine transformation matrix. The RANSAC iteration count was set to 1000, and the reprojection error threshold was set to 3.0 pixels. The post-processed image was transformed to the coordinate system of the pre-processed image using the calculated affine transformation matrix, resulting in the registered post-processed image. The registration residual is usually controlled within 1.5 pixels.

[0029] After geometric registration, this invention employs an adaptive difference algorithm to calculate the pixel-level differences between the two images. Preferably, the images before and after registration are converted to the LAB color space, and the absolute value of the pixel-by-pixel difference is calculated for the L, A, and B channels to obtain a three-channel difference map. , and The differential threshold is determined using an adaptive method: first, the three-channel differential image is Gaussian blurred (kernel size 5×5, standard deviation 1.5) to suppress noise; then, the mean of the channel differences is calculated within the road surface background area (determined by the background class label in the segmentation mask in step S2). and standard deviation Set the threshold to ,in The sensitivity coefficient ranges from 2.0 to 4.0, with the preferred value being... When the difference between any channel exceeds the corresponding threshold, that pixel is identified as a changed pixel. The advantage of this adaptive threshold strategy is that it can automatically adjust the detection sensitivity according to different lighting conditions and road surface texture, avoiding false detections or missed detections caused by a fixed threshold.

[0030] After obtaining the pixel-level change detection results, and combining the pre-operation segmentation mask image output in step S2, this invention calculates the following three core quantitative evaluation indicators. First, the spatiotemporally weighted pollutant removal rate. : ,in: The total number of pollutant categories, a dimensionless integer. ; For the first The weighting coefficient for the difficulty of removing pollutants is dimensionless and ranges from 0.5 to 2.0, specifically for fallen leaves. silt Paper scraps ,plastic bag Oil stains The selection criteria are based on empirical calibration values ​​of the adhesion and cleaning difficulty of various pollutants. The larger the weight, the more difficult it is to remove the pollutant and the higher its contribution to the removal rate. For the first The segmented area of ​​pollutants in the pre-operation image, in pixels or ; For the first The residual area of ​​pollutants in post-operation images, in units of Consistent. When It was judged as excellent at that time. To be qualified, This is considered unqualified. A weighting based on the difficulty of removal is introduced. The technical advantage lies in enabling the evaluation indicators to differentiate the removal contribution of different pollutant types, avoiding the equating easily removable fallen leaves with difficult-to-remove oil stains, and making the evaluation results more consistent with the actual operational difficulty. Secondly, the corner omission rate... : ,in: Preset width for road edge The total area of ​​residual pollutants after the operation within the specified scope, in units of ; The edge detection width is set to 0.3m to 0.5m, preferably. The selection criteria are based on the width of the blind spots at the edge of the road that are difficult for sanitation vehicle brushes to completely cover. The total area of ​​contaminants in the area before the operation, in units of ; To prevent extremely small positive numbers with a denominator of zero. A lower value indicates a more thorough cleaning of the corners and edges. This invention will... Set as the acceptable threshold. The technical effect of this indicator is to specifically quantify the weak link in sanitation operations—the roadside—compensating for the inability of the overall cleanup rate to reflect localized omissions. Third, the roadside residue index. : , in: This represents the total length of the curb in the current road segment, in units of... The length is determined by the GPS track length, with a typical value of 50m to 500m; The coordinates are the position coordinates along the road and in the direction, in units of ; For position Within 0.5m on both sides of the road edge The area density of pollutants, in units of The data was obtained by performing linear density statistics on the area of ​​various pollutants in the roadside region of the segmented mask image along the length of the roadside. For the first The visual impact factor of pollutants is dimensionless and ranges from 0.5 to 3.0, preferably from fallen leaves. silt Paper scraps ,plastic bag Oil stains The selection criteria are based on the degree of negative visual impact on the urban landscape when different types of pollutants remain on the roadside; the larger the value, the stronger the visual impact. A smaller value indicates higher curb cleanliness. The technical advantage of this indicator lies in quantifying curbs as an independent evaluation dimension, filling the gap in existing technologies that lack specific curb assessments. These three indicators together constitute the vector of quantitative evaluation results for operational quality. This vector can be used for quality assessment of a single task, and also as a feedback input signal for the self-optimization control in step S4. In one embodiment of the present invention, a comprehensive task quality score is also defined. Its calculation method is a weighted fusion of three indicators: ,in , , These are the weighting coefficients for each indicator. Overall Score The score ranges from 0 to 100, with scores above 85 considered excellent, 70-85 considered satisfactory, and below 70 considered unsatisfactory. This comprehensive score can be directly used for evaluation and scoring in the sanitation operation assessment and management system.

