Graphic recognition method and obstacle avoidance system for intelligent agricultural robots

Abstract

The present invention provides a graphic recognition method and obstacle avoidance system for intelligent agricultural robots, which relates to the technical field of image pattern recognition. By introducing polarization imaging, the present invention effectively suppresses water droplet highlights, avoids misjudging optical artifacts as diseases, and significantly improves recognition accuracy. For low-confidence targets, the original micro-area thermal excitation physical review mechanism can confirm their true physical properties online, solving the problem of recognition uncertainty; in addition, the dynamic risk volume constructed based on multi-dimensional information, combined with the adaptive speed control of path risk foresight, realizes differentiated intelligent obstacle avoidance for rigid and flexible obstacles. While ensuring safety, this method greatly improves the robot's travel efficiency and smoothness in complex environments.

Description

Technical Field

[0001] The present invention relates to the technical field of image pattern recognition, and in particular to a pattern recognition method and obstacle avoidance system for an intelligent agricultural robot. Background Art

[0002] Currently, mainstream intelligent agricultural robots rely mainly on visual perception technology based on visible light (RGB) or depth (RGB-D) sensors. By being equipped with high-performance computing platforms and running deep learning algorithms, especially convolutional neural networks (CNN), these robots can segment and identify crop organs (such as fruits, leaves, and stems) at the pixel level, thereby achieving maturity and yield assessments or preliminary detection of pests and diseases. In terms of navigation and obstacle avoidance, path planning technologies based on 3D point cloud data, such as using A 、D Algorithms such as RANDOM LITE (RANDOM LITE) or Rapidly Exploring Random Trees (RRT) have become standard for searching collision-free paths within constructed environmental maps. The development of these technologies has significantly advanced the automation of agricultural production and laid the foundation for unmanned farms. However, in real, dynamic, and complex agricultural production environments, especially in scenarios such as high-density strawberry cultivation, the limitations of existing technologies are becoming increasingly prominent, becoming a core bottleneck hindering further improvements in operational efficiency, reliability, and intelligence.

[0003] In the prior art, the publication number is CN118587677A, and the name is an obstacle detection and obstacle avoidance method and system for automatic driving agricultural machinery. The method includes: collecting front images through multiple cameras; detecting the pressure exerted on the baffle by objects in front of the agricultural machinery through a pressure sensor; performing secondary processing on the collected images to obtain images within a preset close range in front of the agricultural machinery; identifying the secondary processed images through a pre-trained neural network detection device; when a preset type of obstacle is identified, the three-dimensional coordinates of the obstacle are calculated using the position information of the obstacle in multiple images collected by multiple cameras, and the obstacle avoidance path is calculated based on the three-dimensional coordinates of the obstacle; when no preset type of obstacle is identified, but the pressure detected by the pressure sensor exceeds a threshold, it is determined that an obstacle appears in front, and an obstacle avoidance operation is performed according to a preset strategy.

[0004] Despite significant progress in existing technologies, they still have profound deficiencies in the robustness of perception and decision-making in complex environments:

[0005] First, at the visual recognition level, traditional RGB imaging technology is highly susceptible to interference from changes in ambient lighting. In particular, in the morning or after irrigation, water droplets and dew on leaves and fruit create strong specular highlights. These highlights appear as bright white spots in the image, and their morphological characteristics are highly similar to the early symptoms of various fungal diseases, such as powdery mildew. This makes neural network models that rely on image features highly susceptible to confusion, generating a large number of false positives. This not only triggers erroneous decisions such as unnecessary pesticide spraying, wasting resources, but also affects accurate judgment of actual disease trends due to data contamination.

[0006] 2. Secondly, in terms of obstacle avoidance and path planning, existing technologies typically treat all detected obstacles as rigid entities with defined boundaries and employ a unified avoidance strategy. However, obstacles in agricultural environments are highly heterogeneous: support poles are rigid and must be strictly avoided; while crop leaves and vines are flexible, and light contact is generally acceptable. Excessive avoidance can significantly compress the traversable space, leading to overly conservative path planning, significantly reducing operational efficiency, and even causing path planning failure in densely populated areas.

[0007] 3. Furthermore, for those "suspicious targets" whose attributes cannot be determined due to low recognition confidence, the existing system lacks an effective online review and confirmation mechanism and can only classify them as high-risk obstacles. This strategy further deteriorates the robot's ability to pass.

[0008] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0009] The purpose of the present invention is to provide an intelligent agricultural robot pattern recognition method and obstacle avoidance system to solve the problems raised in the above background technology.

[0010] To achieve the above object, the present invention provides the following technical solutions:

[0011] The intelligent agricultural robot pattern recognition method includes the following specific steps:

[0012] Step S1: capturing at least two images of a target scene in different polarization directions, and calculating and generating a corrected image that suppresses specular reflection highlight artifacts based on the polarization characteristics of the images;

[0013] Step S2: Inputting the corrected image into a pre-trained neural network model to identify crop targets and output the category of each crop target and its corresponding recognition confidence; when the recognition confidence of a confirmed target is lower than a preset confidence threshold, the area where the crop target is located is determined to be a highly suspicious area;

[0014] Step S3: Performing a physical property review operation on the highly suspicious area, which includes: collecting a temperature change curve of the area through micro-area thermal excitation and high-frequency temperature measurement; fitting the temperature change curve with an exponential decay model to solve the thermal relaxation time constant that characterizes its thermodynamic properties, and finally confirming the true physical properties of the highly suspicious area based on this;

[0015] Step S4: Based on the principle of binocular stereo vision, the three-dimensional spatial coordinates of the robot of the confirmed target in its own coordinate system are calculated; and the confirmed physical properties, recognition confidence, and thermal relaxation time constant information of the crop target are integrated to dynamically construct a risk volume representing the collision risk and its corresponding risk coefficient for the confirmed target;

[0016] Step S5: Map the risk volumes and risk coefficients of all confirmed targets in the workspace onto a rasterized cost map to construct a decision space that comprehensively reflects the distribution of environmental risks; on this cost map, use the path planning algorithm to search and plan an optimal driving path with the minimum total cost, and generate control instructions based on this to drive the robot to move.

[0017] A system for pattern recognition and obstacle avoidance of an intelligent agricultural robot, the system being used to execute the method for pattern recognition of an intelligent agricultural robot, comprising:

[0018] Data acquisition module: used to acquire at least two images of the target scene in different polarization directions, and calculate and generate a corrected image that suppresses specular reflection highlight artifacts based on the polarization characteristics of the images;

[0019] A region determination module is configured to input the corrected image into a pre-trained neural network model to identify crop targets and output the category of each crop target and its corresponding recognition confidence level. When the recognition confidence level of a confirmed target is lower than a preset confidence threshold, the region where the crop target is located is determined to be a highly suspicious region.

[0020] Confirmation module: for performing a physical property review operation on the highly suspicious area, which includes: collecting the temperature change curve of the area through micro-area thermal excitation and high-frequency temperature measurement; fitting the temperature change curve with an exponential decay model to solve the thermal relaxation time constant that characterizes its thermodynamic properties, and finally confirming the true physical properties of the highly suspicious area based on this;

[0021] Risk coefficient calculation module: This module is used to calculate the three-dimensional spatial coordinates of the robot in its own coordinate system based on the principle of binocular stereo vision. It also integrates the confirmed physical properties, recognition confidence, and thermal relaxation time constant information of the crop target to dynamically construct a risk volume representing the collision risk and its corresponding risk coefficient for the confirmed target.

[0022] Result output module: used to map the risk volume and risk coefficient of all confirmed targets in the workspace onto a rasterized cost map to construct a decision space that comprehensively reflects the distribution of environmental risks; on this cost map, a path planning algorithm is used to search and plan an optimal driving path with the minimum total cost, and based on this, control instructions are generated to drive the robot to move.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. First, by introducing polarization imaging technology, specular highlights caused by water droplets and dew are effectively suppressed at the source, greatly reducing the probability of visual recognition models misclassifying optical artifacts as defects, significantly improving the accuracy and robustness of target recognition in complex lighting and humid environments.

