A control method and system of a spherical robot based on artificial intelligence

Abstract

This invention discloses a control method and system for a spherical robot based on artificial intelligence, relating to the field of spherical robot control technology. It includes: receiving multi-dimensional data collected by the spherical robot, performing preprocessing and parsing to obtain an initial dataset; and using a target recognition algorithm to perform target recognition on the data in the initial dataset. Its key technical points are: this invention combines a hierarchical closed-loop control architecture with adaptive optimization technology, integrating the entire process modules of perception, decision-making, execution, and feedback to achieve fully autonomous concealment and stable tracking after deployment. This solves the problems of insufficient intelligence and the need for manual intervention in existing spherical robot control systems, achieving unmanned operation throughout the entire process from target locking and concealment control to stable tracking after deployment. The overall tracking accuracy is high, significantly improving the autonomous combat capability of the spherical robot, expanding its application scope in complex reconnaissance scenarios, and reducing manual operation costs and exposure risks.

Description

Technical Field

[0001] This invention relates to the field of spherical robot control technology, specifically to a control method and system for a spherical robot based on artificial intelligence. Background Technology

[0002] As a new type of reconnaissance equipment with a spherical shell as its core structure and silent rolling drive, the spherical robot has been widely used in complex scenarios such as counter-terrorism, emergency rescue, and battlefield reconnaissance due to its advantages such as compact structure, strong impact resistance, and flexible movement. The spherical shell of the spherical robot is made of high-strength lightweight composite material, which can withstand the impact of being thrown from a height of 5-10 meters without damage. The spherical robot has built-in vision sensors, environmental sensors, and inertial measurement units, and can quickly recover its posture after being thrown and landing. It can adapt to complex terrains such as uneven surfaces and narrow and confined spaces. The spherical robot can achieve 360° movement without blind spots, and its silent operation provides a certain degree of concealment.

[0003] The core functions of a spherical robot are highly dependent on its supporting control system. After being launched, it can quickly adapt to the environment, lock onto the target, and covertly track it. This is the key to achieving efficient reconnaissance and is also one of the core directions for the current development of reconnaissance equipment towards miniaturization, intelligence, and concealment.

[0004] The existing control systems for spherical robots have fundamental flaws, specifically: The core deficiency of current spherical robot control systems lies in the lack of fully autonomous operation capabilities. After being launched, the spherical robot requires human assistance to complete reconnaissance missions and cannot achieve independent covert tracking. Given that the launching scenarios are often complex and unknown environments with variable lighting, dense obstacles, and interference from non-target personnel, all core operations require remote human intervention after the spherical robot lands. Operators must manually perform operations such as attitude calibration, target locking, and selection of concealed positions based on real-time transmitted images, while also manually adjusting vision equipment parameters and movement paths. Manual control suffers from long response delays, making it difficult to quickly respond to sudden situations such as target movement and attitude deviations. Furthermore, it is difficult to accurately balance concealment and shooting effects, and operational errors can easily lead to the spherical robot being exposed and the target being lost, severely impacting reconnaissance efficiency.

[0005] In summary, existing control systems for spherical robots do not meet market demands. Therefore, we propose an artificial intelligence-based control method and system for spherical robots. Summary of the Invention

[0006] To achieve the above objectives, the present invention provides the following technical solution: A control method and system for a spherical robot based on artificial intelligence, comprising: Receive multi-dimensional data collected by a spherical robot, and perform preprocessing and parsing to obtain an initial dataset; A target recognition algorithm is used to perform target recognition on the data in the initial dataset. After the target is recognized, target locking and parameter analysis are performed to obtain the analysis dataset. After the target is locked, a quantitative evaluation algorithm is used to perform a concealment assessment and a shooting feasibility assessment of the current location from multiple dimensions based on the initial dataset and the analysis dataset. The concealment assessment results and shooting assessment results are obtained. If the concealment assessment result is unqualified, the location is adjusted; otherwise, the shooting assessment result is judged to be qualified. If the shooting assessment result is qualified, no action is taken. If the shooting evaluation results are unsatisfactory, an environmental analysis is performed on the initial dataset. Based on the environmental analysis results, the operation of the vision equipment is adjusted to obtain an adjustment and lock report. The adjustment and lock results are evaluated. If the adjustment and lock results are satisfactory, no action is taken; otherwise, position adjustment is performed. Before performing position adjustment, a deep reinforcement learning model is used, combined with multi-dimensional evaluation indicators, to select and evaluate hidden points, screen hidden points, plan candidate paths based on the screened hidden points, evaluate the optimal hidden path and convert it into control commands to drive the spherical robot to perform position adjustment.

[0007] Preferably, the multi-dimensional data collected by the spherical robot includes: environmental images and potential target area images collected by the vision device on the spherical robot; and light intensity, environmental noise, ground slope, and obstacle height collected by several sensors on the spherical robot.

[0008] Preferably, the preprocessing and parsing process includes: performing noise reduction on environmental images and potential target area images; analyzing ground slope and obstacle height; generating a raster map of passable areas and marking potential concealed locations, open areas, and areas easily covered by personnel's line of sight; further combining the noise-reduced image data, illumination intensity, and environmental noise to form a preprocessing set; performing hierarchical processing on the preprocessing set using a classification parsing algorithm to obtain an environmental parameter report and an image analysis report; and summarizing the preprocessing set, environmental parameter report, and image analysis report to form an initial dataset.