[0031] Step S4: Dynamic self-optimization of operating parameters based on pollution characteristics. This step constructs a closed-loop control path from pollutant identification to operating parameter adjustment, realizing real-time adaptive optimization of sanitation vehicle operating parameters. In one embodiment of the present invention, the operating mode of dynamic self-optimization of operating parameters is as follows: the front-facing camera continuously acquires images of the current road surface, and performs real-time inference (inference speed not less than 15fps) through the semantic segmentation network in step S2 to obtain information on the pollution type distribution and coverage density of the current road surface; based on this information, the optimal operating parameters are calculated through a mapping relationship model and sent to the execution mechanism of the sanitation vehicle.

[0032] In one embodiment of the present invention, the mapping relationship model is a multi-input multi-output nonlinear parameter optimization model. Its input feature vector... It consists of three parts: the first part is the pollution type encoding vector. ,in , The The component represents the first... The first part represents the area percentage of pollutants within the current field of view; the second part represents the pollution cover density value. The first part is defined as the ratio of the total area of ​​all pollutants within the current field of view to the total area of ​​the visible road surface. It is dimensionless and ranges from 0 to 1. The third part is the average area value of pollutants. , defined as the arithmetic mean of the areas of the connected regions of all pollutants, in units of The input feature vector is The model's output is a vector of job parameter settings. ,in: This is the brush rotation speed setting, in units of... The value ranges from 200 to 800. ; This is the suction strength setting value, representing the negative pressure value at the nozzle, in units of... The value ranges from 5 to 25. ; Set the sprinkler flow rate, in units of The value ranges from 2 to 15. The mapping relationship model is calculated using the following adaptive parameter optimization function: ,in: This is the first layer weight matrix. The number of neurons in the hidden layer is taken as... ; This is the first layer bias vector; It is the ReLU activation function; This is the weight matrix for the second layer; This is the second layer bias vector; This represents the complete set of learnable parameters for the model. The model employs a two-layer fully connected network structure, which effectively fits the complex nonlinear relationship between pollution characteristics and operational parameters while minimizing computational cost, thus meeting the real-time requirements of automotive embedded platforms.

[0033] Preferably, the present invention sets a parameter safety constraint layer at the model output end, which... , and Each parameter is limited to its own physically feasible range. In practice, the Sigmoid function is used to map the output to the [0, 1] interval, and then linearly scaled to the range of the corresponding parameter. For example, the final output of the brush rotation speed is... ,in This is the original network output. The technical effect of this security constraint mechanism is to prevent abnormal network output from causing the actuator to exceed its physical limits.

[0034] The key innovation of this invention lies in the design of a closed-loop feedback mechanism. After each work segment is completed, step S3 outputs a quantitative evaluation result of the work quality. This signal is fed back to the self-optimization module. The objective function defined in this invention is: ,in: , and The penalty weight coefficient for each evaluation indicator is dimensionless and ranges from 0.1 to 10.0, preferably... , , The selection criteria are based on the order of importance of each indicator in the actual sanitation assessment, with the clearance rate being the most critical and therefore having the largest weight. The squared penalty term for a clearance rate that deviates from the target value by 100%; and These are the squared penalty terms for the corner omission rate and the curb residue index, respectively, with an objective value of 0 for both. Optimize the objective function. The smaller the value, the closer the work quality is to the ideal state.

[0035] Model parameters The online update uses the stochastic gradient descent method: ,in: For the first Model parameters at the next update; The learning rate has a range of values. to Preferred ; This represents the gradient of the objective function with respect to the model parameters. Through this online gradient descent method, the model automatically fine-tunes its parameters after each task feedback, gradually bringing the parameter configuration of subsequent tasks closer to the optimal value. This invention also incorporates a learning rate decay strategy; after every 50 parameter updates, the learning rate is multiplied by a decay factor of 0.95 to prevent parameter oscillation. The core technical effect of this closed-loop feedback mechanism is that it transforms the traditional open-loop operation mode into a closed-loop self-optimization mode, enabling the sanitation vehicle to autonomously adjust its parameters based on actual operation results, continuously improving the quality of its work.