[0025] 2. Secondly, the present invention combines a micro-area thermal excitation physical verification mechanism to provide an online, reliable means of identifying the physical properties of low-confidence suspicious targets. It can accurately distinguish between real biological tissue and non-biological interference, avoiding overly conservative obstacle avoidance due to uncertainty.

[0026] 3. Finally, by constructing a dynamic risk volume based on multi-dimensional information fusion and combining it with an adaptive speed control strategy based on path risk foresight, the robot can adopt differentiated and more intelligent obstacle avoidance behaviors for obstacles with different physical properties. While ensuring safety, it greatly improves the traffic efficiency and operation smoothness in high-density and complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of the overall method flow of the present invention;

[0028] Figure 2 This is a block diagram of the system module of the present invention;

[0029] Figure 3 This is a schematic diagram of the intelligent agricultural robot of the present invention. DETAILED DESCRIPTION

[0030] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0031] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0032] Example 1:

[0033] See also Figure 1 and Figure 3 , the present invention provides a technical solution:

[0034] The pattern recognition method for intelligent agricultural robots is applied to the obstacle avoidance system. The specific steps include:

[0035] Step S1: capturing at least two images of a target scene in different polarization directions, and calculating and generating a corrected image that suppresses specular reflection highlight artifacts based on the polarization characteristics of the images;

[0036] Further explanation: A polarization camera is used to obtain images of four polarization channels of 0°, 45°, 90°, and 135° in the same scene, and the linear polarization degree of each pixel is calculated based on the images. , to generate an optical artifact mask corresponding to the specular reflection area ; and using the optical artifact mask For the original collected RGB image Perform pixel-level weighted correction to obtain a corrected image with suppressed specular highlights. The specific implementation is as follows:

[0037] The polarization camera in the robot vision system can capture images of four polarization channels in a single exposure. The system calculates the linear polarization degree of each pixel from these images. Since the specular reflection light from the water droplet surface is strongly polarized, and the diffuse reflection light from the plant tissue is basically non-polarized light, The value can effectively quantify the intensity of the specular reflection. The value exceeds the preset polarization threshold , the pixel is identified as an optical artifact. All these artifact points constitute an optical artifact mask The system uses this mask to perform weighted correction on the original RGB image, significantly darkening the highlight area, restoring the true color and texture information that is obscured underneath, and finally outputs the corrected image;

[0038] Further explanation: The original RGB image The pixel-level weighted correction is performed, and the correction weight is based on the linear polarization degree of each pixel. and polarization threshold The adaptive correction weight is calculated dynamically based on the relationship between , in order to achieve differentiated suppression of specular reflections of different intensities. The specific implementation contents are as follows:

[0039] For optical artifact masks For each pixel marked as 1, the correction strength is determined by an adaptive correction weight The weight is calculated through a normalized exponential gain function to achieve refined and nonlinear suppression of highlight artifacts.

[0040] For adaptive correction weight The calculation formula is constructed through a standardized exponential function; the formula is based on the linear polarization degree of the pixel Exceeding the polarization threshold The degree of is the independent variable, the exponential function is used to amplify the difference, and then the theoretical maximum input is divided by the function; that is, The maximum value corresponding to is normalized to smoothly map the correction weight to the [0,1] interval;

[0041] Adaptively correct weights The calculation formula is represented as: ;

[0042] Among them, the adaptive correction weight The value range is [0,1]. This value determines the suppression strength of the highlight artifact of the current pixel; is the linear polarization degree of the current pixel, collected by the polarization camera , , , The four channel images are calculated and their value range is [0,1]. It is a preset polarization threshold used to preliminarily determine whether a pixel is an optical artifact. It is a constant pre-calibrated according to the ambient light and target characteristics. Its value range is (0, 1). In this embodiment, it is between 0.7 and 0.9. is the correction sensitivity coefficient, a constant greater than 0, used to adjust the severity of the change in correction weight with polarization degree; e is the base of the natural logarithm; the correction formula for the final output corrected image is represented as: ;in, is the corrected image of the pixel coordinate (x, y) after applying the adaptive correction weights;

[0043] The suppression strength of optical artifacts is proportional to the polarization degree of the artifacts themselves, and is a nonlinear relationship. Using an exponential function can effectively amplify The difference after exceeding the threshold value significantly enhances the suppression effect of strong reflections, while the suppression of weak reflections is relatively mild. and How to set the weights so that they are always in a uniform and meaningful interval of [0,1], ensuring the robustness of the algorithm;

[0044] when The closer the output is to 0: the current pixel Value and The closer it gets, the weaker the reflection. This means the system has detected a weak specular reflection signal, which may be a very small water droplet or just a weak reflection from the wax layer on the surface of the crop leaf at a specific angle.

[0045] Image Fidelity: The trend is towards maximizing the preservation of original image information. Because the correction weight is small, the brightness of that pixel is minimally suppressed. This prevents useful surface textures (such as leaf veins and hairs) from being misidentified as artifacts and overly darkened, thereby improving the fidelity and detail richness of the corrected image.

[0046] Parameter change logic reasoning: when Infinitely close to When the exponential term in the numerator of the formula Approaching , causing the entire numerator to approach 0, so Approaching 0. This design ensures that the system can intervene in the most "gentle" way in the critical area between artifacts and real textures, and its technical effect is to prioritize not damaging image details.

[0047] when The closer the output is to 1: the current pixel The closer the value is to 1, the clearer it is. This is a clear indication of a very strong specular reflection source, such as a large, plump, smooth water droplet facing the light source, whose reflected light is almost completely polarized.

[0048] The trend of "artifact suppression efficiency" is towards maximizing the suppression of optical artifacts. When the correction weight reaches its maximum value, the system applies the strongest brightness suppression to that pixel, striving to completely eliminate the effects of highlights and reveal key information such as possible underlying lesions.

[0049] Parameter change logic reasoning: when When it approaches 1, the numerator of the formula Approaching the denominator , so that the value of the entire fraction approaches 1. In addition, if the correction sensitivity coefficient is increased , the exponential term will grow more rapidly, making It can grow faster from 0 to 1. This design ensures that the system can clear it in the most decisive way when faced with strong interference.

[0050] The core innovation of this scheme is to introduce adaptive correction weights that are positively correlated with the artifact intensity. The system is able to achieve differentiated and appropriate corrections. The technical effect is a significant improvement in the quality of the corrected image: it can both strongly remove severe highlight interference and gently process weak reflection signals that may be confused with real textures, thus achieving an optimal balance between "removing artifacts" and "preserving authenticity." This high-quality image input directly reduces the recognition difficulty of the subsequent neural network model and is a key step in improving the robustness and accuracy of the entire system.

[0051] It should be noted that this embodiment aims to verify the beneficial effect of the proposed "image correction method based on adaptive correction weight" in improving the accuracy of identifying early lesions of strawberry powdery mildew compared with the "traditional fixed threshold correction method" and "uncorrected original image".

[0052] The experiment was conducted in a semi-automated strawberry greenhouse, with the ambient temperature controlled at 22±2°C and the humidity maintained at 85%±5%RH. The experiment was conducted between 6:00 and 7:00 in the morning, when dewdrops of varying sizes are commonly condensed on the leaves and fruit surfaces of the strawberry plants, fully simulating complex optical artifact scenes. The experimental platform was an independently developed agricultural robot, whose vision system was equipped with a Tornado-CAM model POL-5000 snapshot polarization camera, which can capture 5-megapixel images of four polarization channels at 0°, 45°, 90°, and 135° in a single exposure and simultaneously output the corresponding RGB images. The robot's processing core uses NVIDIA-Jetson-AGX-Orin, running pre-deployed image processing algorithms and neural network models.