[0036] Step S5: Construction and Knowledge Iteration of Road Section Cleanliness Image Archives. This step achieves structured storage and knowledge accumulation of long-term data, providing data-driven support for the refined operation management of sanitation vehicles. In one embodiment of the present invention, the data structure of the road section cleanliness image archive adopts a multi-level associative index architecture, with the road section number as the first-level index, the collection date as the second-level index, and the operation batch as the third-level index.

[0037] Each record contains the following associated data: the original image file path and thumbnail of the paired image pair, the segmentation mask and contaminant spatial distribution map output from step S2, and the job quality quantitative assessment result vector output from step S3. The actual operation parameter setting vector used in step S4 The data also includes environmental metadata at the time of collection (including temperature, humidity, wind speed, weather type, and season label). The environmental metadata is automatically obtained through a connection to a meteorological service API, with a time granularity of 1 hour.

[0038] Based on accumulated data, this invention performs cluster analysis on archival data according to season and weather conditions to generate pollution characteristic maps for each road segment. The clustering method adopts a time-series-based hierarchical clustering strategy, consisting of two levels of clustering: the first level divides the data into four clusters according to the seasonal dimension: spring leaf fall period (March to May), summer rainy season (June to August), autumn peak leaf fall period (September to November), and winter icing period (December to February); the second level further subdivides each seasonal cluster into four subclusters according to weather conditions: sunny days, cloudy days, rainy days, and snowy days. The pollution characteristic map within each subcluster is defined as a statistical description of the frequency, average area, and typical distribution location of various pollutants in all collected data for that subclusterter.

[0039] Preferably, the present invention calculates the optimal operational parameter configuration for each road segment-season-weather combination. The optimal parameters are determined using the following knowledge iteration weight update formula: ,in: For the process The optimal job parameter configuration vector after the next job iteration; The total number of assignments is a dimensionless positive integer. For the first The parameter vector actually used in this task; For the first The quality weight factor for this task is calculated as follows: ,in For the first The objective function value of this operation (from step S4) (Calculated) The temperature parameter has a value range of 0.1 to 1.0, preferably... Its technical effect lies in making the work quality better. The smaller the parameter configuration, the higher its weight, thus gradually converging the optimal parameter to the configuration with the best actual effect.

[0040] In this invention, the knowledge iteration update cycle is set to perform a batch update once a week. During each update, all seasonal-weather combination subclusters of all road segments are traversed, and the newly accumulated operational data is used to recalculate... The updated optimal parameter configuration is automatically synchronized to the initialization parameters of the mapping model in step S4. The technical effect of this knowledge iteration mechanism is that, through long-term data-driven parameter optimization, sanitation vehicles can directly call upon historical optimal parameter configurations as the starting point for operations when facing specific road sections, seasons, and weather conditions, shortening the self-optimization convergence time and improving the quality of the first pass. It is worth noting that an abnormal data filtering mechanism is also introduced during the knowledge iteration process; when the objective function value of a certain operation... If the value exceeds three standard deviations of the historical mean of a subcluster, the data from that operation will be marked as an outlier and excluded from the optimal parameter calculation. This is to prevent outlier data caused by occasional factors such as equipment failure or extreme weather from polluting the long-accumulated knowledge base. In addition, the system will periodically monitor the convergence of parameters for each subcluster. When the change in the optimal parameters of a subcluster is less than 2% of the initial value for four consecutive weeks, the parameters of that subcluster are considered to have converged sufficiently. Subsequent recalculation will only be triggered when new significantly biased data is added, in order to save cloud computing resources.

[0041] Furthermore, the archive also supports cross-segment knowledge transfer. For newly included road segments, this invention searches for road segments with similar attributes in the existing archive based on their pavement material, road width, and surrounding environment type, and uses their optimal parameter configuration as the initial recommended parameters for the new road segment, achieving rapid parameter initialization in cold start scenarios. Similarity retrieval uses Euclidean distance as a metric, and the attribute vector includes five dimensions: pavement material code, road width, green coverage, surrounding building density, and traffic flow level. Preferably, the cold start parameters for the new road segment are not directly taken from the parameters of the most similar road segment, but rather a weighted average of the optimal parameters from the top three most similar road segments, with the weight proportional to the similarity. This strategy improves the robustness of the cold start parameters and avoids excessive deviation in initial parameters due to the specificity of a single reference road segment.

[0042] In one embodiment of the present invention, the data storage of the archive employs a hierarchical compression strategy to control storage space usage. Specifically, data from the past 30 days retains complete original paired image files and segmentation masks, with a daily storage volume of approximately 200MB per segment; data from 30 to 180 days retains only segmentation masks and evaluation results, deleting the original image files; data older than 180 days is further compressed into statistical summary information, retaining only the pollution characteristic statistics and optimal parameter configurations for each cluster. This hierarchical compression strategy keeps the system's storage space growth rate at a controllable level of approximately 10GB per month during long-term operation.

[0043] By combining the above five steps, this invention forms a complete closed loop of "acquisition-analysis-evaluation-optimization-accumulation". The output of step S1 provides paired image input for steps S2 and S3; the segmentation result of step S2 is used for differential quantization analysis in step S3, and simultaneously for real-time contamination identification in step S4; the evaluation result of step S3 is fed back to step S4 for parameter correction; the knowledge accumulation in step S5 updates the mapping model of step S4 in reverse, and the accumulated data can be used for incremental training of the segmentation network in step S2, forming a deeply coupled multi-layered closed-loop collaborative system. The data flow between each step of this invention is clear and inseparable; the absence of any single step will lead to a functional degradation of the overall closed-loop system, demonstrating the deep coupling and synergistic effect between the steps.

[0044] Reference Figure 2 The present invention also provides an image analysis-based sanitation vehicle operation quality assessment and self-optimization system, which corresponds one-to-one with each step in the above method embodiments and includes the following functional modules.

[0045] The dual-view image synchronous acquisition module, corresponding to step S1, includes a front-mounted industrial camera installed at the front of the sanitation vehicle and a rear-mounted industrial camera installed at the rear. The two cameras achieve frame-level synchronous acquisition through a hardware trigger controller. Preferably, both the front and rear cameras are industrial-grade CMOS cameras supporting GigE Vision or USB 3.0 interfaces, with a protection rating of at least IP67 and an operating temperature range of -20℃ to 60℃ to adapt to all-weather outdoor sanitation operation environments. This module also includes an RTK differential GPS positioning unit and a precise timing unit, used to assign centimeter-level geographic coordinates and millisecond-level timestamps to each frame of image. The image pairing engine is responsible for executing a spatiotemporal pairing algorithm, automatically completing the pairing and association of images before and after the operation based on GPS coordinate distance and time delay constraints, and outputting a set of paired image pairs to subsequent modules. Furthermore, this module also integrates an image quality preprocessing subunit, performing CLAHE illumination normalization and distortion correction operations to ensure the consistency of image quality input to subsequent analysis modules. In terms of data transmission, the module transmits the acquired image data to the vehicle edge computing platform via a gigabit Ethernet interface. The transmission delay of a single frame image is no more than 5ms, and the data transmission bandwidth margin is sufficient to meet the requirements of simultaneous transmission of dual video streams without causing data frame loss or buffer overflow.

[0046] The road pollutant semantic segmentation and classification module, corresponding to step S2, is deployed on an onboard edge computing platform, preferably using an NVIDIA Jetson AGX Orin or an embedded GPU platform with equivalent computing power (no less than 200 TOPS INT8). This module loads a pre-trained road pollutant semantic segmentation network model, receives images output from the acquisition module, performs forward inference using an encoder-decoder structure, and outputs a pixel-level segmentation mask. The model file is optimized for inference using the TensorRT acceleration engine, quantizing the floating-point model to INT8 precision, increasing the inference speed to more than three times that of the original PyTorch model while maintaining a segmentation accuracy loss of no more than 0.5%. Simultaneously, this module integrates multi-granularity classification subunits, utilizing color and texture features to achieve two-level fine-grained classification of pollutants, generating a spatial distribution map of pollutants including category labels, area statistics, and centroid coordinates. This module supports a dual-channel parallel processing mode: one channel processes the pre-processing images of paired image pairs for offline evaluation, and the other channel processes the real-time video stream from the front camera for online self-optimization control. Both channels share the same segmentation network weights but run the inference process independently. The inference latency of this module is controlled within 50ms, which meets the requirements of real-time processing.