[0053] Six "Hongyan" strawberry plants, in good health but confirmed to be infected with early-stage powdery mildew, were selected as test subjects. Agricultural technicians manually marked and microscopically verified 100 early-stage powdery mildew lesions (1-3 mm in diameter, covered with a white, powdery mold layer) on these plants as "true positives," and 100 dewdrops similar in size and shape to early-stage lesions as "false positives."

[0054] Detailed implementation process:

[0055] The robot followed a pre-set path, moving in front of six test plants in sequence and hovering 40 cm above the canopy. For each target area containing a known lesion or dew drop, the robot's vision system performed the following operations:

[0056] 1. Data acquisition: The polarization camera shoots the target area to obtain the original RGB image And images of the four polarization channels , , , .

[0057] 2. Image processing: The same set of collected raw data is fed into three different processing flows in parallel:

[0058] Group A is the control group 1-original image: without any processing, the original RGB image is directly Feed into subsequent recognition model.

[0059] Group B is the control group 2-Traditional method: using fixed threshold correction method. First calculate , set a fixed polarization threshold and fixed modification weights For all The brightness value of the pixel is multiplied by 1-0.9=0.1 to generate the corrected image.

[0060] Group C is the method of the present invention: the adaptive weight correction method proposed by the present invention is used. , and set the same polarization threshold But the modified weight By the formula Dynamic calculation, where the sensitivity coefficient is corrected Set to 5.0. Use this dynamic weight to correct the image and generate the final corrected image .

[0061] The images output by the three processing steps above were fed into a pre-trained YOLOv5s object recognition model specifically trained for strawberry pests and diseases. The model's recognition results for 100 true-positive lesions and 100 false-positive dewdrops were recorded.

[0062] 4. Data quantification and analysis: Statistics are collected on the recognition results, mainly evaluating the following key performance indicators:

[0063] True lesion detection rate (TPR): the number of correctly identified lesions / total number of lesions.

[0064] Dewdrop false positive rate (FPR): the number of dewdrops mistakenly identified as lesions / total number of dewdrops.

[0065] Image Entropy: Calculates the average entropy of each group of output images. A higher entropy generally indicates a richer image with more detail and texture information.

[0066] Peak Signal-to-Noise Ratio (PSNR): Using an "ideal image" with finely crafted highlights removed as a benchmark, the PSNR value of each corrected image is calculated. A higher PSNR value indicates less image distortion and higher quality.

[0067] Structural Similarity (SSIM): Using the "ideal image" as a benchmark, the SSIM value of each corrected image is calculated. SSIM better reflects the fidelity of the image in terms of structure, brightness, and contrast. Values ​​closer to 1 are better.

[0068] Comprehensive evaluation F1 score (F1-Score): A comprehensive evaluation index calculated by combining TPR and FPR. ,in , The higher the F1 score, the better the overall performance of the model.

[0069] By repeating the test and conducting data statistics on 200 target points of these 6 samples, the following comparative data table was finally obtained.

[0070] Table 1 Study on the effectiveness of image correction methods based on adaptive correction weights:

[0071]

[0072] It can be clearly seen from the above table data:

[0073] 1. Recognition Performance: The method of the present invention achieved the highest true spot detection rate of 96% and the lowest false dew drop detection rate of 4%. This directly demonstrates that the adaptive weighting method produces higher image quality, providing a cleaner and more reliable input. In contrast, the original image was severely affected by highlights, resulting in a large number of false positives and missed positives. While the traditional fixed threshold method offers some improvements, its one-size-fits-all approach still results in loss of detail in some spots (TPR fails to reach the maximum) or inadequate correction of weakly reflective dew drops (FPR remains at 22%). The proposed method achieved an F1 score of 0.96, far exceeding both control groups, demonstrating its comprehensive performance advantages in practical applications.

[0074] 2. Image Quality: The proposed method achieved a significantly higher average image information entropy of 7.05 than the traditional method (6.58), demonstrating that it suppresses highlights while better preserving image detail and texture. This is crucial, as early-stage lesion detection often relies on subtle texture changes. Furthermore, the highest PSNR (34.7dB) and SSIM (0.95) values ​​objectively quantify that the image output by the proposed method is closest to the "ideal noise-free image," with minimal distortion and the highest fidelity.

[0075] In summary, this embodiment fully and powerfully demonstrates through specific and quantifiable experimental data that the "image correction method based on adaptive correction weights" proposed in the present invention can significantly improve the quality of the corrected image in complex optical environments compared to the existing technology.

[0076] Step S2: Inputting the corrected image into a pre-trained neural network model to identify crop targets and output the category of each crop target and its corresponding recognition confidence; when the recognition confidence of a confirmed target is lower than a preset confidence threshold, the area where the crop target is located is determined to be a highly suspicious area;

[0077] Further explanation: Further explanation: The corrected image Input into a pre-trained neural network model for crop target recognition, and obtain the category of each recognized target and its corresponding recognition confidence output by the neural network model ; and the recognition confidence With a preset confidence threshold When compared, Below When the crop target is identified, the area where it is located is determined to be a highly suspicious area. The specific implementation contents are as follows:

[0078] After obtaining the corrected image with the highlight interference removed After that, it is further fed into a pre-trained deep learning model. This model is responsible for analyzing the image content, detecting and classifying various crop targets of interest. In this embodiment, crop targets include but are not limited to "strawberry fruit", "leaves", "powdery mildew spots", etc. For each detected crop target, the model not only gives its category judgment, but also gives a recognition confidence score between 0 and 1. , the recognition confidence represents the "confidence" of the model in its judgment; when Below a preset When the system determines that the model's judgment is not reliable enough and there is ambiguity, it marks the crop target area as a "highly suspicious area" and records its location information in the image, waiting for further accurate diagnosis.

[0079] It should be noted that the pre-trained deep learning model in this embodiment is YOLOv5 or Faster-R-CNN; the confidence threshold The initial value is 0.75.

[0080] Further explanation: The confidence threshold for determining highly suspicious areas is a dynamic confidence threshold for each category obtained by querying or calculating from a preset category-threshold mapping table based on the specific category of each identified target output by the neural network model. ;

[0081] Specifically: The system has a built-in "category risk-threshold mapping table" that pre-sets a basic confidence threshold for each possible crop target category (including fruit, leaves, spots, etc.) and a category risk factor ; Dynamic confidence threshold for final comparison , which is calculated based on these two parameters through a specific function;

[0082] Dynamic confidence threshold The calculation formula is constructed through a transformation model based on the S-type logistic function. The model is based on the preset category-based confidence threshold. As the center, using the category risk coefficient To adjust the shape of the S-curve, an abstract risk level is smoothly and nonlinearly converted into a specific dynamic confidence threshold in the interval (0,1); dynamic confidence threshold The characterization formula is: ;

[0083] in, is the dynamic confidence threshold, and its range is , in this embodiment, it is set to (0.5, 0.95), which is used to identify the confidence level of the neural network output. Make a comparison. It is the preset highest threshold, representing the system's highest "alertness" level for the most critical and ambiguous targets; It is the preset minimum threshold, representing the system's basic "trust" level for the simplest and clearest goals; : Class risk coefficient, a constant greater than 0, preset in the mapping table. This coefficient represents the "diagnostic importance" or "inherent ambiguity" of a target class. For example, the value of "powdery mildew lesions" is It will be much larger than the "ripe strawberry" . It is the numerical code corresponding to the currently identified target category; "leaves", "fruits", "spots", etc. are converted into numerical values ​​through one-hot encoding or label encoding. It is the center point or turning point of the category code, which is used to locate the area where the threshold changes most dramatically on the S-shaped curve; e is the base of the natural logarithm.

[0084] This ensures that even for the simplest targets, the system retains a minimum review standard to prevent obvious errors in extreme cases from being overlooked. This prevents the threshold from being too high, causing all targets to be judged suspicious and paralyzing the review mechanism. This dynamic range ensures that the threshold is both discriminatory and remains within a practical and operational range.