[0047] The job performance difference analysis and evaluation module, corresponding to step S3, receives paired image pairs and the pre-job segmentation mask as input. The geometric registration subunit performs SIFT feature point detection, KNN matching, and RANSAC affine transformation calculation to achieve sub-pixel-level alignment of the images before and after the job. It is worth noting that this invention also introduces an image registration quality evaluation mechanism in actual engineering implementation. When the RANSAC in-point ratio is less than 60% or the registration residual exceeds 3.0 pixels, the paired image pair will be marked as registration failure and excluded from subsequent analysis processes to avoid false difference results caused by registration errors. The difference calculation subunit performs adaptive threshold difference processing in the LAB color space and performs classification statistics in conjunction with the segmentation mask. The evaluation index calculation subunit calculates the spatiotemporally weighted pollutant removal rate according to the formulas described in the method embodiment. Corner omission rate and curb residue index Output a vector of quantitative evaluation results for job quality. This module also provides a visualization output function, which overlays the difference results of the images before and after the operation in the form of a heat map. Different colors represent the areas where different types of pollutants have been removed and remain, making it easier for sanitation managers to intuitively understand the operation results.

[0048] The dynamic self-optimization module for operation parameters, corresponding to step S4, is the core control module of this system. Its forward control path receives the real-time inference results from the semantic segmentation and classification module, extracts features such as pollution type encoding vectors, coverage density, and average area, calculates them using a mapping relationship model, and outputs the operation parameter setpoints. These setpoints are then sent to the sanitation vehicle's brush motor controller, vacuum pump controller, and sprinkler valve controller via the CAN bus interface. Its feedback control path receives the output from the evaluation module... The vector is used to calculate the objective function and perform online gradient descent to update the mapping model parameters. The control cycle of the forward path is 100ms, and the update cycle of the feedback path is executed once after each work segment is completed. In one embodiment of the invention, the module is also provided with a manual intervention interface. When the automatic adjustment result does not meet the driver's expectations, the driver can manually fine-tune each work parameter through the vehicle touch screen. The manually adjusted parameter values ​​are also recorded in the archive as training samples for knowledge iteration. In addition, the module has built-in safety protection logic. When the adjustment range of any work parameter exceeds 20% of the rated range in a single control cycle, the system automatically triggers a gradual transition mechanism to distribute the parameter change smoothly across multiple control cycles to prevent mechanical impact on the working mechanism.

[0049] The road section cleanliness archive and knowledge iteration module, corresponding to step S5, connects to the cloud server via a 4G / 5G wireless communication link to upload image data, segmentation results, evaluation results, and operational parameters collected by the vehicle to the cloud archive in real time. Preferably, to reduce communication bandwidth usage, the original images are uploaded after JPEG compression (compression quality factor set to 85) on the vehicle, and the segmentation mask is uploaded in PNG lossless compression format. The upload data size of a single frame paired image is controlled within 500KB. The cloud server is responsible for performing structured storage of archive data, hierarchical clustering analysis of seasonal and weather dimensions, periodic batch calculation of optimal parameter configuration, and knowledge iteration weight updates. The updated optimal parameter configuration is sent to the vehicle self-optimization module via the wireless link, completing the closed-loop knowledge cycle of "vehicle collection - cloud analysis - vehicle application". In addition, this module provides a web-based visual management interface for sanitation management personnel to view the historical cleanliness trends of each road section, operational parameter evolution curves, and cross-road section comparative analysis reports. It supports multi-dimensional query and export functions by road section, by time period, and by weather conditions.

[0050] The aforementioned modules are coupled and collaborate through standardized data interfaces. The dual-view acquisition module provides basic image data for the semantic segmentation and differential analysis modules; the output of the semantic segmentation module simultaneously serves the classification statistics of the differential analysis module and the real-time control of the self-optimization module; the evaluation results of the differential analysis module are both a direct output of job quality and a feedback input to the self-optimization module; the parameter updates of the self-optimization module are supported by long-term data from the knowledge iteration module; and the accumulation of knowledge in the knowledge iteration module, in turn, improves the incremental training accuracy of the segmentation module. This multi-level, deeply coupled design ensures that the system modules are not simply sequentially connected, but rather form a collaborative evolutionary architecture with positive feedback characteristics, achieving a system-level performance gain of 1+1>2.