[0085] Different categories of crop targets have different risks of misjudgment and inherent difficulty of identification, so they should not be measured using a unified standard. The S-type logistic function is suitable for mapping discrete category information to a continuous threshold space. Category risk coefficient It directly controls the steepness of the curve and perfectly quantifies the core idea that "the higher the risk, the stricter the standard", providing solid mathematical and logical support for the setting of the threshold.

[0086] when The closer the output When: The target category identified by the neural network has a lower category risk coefficient This corresponds to targets with distinct features and simpler backgrounds, such as “ripe, unobstructed strawberry fruit”.

[0087] The changing trend in "operational efficiency" is toward maximizing the smoothness and speed of the operational process. This is because the lower the confidence threshold, the more likely the recognition results for such targets will be directly accepted, eliminating the need for time-consuming physical verification procedures, significantly speeding up the robot's overall operational cycle.

[0088] When the category risk factor The smaller the value, the flatter the S-shaped function curve becomes. Whatever the value, the exponential term in the denominator approaches , so that the entire fractional part approaches In order to make it close to ,Adjustment Or directly use a simplified linear model. But in the S-type model framework, This means that the threshold is insensitive to changes in category and will stabilize at a low level. This design ensures that the system adopts a "quick pass" strategy for low-risk, easily identifiable targets, and its technical effect is to prioritize efficiency.

[0089] when The closer the output When: The identified target category has a higher category risk coefficient This clearly points to targets that are crucial to agricultural production and easily misjudged, such as "early powdery mildew lesions" and "tiny insect pests."

[0090] The changing trend of "setting the technical goal as diagnostic accuracy" is to maximize the reliability and precision of diagnostic results. Because the confidence threshold is raised to the highest level, uncertain judgments are intercepted and forced into a more precise physical review process.

[0091] When the category risk factor The larger the value, the steeper the S-shaped function curve becomes. The farther away the value , the exponential term in the denominator will become larger or smaller, making the entire fractional part closer to 0 or , so that The closer to its extreme value or Specifically, for the high-risk category, The value is designed to make the exponential term tend to the infinity side, causing the fraction to approach 0, and finally The closer Its technical effect is to prioritize ensuring that no key abnormal points are missed.

[0092] The core value of this solution lies in its transformation of the post-identification "triage" process from a static, extensive management model to a dynamic, sophisticated, risk-based intelligent decision-making process. By setting different dynamic thresholds for different target categories, the system optimizes computing resources, focusing its most valuable precision diagnostic capabilities on resolving the most critical and ambiguous issues. This not only significantly improves the detection rate and accuracy of key targets such as pests and diseases, but also avoids wasting time on simpler targets, thereby significantly improving the robot's overall operational efficiency while maintaining high reliability.

[0093] Step S3: Performing a physical property review operation on the highly suspicious area, which includes: collecting a temperature change curve of the area through micro-area thermal excitation and high-frequency temperature measurement; fitting the temperature change curve with an exponential decay model to solve the thermal relaxation time constant that characterizes its thermodynamic properties, and finally confirming the true physical properties of the highly suspicious area based on this;

[0094] Further explanation: Further explanation: Triggering a review operation on the highly suspicious area, the review operation includes controlling a microfocus laser to emit a laser pulse of preset energy to the highly suspicious area, and synchronously using a microbolometer to monitor the temperature change curve of the area after the laser pulse. , through the temperature change curve Perform exponential decay model fitting to extract the thermal relaxation time constant that characterizes the heat dissipation rate , and according to the The actual physical properties of the highly suspicious area are finally confirmed by the numerical value; the specific implementation content includes:

[0095] When an area is judged to be "highly suspicious", the robot will activate the precision detection head integrated on its end effector. The detection head is aimed at the area. First, the microfocus laser emits a short and low-energy laser pulse. In this embodiment, "short" and "low energy" are "1ms" and "μJ level" respectively; then the crop target point is heated slightly and non-destructively. At the same time, the coaxial high-frequency microbolometer begins to accurately record the temperature of the point from the peak after heating at a sampling rate of kHz. The cooling process over time forms a temperature change curve Due to the large difference in thermodynamic properties between water (fast evaporation) and plant tissue (slow conduction), their cooling rates are also completely different. The data were fitted with an exponential model of Newton's law of cooling, allowing the thermal relaxation time constant to be accurately calculated. This constant directly quantifies the rate of heat dissipation. By comparing it with the preset baseline values ​​of "water" and "plant tissue", a final and definitive judgment can be made on the true physical properties of the suspicious point.

[0096] Further explanation: Based on the As well as the pre-calibrated reference time constants of water and healthy plant tissue, a continuously changing physical reality index in the interval (0,1) is calculated through a nonlinear S-type logistic function. , to quantify the degree of confidence that the area is real plant tissue; physical authenticity index The larger the value, the higher the probability that the area has plant tissue with normal physiological activity. The specific contents include:

[0097] Upgrading the output of physical verification from a binary label (water / tissue) to a continuous quantitative indicator - the Physical Authenticity Index The calculation of this index relies on two pre-calibrated reference parameters in the laboratory: the typical thermal relaxation time constant of pure water is set to ; and the typical thermal relaxation time constant for a healthy, well-watered target crop is assumed to be ;

[0098] Physical authenticity index of this embodiment The calculation formula is based on a standard S-type logistic function; specifically, the measured thermal relaxation time constant In a and It is positioned in a defined coordinate system and nonlinearly mapped to a probability or credibility space of (0,1) through an adjustable sensitivity parameter, thereby obtaining a continuous indicator that can smoothly reflect the transition from "absolute non-organization" to "absolute real organization";

[0099] Physical Reality Index The calculation formula is represented as: ;

[0100] in, It is a physical authenticity index with a valid value range of (0,1). This index quantitatively describes the credibility of the measured object being a real plant tissue; It is the thermal relaxation time constant of the currently highly suspicious area actually measured and calculated through physical verification operations. is the time constant for determining the center point, which represents the theoretical dividing line between the "water" and "plant tissue" states; it is calculated from a pre-calibrated reference value: k1 is the exponential sensitivity coefficient, which is a constant greater than 0 and is used to adjust right The steepness of the response to the change. The larger the k1, exist The smaller the k1, the smoother the transition and the more vague the description of the intermediate state.

[0101] Physically, This directly reflects the thermodynamic properties of the substance. Logically, the closer an object's thermodynamic properties are to those of a healthy plant, the less likely it is to be a water droplet.

[0102] when The closer the output is to 0: The smaller the , the closer to or even smaller than The higher the probability that the “highly suspicious area” is actually a water droplet or other non-living matter with an extremely fast heat dissipation rate;

[0103] The trend toward "setting the technical goal as false alarm reduction" is towards maximizing the rejection of false alarms caused by optical artifacts. The system confirms with high confidence that the artifact is a physical artifact, allowing it to be decisively excluded from subsequent risk assessment and work planning.

[0104] when The smaller the When the exponential term becomes a larger positive number, resulting in tends to infinity, the denominator tends to infinity, so The closer it is to 0. This design ensures that when the physical measurement results clearly point to "non-plant" characteristics, the system can give a deterministic negative judgment close to zero. Its technical effect is to provide a clean, error-free input for subsequent decision-making.

[0105] when The closer the output is to 1, the Bigger than , approaching or even exceeding the healthy The higher the probability that the area contains plant tissue with normal physiological activity;

[0106] The changing trend of "setting the technical goal to suppress missed detection rate" is as follows: the system uses physical means to save a real target (such as an early, inconspicuous lesion) that is given a low confidence score due to atypical morphology, poor lighting, etc., to ensure that it will not be missed.

[0107] when Bigger than When the exponential term Become a negative number with a larger absolute value, resulting in tends to 0, and the denominator tends to 1, so The closer it approaches 1. The larger the exponential sensitivity coefficient k1, the faster this convergence. This design ensures that when the physical measurement results clearly point to "plant" characteristics, the system can provide a definitive, positive judgment close to 1. The technical effect is to provide decisive and irrefutable evidence for the final identification of key targets.

[0108] This solution will physically verify by outputting a continuous physical authenticity index , the system obtains unprecedented quantitative information about the nature of the target. This not only can eliminate the interference of artifacts such as water droplets with a high probability, but also provides a key, quantitative input for subsequent steps for more refined risk assessment. For example, a A slightly dehydrated leaf target of 0.7 and a The healthy leaf target of 0.95 can be assigned different risk weights during path planning. This makes the entire system's decision chain more data-driven, more intelligent, and closer to the complex reality of organisms.

[0109] This embodiment introduces the "physical authenticity index" The core innovation of the physical reality index is a continuous value in the interval (0,1). The value is mapped to a standardized "credibility" or "healthiness" scale. This upgrade from "qualitative classification" to "quantitative assessment" enables the system to not only distinguish "authenticity" but also assess "status," providing unprecedented, higher-dimensional input information for subsequent risk assessment and decision-making.

[0110] Step S4: Based on the principle of binocular stereo vision, the three-dimensional spatial coordinates of the robot of the confirmed target in its own coordinate system are calculated; and the confirmed physical properties, recognition confidence, and thermal relaxation time constant information of the crop target are integrated to dynamically construct a risk volume representing the collision risk and its corresponding risk coefficient for the confirmed target;

[0111] Further explanation: At least two cameras with a preset baseline distance B1 in space are used to synchronously acquire an image pair of a crop target, and the disparity d1 of the target with confirmed attributes in the left and right images is calculated by a stereo matching algorithm, and combined with the camera focal length and the baseline distance B1, and calculate its three-dimensional space coordinates in the robot's own coordinate system ; Then, the category attributes of the crop target and the neural network recognition confidence are integrated and the physical reality index , through a preset fusion model, a risk volume is generated for the crop target and its corresponding risk factor The specific implementation contents are as follows:

[0112] Three-dimensional space coordinates The method of obtaining is: by performing stereo calibration on the binocular camera system, accurately solving the intrinsic parameter matrix K2 of each camera, the distortion coefficient D1, and the relative posture relationship between the two cameras, that is, the rotation matrix R and the translation vector T, where the modulus of the translation vector is the baseline distance Next, the hand-eye calibration process is used to further determine the rigid transformation matrix between the camera coordinate system and the robot body coordinate system. ; Specifically represents the homogeneous transformation matrix required to transform a point from the camera coordinate system to the robot coordinate system; "cam" and "robot" represent the camera and robot respectively;

[0113] Secondly, when the robot is running, the system enters the online solution phase. Whenever a pair of left and right images are synchronously acquired, efficient stereo correction is immediately performed using the calibrated parameters to generate an image pair with pixel rows aligned. Subsequently, a dense disparity map is calculated on the corrected images using the semi-global block matching (SGBM) algorithm. For a recognized target, its pixel coordinates in the left image are is used to query the disparity map, The "L" in the equation represents the "left", so that the accurate parallax d1 is obtained. Based on the principle of triangulation and the focal length of the camera, and the baseline distance B1, by the formula Calculate the depth of the target in the camera coordinate system and obtain its complete three-dimensional coordinates ; The subscript index in is the abbreviation of "cam"; finally, through matrix multiplication , seamlessly converting the coordinates into the robot's own three-dimensional space coordinates.

[0114] After the target's true identity is confirmed, the system enters the three-dimensional space modeling and risk assessment stage; first, using the binocular vision system, the system calculates its precise three-dimensional space coordinates. ; Next, the system dynamically generates a "risk volume" for this target in three-dimensional space ; Define “risk volume” Based on various attributes and associated risk factors The geometric bounding box is determined dynamically by a comprehensive evaluation function based on the various properties of the target; these properties include: the rigidity / flexibility represented by the physical properties of the target, the recognition confidence of the neural network for its classification , if verified, also includes the physical authenticity index ;

[0115] In this embodiment, a branch that is identified as a rigid branch and has a risk factor of The higher the crop target, the will be larger; and a leaf that is identified as flexible and The lower the target, the It will get smaller.

[0116] It should be noted that: after obtaining the physical authenticity index that characterizes the physical properties Then, it is used as the key modulation factor to determine the inherent category risk corresponding to the crop target category output by the neural network. and recognition confidence Perform cross-weighted fusion to generate a risk volume ,This volume uniformly quantifies the final risk level of each identified target;

[0117] Risk Volume The calculation formula is based on a "hierarchical fusion model under evidence modulation". The model first divides the inherent risk category into a preset category representing the inherent risk level of the target. and the recognition confidence given by the neural network Multiply, then, the formula introduces the physical reality index representing the “ground truth” obtained by the physical review in step S3 , as the decisive modulation coefficient of the aforementioned risk assessment value, by multiplying the three together, we finally get a standardized risk volume that integrates all evidence;

[0118] For each identified crop target, its risk volume Calculated using the following formula:

[0119] ;

[0120] in, is the risk volume, which ranges from [0,1). This is a final, standardized risk measure used to guide all subsequent decisions; is the inherent category risk, with a value range of [0,1). This is a pre-set dimensionless parameter stored in the system lookup table. Based on agronomic knowledge, it assigns an inherent risk level to each crop target category (e.g., "ripe fruit," "leaves," "powdery mildew lesions," "aphids"). In this example, The value of 0.9 represents powdery mildew lesions, while It is set to 0.05 to represent healthy leaves; It is the recognition confidence, whose value range is (0,1). It is output by the neural network model in step S2 and represents the "confidence level" of category judgment. is the physical authenticity index, which has a value range of (0,1) and is calculated by the physical property verification operation in step S3. For targets that have not been physically verified, that is, Above dynamic threshold The system will assign a default confidence score of 0.95 to the target, indicating that in the absence of contrary evidence, it tends to believe high confidence judgments.

[0121] when When the output approaches 0, this happens for three main reasons:

[0122] 1): If the value is low, it identifies healthy leaves that are harmless; 2) The value is low and the neural network model is not sure what it is; 3) The value is low, and physical verification confirms that it is a water drop or artifact.

[0123] when When the output approaches 1: 、 and All three parameters simultaneously approach 1. This indicates that the system has identified a dangerous target, and physical verification has confirmed with high confidence that it is a real biological tissue. A critical response is triggered with the highest priority and the highest certainty. This is an "alarm" level signal, indicating that the system has discovered a serious, triple-verified threat that cannot be ignored and requires immediate and decisive action.

[0124] Because it is a multiplicative relationship, only when all links in the chain of evidence provide top-level support, Even if any one of the links is slightly lacking; Only 0.8, or Only 0.9, the final This "unanimous approval" mechanism ensures that the system will only issue the highest level of alert when it has overwhelming evidence. Its technical effect is to greatly improve the robustness and reliability of key decisions and avoid the occurrence of "false alarms";

[0125] The formula, at its core, isn't a simple arithmetic multiplication of three undifferentiated values; rather, it's a physical and logical process based on conditional probability and the chain of evidence, quantitatively modeling the abstract concept of "risk." It resolves the problem of "conceptual differences": the multiplication of three clearly related conditional probabilities produces a joint probability with clear meaning. It embodies the logic of the "chain of evidence": this model perfectly embodies the logic of "veto" and "evidence accumulation."

[0126] Veto: If physical measurements confirm the target is false , then no matter how convinced the neural network model is that it is a high-risk target , , the final joint probability will approach 0. Only when the three conditions of "high inherent risk," "accurate neural network model judgment," and "physical confirmation" are simultaneously met with high probability will their product be high. A decrease in the probability of any link will lead to a proportional decrease in the final risk confidence. This demonstrates that the strength of a chain of evidence depends on its weakest link.