[0051] To verify the technical effectiveness of this invention, field tests were conducted in a city's urban road sanitation operation scenario. The test environment included 50km of urban main roads and secondary roads, with road surface types covering asphalt and cement concrete, and road widths ranging from 6m to 20m. The surrounding environment of the test sections included typical urban scenarios such as commercial areas, residential areas, and green belts, with a full range of common road pollutants. The test covered three weather conditions: sunny, cloudy, and light rain, with temperatures ranging from 5℃ to 35℃, and the test period spanned 3 months, completing 180 working days of data collection. The test vehicle was an 8t-class sanitation sweeper equipped with the system of this invention, with an onboard computing platform using NVIDIA Jetson AGX Orin, increasing the vehicle's power consumption by no more than 80W. The test dataset contained 12,600 pairs of paired images, independently annotated at the pixel level by 3 professional annotators as ground truth labels. Annotation consistency was evaluated using Cohen's Kappa coefficient, with a result of 0.87, indicating reliable annotation quality.

[0052] In terms of semantic segmentation accuracy, the proposed road pollutant semantic segmentation network achieves an average intersection-union ratio (mIoU) of 78.3% across five pollutant categories: 82.1% for fallen leaves, 76.5% for silt, 74.8% for paper scraps, 79.2% for plastic bags, and 78.9% for oil stains. Compared to the baseline model (mIoU of 72.6%) without boundary-aware loss and class-balancing loss, the multi-task joint loss function of this invention brings a 5.7 percentage point improvement in mIoU, validating the effectiveness of the boundary-aware and class-balancing mechanisms. Furthermore, comparative experiments with the standard UNet structure (mIoU of 70.2%) and the PSPNet structure (mIoU of 74.1%) demonstrate that the DeepLabV3+ improved architecture adopted in this invention has significant advantages in road pollutant segmentation tasks. The overall accuracy of multi-granularity classification on the validation set was 92.6%, with a first-level classification accuracy of 97.1% and a second-level classification accuracy of 91.8%. Confusion matrix analysis showed that the most easily confused category pair was mud and sand-oil, with a mutual misclassification rate of approximately 5.3%.

[0053] Regarding the accuracy of operational quality assessment, using the average score of three sanitation operation experts as a benchmark, the Pearson correlation coefficient between the spatiotemporally weighted pollutant removal rate calculated by this invention and the expert scores reached 0.94, the correlation coefficient between the corner omission rate and the expert scores was 0.91, and the correlation coefficient between the curb residue index and the expert scores was 0.89. Compared with the scheme in Comparative Document 1 that uses overall image similarity plus garbage counting, the assessment dimensions of this invention are more comprehensive, and the detection rate of local omission problems is improved by 37.5%. Especially in the road corner and curb areas, the scheme in Comparative Document 1, due to the lack of specific assessment indicators, has an omission detection rate of only 52.1%, while this invention, through specific assessments of corner omission rate and curb residue index, improves the omission detection rate in this area to 89.6%.

[0054] Regarding the effectiveness of self-optimization control, after three consecutive months of closed-loop operation, the average cleaning rate of sanitation vehicles on the same road sections increased from the initial 81.2% to 91.7%, an increase of 10.5 percentage points. The corner / edge omission rate decreased from the initial 28.6% to 14.3%, a decrease of 14.3 percentage points. The curb residue index decreased from the initial 0.42 to 0.15. Water consumption decreased by 18.3% while the cleaning rate improved, and brush wear rate decreased by 12.7%. These data indicate that the self-optimization mechanism not only improved operational quality but also effectively reduced resource consumption and equipment wear, achieving a dual optimization of quality and economy. It is particularly noteworthy that the self-optimization module can automatically select differentiated parameter configurations when facing different types of pollution. For example, when facing road sections with dense oil stains, it can automatically increase the brush speed to 600 r / min and increase the water flow rate to 12 L / min, while when facing road sections with mainly fallen leaves, it can appropriately reduce the speed to 350 r / min and reduce the water flow rate to 5 L / min. This kind of targeted adjustment is something that traditional fixed parameter schemes cannot achieve.