[0127] Therefore, the formula for calculating the “risk volume” is Its rationale stems not from simple numerical calculations but rather from its mathematical modeling of the core question, "the joint probability of a target posing a real risk," under the assumption of conditional independence. It logically and self-consistently integrates observational evidence from diverse sources and physical / conceptual dimensions within the unified framework of probability theory, resulting in a quantitative indicator that is informative, logically defensible, and can directly guide subsequent decision-making.

[0128] Further explanation: The risk factor It is calculated through a cross-weighted fusion mechanism; in this mechanism, it is used to calculate the recognition confidence The function of weight, whose independent variable is the physical reality index ; and used to calculate A function of the weight, whose independent variable is the recognition confidence , thereby achieving dynamic and mutual adjustment between the weights of the two information sources. The specific implementation content includes:

[0129] Construct a cross-weighted fusion model that can reflect the mutual constraints and complementary relationships between information sources. The core of the model is two weight functions: and , which are used to determine the recognition confidence and physical reality index Calculating the risk factor Dynamic weight when

[0130] Risk Factor The calculation formula is constructed through a cross-weighted linear combination model. This model converts the normalized recognition confidence and physical reality index As the basic risk factor, the weight of each factor is not constant, but is dynamically generated by another factor through a decreasing exponential function. Specifically, The weight of The increase and decrease of The weight of Finally, the final risk coefficient is obtained by summing the two weighted factors and normalizing them.

[0131] The normalized recognition confidence Characterized by: ;

[0132] The cross-weighted fusion formula is represented as: ;

[0133] in, is the risk coefficient, and its value range is the interval (0,1). is the normalized recognition confidence, with a range of (0,1). It represents the confidence level of the crop object as "real" from the visual perspective; is the physical authenticity index, with a value range of (0,1). It represents the degree of confidence that the crop target is "real" from the physical measurement level;

[0134] As The dynamic weight of When , the weight approaches 0, indicating that the importance of visual level judgment is lower; As The dynamic weight of . When the normalized recognition confidence When , the weight approaches 0, indicating that the importance of physical measurement is lower;

[0135] Denominator is a normalization factor that ensures that the overall risk assessment remains on a reasonable scale even when the weights change; The weight is ,when The closer it is to 1, The smaller the weight, the greater the risk. And vice versa. This cross-dependency enables the model to intelligently determine which information source is more decisive in the current situation, thereby making the most robust risk assessment. The fractional structure ensures its inherent normalization properties, rigorous logic, and stable calculations.

[0136] when The closer the output is to 0, and hour, The closer it is to 0. This means that both visual and physical measurements consider the crop target to be "fake";

[0137] The changing trend of "setting the technical goal as path planning efficiency" is towards maximizing the neglect of non-risk targets. The closer a crop target's risk factor is to zero, the lower the cost on the subsequent cost map. The robot can completely ignore its presence during path planning, resulting in shorter, more efficient paths.

[0138] when and When , the numerator approaches 0 and the denominator is a non-zero constant, so This design ensures that the risk of double negation is completely eliminated, which is completely consistent with the decision-making logic.

[0139] when The closer the output is to 1, the more conservative and safest strategy is adopted. The larger the value, the higher the risk level of the division.

[0140] when and When the numerator approaches 1, the denominator approaches 1, so It not only assesses the risk of the target itself, but also the risk of the system's "perception." When two reliable information sources provide sharply conflicting information, the system does not favor either side, but instead quantifies this "uncertainty" as a higher risk level. This is an extremely advanced and secure fault-tolerance mechanism.

[0141] By introducing a cross-weighted fusion mechanism, the system goes beyond simply linearly superimposing information and instead implements a dynamic, intelligent "arbitration." It dynamically adjusts the importance of different information sources based on their "certainty," resulting in a risk assessment that's closest to the truth. This approach improves the accuracy and robustness of risk assessments.

[0142] Step S5: Map the risk volumes and risk coefficients of all confirmed targets in the workspace onto a rasterized cost map to construct a decision space that comprehensively reflects the distribution of environmental risks; on this cost map, use the path planning algorithm to search and plan an optimal driving path with the minimum total cost, and generate control instructions based on this to drive the robot to move.

[0143] Further explanation: The risk volume of all confirmed targets Projecting to a 2D or 3D rasterized costmap The cost value of each grid is The risk values ​​of all risk volumes falling within the grid are accumulated or the maximum value is taken; then, the cost-based path planning algorithm is used to calculate the path on the cost map. Search for a total path cost from the robot's current position to the target operation point Minimum optimal driving path ; Finally, the system Generate driving instructions and calculate the accumulated risk ahead of the path in real time to dynamically adjust the robot's upper speed limit. Specific implementation content includes:

[0144] This embodiment adopts A Path planning algorithm; the system integrates the "risk volume" and risk coefficient generated by all confirmed targets into a unified environmental cost map The “passing cost” in high-risk areas is high, and the cost in low-risk areas is low. Path planning algorithm, finds a total path cost on this cost map The robot will ultimately follow this calculated optimal path. This cost combines the physical length of the path and the total risk along the way. The system continuously monitors the cost distribution of the area ahead along this path. If it is about to enter a high-cost area, it will smoothly reduce speed in advance and smoothly resume speed after exiting, achieving refined adaptive speed control.

[0145] Further explanation: The dynamic adjustment of the robot's maximum driving speed limit is achieved by Calculate the preset forward distance ahead of the robot's current position in real time The cumulative integral of the path cost value within , thereby generating a path risk prospective index , and based on this index, a nonlinear mapping function is used to determine the current instantaneous maximum allowable speed The specific implementation contents include:

[0146] Path Risk Forward Index The calculation formula is based on an integral model along the optimal path. Starting from the optimal driving path Integrate forward until the preset look-ahead distance , accumulate the grid cost value at each point position s on the path ; To standardize the index, the integral result will be divided by the maximum possible cumulative cost; that is, all grids within the forward distance have the maximum cost value, thus obtaining a standardized index in the interval [0,1] that is independent of the specific path and cost.

[0147] The path risk forward index is characterized as: ;

[0148] The instantaneous maximum permissible speed Characterized by: ;

[0149] in, is the path risk look-ahead index, which ranges from [0,1]. This index quantitatively describes the risk level of the area the robot is about to traverse; The robot follows the optimal path Current location.

[0150] is the forward distance, a preset constant representing how far the robot "looks forward" in meters. In this embodiment, it is set to the length of one robot body. is the cost value of the cost map grid corresponding to position s on the optimal driving path;

[0151] It is the line integral operation along the path. In the discrete grid map, this is equivalent to the line integral operation along the path. Sum the cost values ​​of all passed grids within the distance; Is the maximum possible cost value for a single grid in the cost map. Used for normalization to ensure The value range is between [0,1].

[0152] is the calculated, instantaneous maximum permissible speed of the robot. It is the basic maximum speed of the robot, that is, the maximum speed in an ideal obstacle-free environment. is the speed attenuation factor, a constant greater than 0, used to adjust the sensitivity of speed to the risk index. This will cause the speed to drop slowly when the risk is low and drop sharply when the risk is high; The opposite is true.

[0153] Integral Model Quantifies the "cumulative risk of the path ahead". Dividing by the denominator to normalize ensures the fairness of the index. Instantaneous maximum allowable speed formula It is a risk-based de-weighting model, and its logic is: the basic speed is multiplied by a "safety factor", and this safety factor decreases as the risk ahead increases.

[0154] when The closer the output is to 0: the robot is ahead of the current position On the optimal driving path within the distance, the smaller the cost, that is The closer it is to 0, the more open and safe the area ahead is. The better the road conditions ahead are judged by the system, the less deceleration measures will be implemented.