[0055] Regarding the effectiveness of knowledge iteration, after three months of data accumulation, the optimal parameter configurations for each road segment under different seasonal and weather conditions gradually converged. Using historically optimal parameters as the initial configuration, the first-pass clearance rate for the new round of operations reached 88.5%, an improvement of 9.2 percentage points compared to the cold start scenario without historical parameters (79.3% first-pass clearance rate). The number of iterations required for self-optimization convergence decreased from an average of 8 to 3. Cross-segment knowledge transfer experiments showed that for new road segments with similar attributes, the first-pass clearance rate after migrating the initial parameters was 85.7%, significantly higher than the cold start scenario without any prior knowledge, fully validating the practical value of knowledge iteration and cross-segment transfer mechanisms.

[0056] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for quality assessment and self-optimization of sanitation vehicle operations based on image analysis, characterized in that, Includes the following steps: Step S1, Dual-view image synchronous acquisition and spatiotemporal pairing: The front-mounted camera installed at the front of the sanitation vehicle acquires the original road surface image before the operation, and the rear-mounted camera installed at the rear of the sanitation vehicle acquires the road surface image after the operation. The GPS coordinates and acquisition timestamps corresponding to each image frame are obtained. Based on the spatial distance and time difference of the GPS coordinates, the spatiotemporal pairing of the images before and after the operation of the same road segment is performed to generate a set of paired image pairs. Step S2, Semantic segmentation and multi-granularity classification of road pollutants: The road surface image before operation in the paired image pair is input into the pre-trained semantic segmentation network for pixel-level parsing, and the segmentation mask map of each pollutant category is output. Color features and texture features are extracted for each pollutant region to achieve multi-granularity classification and generate a spatial distribution map of pollutants. Step S3, Image Differential Analysis and Quantitative Evaluation of Operation Effect: Geometric registration is performed on the pre-operation and post-operation images in the paired image pairs. The pixel-level differences of the registered image pairs are calculated. Combined with the segmentation mask image, the removal area and residual area of ​​various pollutants are statistically analyzed, and the pollutant removal rate, corner omission rate and curb residual index are calculated. Step S4, Dynamic self-optimization of operation parameters based on pollution characteristics: Based on the road surface images collected in real time by the front camera, the pollution type and coverage density of the current road surface are identified through a semantic segmentation network. The brush rotation speed, suction intensity and water flow of the sanitation vehicle are dynamically adjusted based on the mapping relationship model. The evaluation results output in step S3 are used as feedback signals to correct the parameters of the mapping relationship model in real time.

2. The image analysis-based method for evaluating and self-optimizing the operation quality of sanitation vehicles according to claim 1, characterized in that, In step S1, both the front camera and the rear camera are industrial cameras with a resolution of not less than 1920×1080 pixels. The installation height of the front camera is 1.5m to 2.5m above the road surface, and the downward angle is 30° to 60°. The rear camera maintains the same installation parameters as the front camera. In the spatiotemporal pairing, the matching distance threshold for GPS coordinates is set to 0.5m to 2.0m, and the matching time window for the collected timestamps is set to the driving time of the vehicle passing through the work section.

3. The image analysis-based method for evaluating and self-optimizing the operation quality of sanitation vehicles according to claim 1, characterized in that, In step S2, the road pollutant semantic segmentation network adopts an encoder-decoder architecture. The encoder extracts multi-scale features based on the ResNet-50 backbone network, and the decoder uses a hollow spatial pyramid pooling module to fuse multi-scale contextual information. The multi-granularity classification includes primary classification and secondary classification. The primary classification divides pollutants into three major categories: solid waste, liquid stains, and mixed pollutants. The secondary classification further subdivides each major category into specific pollutant types.

4. The image analysis-based method for evaluating and self-optimizing the operation quality of sanitation vehicles according to claim 1, characterized in that, In step S3, the pollutant removal rate is calculated based on the ratio of the pixel area difference between the segmented mask before and after the operation to the total area of ​​pollutants before the operation; the corner omission rate is calculated based on the ratio of the residual pollutant area within a preset width range of the road edge to the pollutant area in that area before the operation; the roadside residual index is calculated based on the weighted area density of residual pollutants within a preset distance on both sides of the roadside line.

5. The image analysis-based method for evaluating and self-optimizing the operation quality of sanitation vehicles according to claim 1, characterized in that, In step S2, the semantic segmentation network is trained using a multi-task joint loss function, which includes a weighted sum of a pixel-level cross-entropy loss term, a boundary-aware loss term, and a class-balanced loss term. The boundary-aware loss term applies enhancement weights to the boundary pixels of the pollutant region, and the class-balanced loss term dynamically adjusts the loss weights according to the sample proportions of each type of pollutant.

6. The image analysis-based method for evaluating and self-optimizing the operation quality of sanitation vehicles according to claim 1, characterized in that, In step S3, the geometric registration adopts an affine transformation method based on feature point matching. By detecting and matching SIFT feature points in the images before and after the operation, the affine transformation matrix is ​​calculated to achieve sub-pixel level alignment. The pixel-level difference calculation adopts an adaptive difference algorithm, which calculates the difference for each color channel and performs thresholding processing. The threshold is adaptively determined based on the gray-scale statistical characteristics of the road surface background.

7. The image analysis-based method for evaluating and self-optimizing the operation quality of sanitation vehicles according to claim 1, characterized in that, In step S4, the mapping relationship model is a multi-input multi-output nonlinear regression model. Its inputs include pollution type encoding vector, pollution coverage density value and average area value of pollutants, and its outputs include brush rotation speed setting value, suction intensity setting value and water flow rate setting value. The real-time correction adopts the online gradient descent method, and uses the deviation between the operation quality quantitative evaluation result and the preset target value as the loss function to update the parameters.

8. The image analysis-based method for evaluating and self-optimizing the operation quality of sanitation vehicles according to claim 1, characterized in that, Step S1 further includes an image quality preprocessing step: performing illumination normalization and distortion correction on the acquired image. The illumination normalization adopts an adaptive histogram equalization method to eliminate the influence of illumination conditions at different times. The distortion correction eliminates radial and tangential distortion of the image based on a pre-calibrated camera intrinsic parameter matrix.

9. The image analysis-based method for evaluating and self-optimizing the operation quality of sanitation vehicles according to claim 1, characterized in that, The process also includes step S5, which involves constructing and iterating a road segment cleanliness image archive. This involves associating and storing paired image pairs, pollutant distribution maps, assessment results, and operational parameters for each road segment to build an archive. Then, it involves clustering analysis based on seasons and weather conditions to generate pollution feature maps and optimal operational parameter configurations. Finally, it involves feeding back the optimal parameter configurations to update the mapping relationship model in step S4. The clustering analysis uses a time-series-based hierarchical clustering method. Pollution feature data collected at different times for the same road segment are divided into four clusters based on the seasonal dimension: spring leaf fall period, summer rainstorm period, autumn leaf fall peak period, and winter icing period. Within each cluster, the data is further subdivided into sunny, cloudy, rainy, and snowy subclusters based on weather conditions. Each subcluster corresponds to a set of optimal operational parameter configurations.

10. A sanitation vehicle operation quality assessment and self-optimization system based on image analysis, used to implement the sanitation vehicle operation quality assessment and self-optimization method based on image analysis as described in claim 9, characterized in that, include: The dual-view image synchronous acquisition module is used to acquire road surface images before and after the operation by using a front-mounted camera installed at the front of the sanitation vehicle and a rear-mounted camera installed at the rear of the sanitation vehicle, and to perform spatiotemporal pairing of images before and after the operation on the same road segment based on GPS coordinates and acquisition timestamps to generate a set of paired image pairs. The road pollutant semantic segmentation and classification module is used to perform pixel-level semantic segmentation on the original road surface image before operation in the paired image pair, output segmentation mask images of various pollutants, and realize multi-granularity classification of pollutants through color feature analysis and texture feature extraction to generate a pollutant spatial distribution map. The task performance difference analysis and evaluation module is used to perform geometric registration and pixel-level difference calculation on paired image pairs. Combined with the segmentation mask map, it counts the removal area and residual area of ​​various pollutants and calculates the pollutant removal rate, corner omission rate and curb residue index. The dynamic self-optimization module for operation parameters is used to identify the current pollution type and coverage density based on the road surface images collected in real time by the front camera, dynamically adjust the brush rotation speed, suction intensity and water flow rate based on the mapping relationship model, and receive the evaluation results output by the operation effect differential analysis and evaluation module as feedback signals to correct the mapping relationship model. The road section cleanliness profile and knowledge iteration module is used to associate and store the image data, evaluation results and operation parameters of each road section, generate pollution feature maps and optimal operation parameter configurations by clustering analysis according to season and weather conditions, and feed back the optimal operation parameter configurations to update the operation parameter dynamic self-optimization module.

Images