[0155] when The closer it is to 0, .therefore, The robot is allowed to travel at its designed maximum basic speed. This design ensures that under the premise of safety, the robot can give full play to its movement performance, shorten the non-operational travel time, and thus improve the overall operation efficiency; it should be noted that, " "It means approaching;

[0156] when The closer the output is to 1: the robot is ahead The optimal driving path within the distance is more occupied by high-cost grids, that is, The closer This indicates that the robot is about to enter or is currently traversing a more crowded, complex, or high-risk area.

[0157] Operational safety tends to be maximized. The more severe the road conditions ahead, the more the system determines that the speed strategy to be adopted should focus on safe deceleration.

[0158] when The closer it is to 1, .therefore, The robot is instructed to slow down to a near-stop state. This design ensures that when faced with a foreseeable and extremely high risk of collision, the system can enforce the safest behavior mode, buying valuable reaction time for the perception and decision-making systems to avoid collisions. Speed ​​attenuation factor The adjustment can further optimize the deceleration behavior to make it more in line with actual needs; when a more "aggressive" braking effect is required, by increasing accomplish.

[0159] This embodiment introduces a path risk prospect index The robot's speed control is no longer a passive response, but an active prediction. This represents a quantum leap from "speed control based on current environmental complexity" to "speed control based on future path risk." This refined speed control strategy, tightly coupled with path planning results, makes the robot's motion in complex environments more intelligent. It can advance when it's safe to do so, while also reducing speed to ensure safety before danger strikes. This perfectly balances the core conflict between efficiency and safety in agricultural robot operations.

[0160] Example 2:

[0161] This experiment aims to simulate and verify the performance of an autonomous navigation agricultural robot in a high-density, complex strawberry greenhouse, performing a navigation task from a starting point to a designated work point (a ripe strawberry to be picked). The test environment is a 10m x 5m simulated greenhouse area, in which a variety of simulated obstacles are randomly but densely arranged, including: PVC pipes representing rigid support rods, with inherent risk category of PVC pipes set as , representing flexible but still cautiously avoided strawberry plant leaves, set , low reflectivity patches representing harmless ground water stains or optical artifacts, set The experimental robot platform is equipped with a binocular stereo camera, a computing platform and a wheeled drive chassis. The robot's basic maximum speed Set to 1.0m / s, look-ahead distance Set to 0.5 meters, speed attenuation factor Set to 1.5.

[0162] The experiment was conducted in two groups: a control group and the present invention group;

[0163] The control group uses a relatively advanced adaptive speed control method in the prior art: a speed control strategy based on the complexity of the global environment. This strategy first constructs a cost map that is exactly the same as the invention group. , and plans the optimal path. It then calculates the total cost within a fixed 1.5-meter radius circular area around the robot, using this as an "environmental complexity index." When this index exceeds a preset threshold, the robot's speed is forced to slow down to a fixed, low speed of 0.3 m / s. If it falls below the threshold, the robot returns to its maximum speed of 1.0 m / s. This speed regulation is step-by-step and unpredictable.

[0164] The present invention group fully adopts the "path risk prospective index based on the present invention" "Adaptive speed control strategy. The system plans the optimal path After that, the accumulated cost value on the 0.5-meter path ahead of the robot is calculated in real time to generate a continuously changing Index, and based on this, the formula Continuously and smoothly adjust the instantaneous maximum permissible speed.

[0165] The test scenario was designed with six representative path segments (test objects), each of which contained different combinations and distributions of obstacles, aiming to comprehensively examine the coping capabilities of the two strategies.

[0166] Path segment 1 is an open area: there are few obstacles ahead on the path.

[0167] Path segment 2 is before entering a dense area: the path is about to enter a complex area consisting of multiple PVC pipes and leaves.

[0168] Path segment 3 is through dense areas: the path zigzags between PVC pipes and foliage.

[0169] Path segment 4 is a crossing artifact area: there are multiple low-risk targets ahead of the path that have been identified and confirmed as harmless water stains.

[0170] Path segment 5 is an emergency avoidance: a high-risk PVC pipe that had not been detected before suddenly appeared in the front of the path.

[0171] Path segment 6 is exiting the dense area: the path is about to leave the complex area and enter the open area.

[0172] For each path segment, the following key performance indicators are recorded in detail: the robot's average speed, standard deviation of speed, minimum safe distance (the instantaneous minimum distance between the robot and the nearest high-risk obstacle), and total travel time. By comparing these data, we can objectively and quantitatively evaluate the superiority of our invention over existing technologies.

[0173] Table 2 Optimal driving path and beneficial study of generating control instructions to drive robot movement:

[0174]

[0175] By comparing the data in the above table, we can clearly see the significant beneficial effects brought about by the present invention:

[0176] 1. Higher operational efficiency and intelligent speed decision-making:

[0177] In "Segment 3" of the densely populated area, the average speed of the group using the invention was 0.42 m / s, significantly higher than the control group's 0.3 m / s, reducing travel time by 28%. This is because the invention can fine-tune the speed based on the actual distribution of risks along the route, rather than simply reducing it to the minimum speed.

[0178] When crossing the artifact area "Segment 4," the average speed of the invention group, 0.95 m / s, has little impact because it correctly assesses the low risk of the path ahead. However, the existing system may mistakenly reduce speed due to the increase in the number of detected targets.

[0179] When exiting the densely populated area "Segment 6", the inventive group was able to predictably and smoothly accelerate in advance to an average speed of 0.68 m / s, while the control group had to completely leave the complex area before resuming high speed, resulting in the travel time being more than doubled.

[0180] 2. Higher safety and smoothness:

[0181] In two key scenarios, "Chapter 2" before entering a crowded area and "Chapter 5" for emergency avoidance, the minimum safe distance of the invention group (35cm, 38cm) was much greater than that of the control group (18cm, 15cm). This is due to the predictive deceleration, which allows the robot to reach a lower, safer speed when approaching danger, giving it more room to react.

[0182] It can be seen from the speed standard deviation data that the speed changes of the present invention group are smooth, with a standard deviation of less than 0.15, while the control group has a sharp step at the speed switching point, with a standard deviation as high as 0.45. This sudden acceleration and deceleration not only has an impact on the robot hardware, but also increases the uncertainty of the movement.

[0183] The data from this example strongly demonstrates that, compared to existing speed control strategies based on global environmental complexity, the adaptive speed control method based on path risk foresight proposed in this paper achieves a better balance between robot operating efficiency and safety in complex environments. Through a refined and predictive mechanism, it makes the robot's motion more fluid, intelligent, and reliable, demonstrating significant practical value and technological advancement.

[0184] Example 3:

[0185] See also Figure 2 and Figure 3 , a smart agricultural robot pattern recognition and obstacle avoidance system, the system is used to execute the smart agricultural robot pattern recognition method, comprising:

[0186] Data acquisition module: used to acquire at least two images of the target scene in different polarization directions, and calculate and generate a corrected image that suppresses specular reflection highlight artifacts based on the polarization characteristics of the images;

[0187] A region determination module is configured to input the corrected image into a pre-trained neural network model to identify crop targets and output the category of each crop target and its corresponding recognition confidence level. When the recognition confidence level of a confirmed target is lower than a preset confidence threshold, the region where the crop target is located is determined to be a highly suspicious region.

[0188] Confirmation module: for performing a physical property review operation on the highly suspicious area, which includes: collecting the temperature change curve of the area through micro-area thermal excitation and high-frequency temperature measurement; fitting the temperature change curve with an exponential decay model to solve the thermal relaxation time constant that characterizes its thermodynamic properties, and finally confirming the true physical properties of the highly suspicious area based on this;

[0189] Risk coefficient calculation module: This module is used to calculate the three-dimensional spatial coordinates of the robot in its own coordinate system based on the principle of binocular stereo vision. It also integrates the confirmed physical properties, recognition confidence, and thermal relaxation time constant information of the crop target to dynamically construct a risk volume representing the collision risk and its corresponding risk coefficient for the confirmed target.

[0190] Result output module: used to map the risk volume and risk coefficient of all confirmed targets in the workspace onto a rasterized cost map to construct a decision space that comprehensively reflects the distribution of environmental risks; on this cost map, a path planning algorithm is used to search and plan an optimal driving path with the minimum total cost, and based on this, control instructions are generated to drive the robot to move.

[0191] It should be noted that: All calculation formulas in this application document use regression analysis including but not limited to machine learning algorithms to deeply analyze the relevant parameters collected and identify their natural trends and relationships. Use professional software, such as Python's Scikit-learn library or R language, to automatically generate mathematical models that match the data. Then, objectively evaluate the performance of the model through methods such as cross-validation, and combine continuous feedback and optimization to ensure that the created formula truly reflects the inherent laws of the data, thereby ensuring its effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula are dimensionally non-dimensionalized within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless technical means include but are not limited to Min-Max-Normalization and Z-Score standardization;

[0192] The technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random-access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0193] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0194] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A pattern recognition method for an intelligent agricultural robot, applied to an obstacle avoidance system, characterized in that: The specific steps include: Step S1: capturing at least two images of a target scene in different polarization directions, and calculating and generating a corrected image that suppresses specular reflection highlight artifacts based on the polarization characteristics of the images; Step S2: Inputting the corrected image into a pre-trained neural network model to identify crop targets and output the category of each crop target and its corresponding recognition confidence; when the recognition confidence of a confirmed target is lower than a preset confidence threshold, the area where the crop target is located is determined to be a highly suspicious area; Step S3: Performing a physical property review operation on the highly suspicious area, which includes: collecting a temperature change curve of the area through micro-area thermal excitation and high-frequency temperature measurement; fitting the temperature change curve with an exponential decay model to solve the thermal relaxation time constant that characterizes its thermodynamic properties, and finally confirming the true physical properties of the highly suspicious area based on this; Step S4: Based on the principle of binocular stereo vision, the three-dimensional spatial coordinates of the robot of the confirmed target in its own coordinate system are calculated; and the confirmed physical properties, recognition confidence, and thermal relaxation time constant information of the crop target are integrated to dynamically construct a risk volume representing the collision risk and its corresponding risk coefficient for the confirmed target; Step S5: Map the risk volumes and risk coefficients of all confirmed targets in the workspace onto a rasterized cost map to construct a decision space that comprehensively reflects the distribution of environmental risks; on this cost map, use the path planning algorithm to search and plan an optimal driving path with the minimum total cost, and generate control instructions based on this to drive the robot to move.

2. The intelligent agricultural robot pattern recognition method according to claim 1, characterized in that: Using a polarization camera to acquire images of four polarization channels of 0°, 45°, 90°, and 135° for the same scene, and calculating the linear polarization degree of each pixel based on the images to generate an optical artifact mask corresponding to the specular reflection area; and performing pixel-level weighted correction on the originally acquired RGB image using the optical artifact mask to obtain a corrected image with suppressed specular highlights; The pixel-level weighted correction of the original collected RGB image is performed, and its correction weight is an adaptive correction weight dynamically calculated according to the relationship between the linear polarization degree of each pixel and the polarization degree threshold, so as to achieve differentiated suppression of mirror reflections of different intensities.

3. The intelligent agricultural robot pattern recognition method according to claim 2, characterized in that: The confidence threshold used to determine highly suspicious areas is a dynamic confidence threshold for each category that is queried or calculated from a preset category-threshold mapping table based on the specific category of each identified target output by the neural network model.

4. The intelligent agricultural robot pattern recognition method according to claim 3, characterized in that: triggering a review operation for the highly suspicious area, the review operation comprising controlling a microfocus laser to emit a laser pulse of preset energy toward the highly suspicious area, and simultaneously using a microbolometer to monitor a temperature change curve of the area after the laser pulse is applied, wherein the temperature change curve is fitted with an exponential decay model to extract a thermal relaxation time constant representing a heat dissipation rate, and finally confirming the true physical properties of the highly suspicious area based on the thermal relaxation time constant; Based on the thermal relaxation time constant and pre-calibrated reference time constants of water and healthy plant tissue, a continuously varying physical authenticity index is calculated through a nonlinear S-type logistic function to quantify the degree of confidence that the area is real plant tissue; The larger the value of the physical authenticity index, the higher the probability that the area contains plant tissue with normal physiological activity.

5. The intelligent agricultural robot pattern recognition method according to claim 4, characterized in that: Using at least two cameras with a preset baseline distance in space, synchronously acquire a pair of images of the crop target. Using a stereo matching algorithm, the parallax of the target with confirmed attributes in the left and right images is calculated. Combining the camera focal length and baseline distance, the three-dimensional spatial coordinates of the target in the robot's own coordinate system are then calculated. Then, a risk volume and its corresponding risk coefficient are generated for the crop target through a preset fusion model, combining the crop target's category attributes, the recognition confidence based on the neural network, and the physical reality index. The risk coefficient is calculated through a cross-weighted fusion mechanism; In this mechanism, the function used to calculate the recognition confidence weight has the physical authenticity index as its independent variable, while the function used to calculate the physical authenticity index weight has the recognition confidence as its independent variable, thereby realizing dynamic and mutual adjustment between the weights of the two information sources.

6. The intelligent agricultural robot pattern recognition method according to claim 5, characterized in that: Project the risk volumes of all confirmed targets onto a two-dimensional or three-dimensional gridded cost map. The cost value of each grid is determined by summing or maximizing the risk values ​​of all risk volumes within the grid. Then, a cost-based path planning algorithm is used to search for an optimal driving path from the current position of the robot to the target operation point with the minimum total path cost on the cost map; Ultimately, the system generates driving instructions based on the optimal driving path and also calculates the risk accumulation ahead of the path in real time to dynamically adjust the upper limit of the robot's driving speed.

7. The intelligent agricultural robot pattern recognition method according to claim 6, characterized in that: The dynamic adjustment of the robot's maximum driving speed limit is achieved by calculating in real time the cumulative integral of the path cost value within a preset forward distance ahead of the robot's current position on the planned optimal driving path, thereby generating a path risk forward index, and based on the index, determining the current instantaneous maximum allowable speed through a nonlinear mapping function.

8. An intelligent agricultural robot pattern recognition and obstacle avoidance system, characterized by: The system is used to execute the intelligent agricultural robot pattern recognition method according to any one of claims 1 to 7, comprising: Data acquisition module: used to acquire at least two images of the target scene in different polarization directions, and calculate and generate a corrected image that suppresses specular reflection highlight artifacts based on the polarization characteristics of the images; A region determination module is configured to input the corrected image into a pre-trained neural network model to identify crop targets and output the category of each crop target and its corresponding recognition confidence level. When the recognition confidence level of a confirmed target is lower than a preset confidence threshold, the region where the crop target is located is determined to be a highly suspicious region. Confirmation module: for performing a physical property review operation on the highly suspicious area, which includes: collecting the temperature change curve of the area through micro-area thermal excitation and high-frequency temperature measurement; fitting the temperature change curve with an exponential decay model to solve the thermal relaxation time constant that characterizes its thermodynamic properties, and finally confirming the true physical properties of the highly suspicious area based on this; Risk coefficient calculation module: This module is used to calculate the three-dimensional spatial coordinates of the robot in its own coordinate system based on the principle of binocular stereo vision. It also integrates the confirmed physical properties, recognition confidence, and thermal relaxation time constant information of the crop target to dynamically construct a risk volume representing the collision risk and its corresponding risk coefficient for the confirmed target. Result output module: used to map the risk volume and risk coefficient of all confirmed targets in the workspace onto a rasterized cost map to construct a decision space that comprehensively reflects the distribution of environmental risks; on this cost map, a path planning algorithm is used to search and plan an optimal driving path with the minimum total cost, and based on this, control instructions are generated to drive the robot to move.

Images