[0009] Preferably, the target recognition algorithm is used to perform target recognition on the data in the initial dataset. After the target is identified, target locking and parameter analysis are performed, including: performing real-time frame parsing on the image data in the initial dataset using the target recognition algorithm to extract human contour features and motion trajectory features; calculating the matching score of human contour features using a threshold judgment method; comparing the matching score with a preset score threshold; if the matching score is lower than the preset score threshold, it is determined to be a non-target and non-target marking is performed; if the matching score is not lower than the preset score threshold, it is determined to be a target and target marking locking is performed; establishing a coordinate system with the current position of the spherical robot as the origin, and locking the real-time coordinates of the target based on the coordinate positioning algorithm; analyzing the motion trajectory features of non-targets and target personnel, marking the movement speed, target orientation, line-of-sight coverage area and movement direction, and summarizing them to form an analysis dataset.

[0010] Preferably, the quantitative evaluation algorithm performs a concealment assessment of the current position from multiple dimensions based on the initial dataset and the analysis dataset, including: in the target line-of-sight avoidance dimension, calculating the angle between the target orientation and the current position, and performing a target line-of-sight avoidance score based on the angle; in the non-target visibility dimension, calculating the average overlap of all non-target line-of-sight coverage areas with the current position, and performing a non-target visibility score based on the average overlap; in the occlusion recognition dimension, calculating a contrast score based on the current position's environmental color and the spherical robot's color, and calculating an occlusion volume ratio score based on the occlusion volume within the current position's range; processing the contrast score and occlusion volume ratio score using a weighted summation method to obtain an occlusion recognition score; performing a weighted summation on the target line-of-sight avoidance score, the non-target visibility score, and the occlusion recognition score to obtain a concealment score; and comparing the concealment score with a preset evaluation interval to obtain the concealment assessment result.

[0011] Preferably, a quantitative evaluation algorithm is used to perform a shooting feasibility assessment of the current location based on the initial dataset and the analysis dataset from multiple dimensions, including: in the shooting range dimension, assessing whether the target is within the shooting range based on the field of view of the vision device, the straight-line distance between the spherical robot and the target, and the real-time coordinates of the target; in the image clarity dimension, performing a comprehensive assessment based on the light intensity, the parameters of the vision device, and the image analysis report; in the target's proportion in the image dimension, extracting the ratio of the target's pixel area in the image to the total pixel area of ​​the image based on the image segmentation algorithm, and performing an assessment; in the target occlusion dimension, extracting the ratio of the pixel area of ​​the occluded part of the target to the pixel area of ​​the target in the image based on the target recognition algorithm, and performing an assessment; summarizing the evaluation results of the four dimensions and performing a judgment to obtain the shooting evaluation result.

[0012] Preferably, environmental analysis is performed on the initial dataset, and the operation control of the vision device is performed based on the environmental analysis results, including: extracting the light intensity and judging the lighting environment; adjusting the corresponding operation mode of the vision device based on the lighting environment; calculating the field of view adjustment value of the vision device based on the distance between the current position and the target; calculating the horizontal and vertical offsets based on the real-time coordinates of the target and the current orientation of the vision device, and calculating the rotation speed and rotation direction of the gimbal based on the offsets; converting the calculated field of view adjustment value of the vision device, the rotation speed and rotation direction of the gimbal into adjustment commands, and sending them to the spherical robot to perform the field of view control operation and gimbal adjustment; after adjustment, detecting whether the vision device is pointing towards a reflective surface or the direction of the person's line of sight; if the vision device is pointing towards a reflective surface or the direction of the person's line of sight, then performing a fine adjustment of the vision device orientation and re-detecting; otherwise, no adjustment is made.

[0013] Preferably, the selection and evaluation of concealed locations includes: classifying and organizing the data in the initial dataset and the analysis dataset, inputting them into a deep reinforcement learning model to perform format standardization and outlier removal, obtaining raster map data and parameter vectors; performing normalization processing on the raster map data and parameter vectors to obtain normalized first raster map spatial features and first parameter feature vectors; extracting spatial correlation features of obstacle distribution and shadow coverage from the first raster map spatial features, performing compression processing, and fusing them with the first parameter feature vector to obtain a first fused feature vector; combining a pre-set decoupled reward function, performing scoring calculations on each concealed location from four dimensions: concealment, shooting effect, distance, and environmental adaptation, to obtain a comprehensive score for the concealed location; comparing the comprehensive score of the concealed location with a preset screening threshold, and removing concealed locations with comprehensive scores lower than the preset screening threshold to obtain a set of concealed locations.

[0014] Preferably, the process of planning candidate paths based on the selected concealed locations, evaluating the optimal concealed path, and converting it into control instructions includes: classifying and organizing the data from the initial dataset, analysis dataset, and concealed location set, inputting them into a deep reinforcement learning model to perform format standardization and outlier removal, obtaining normalized second grid map spatial features and second parameter feature vectors; extracting the spatial features of obstacle distribution, shadow area continuity, and spatial range correlation of non-target line-of-sight coverage from the second grid map spatial features, performing compression processing, and fusing them with the second parameter feature vector to obtain a second fused feature vector; planning candidate paths, converting the second fused feature vector of the candidate paths into Q-values, comparing the Q-values, selecting the three candidate paths with the highest Q-values ​​for quantization verification, optimizing the nodes of the candidate paths, recalculating the Q-values, comparing the recalculated Q-values, and selecting the path with the highest Q-value as the optimal concealed path.

[0015] A control system for an artificial intelligence-based spherical robot includes a data acquisition module, a target recognition module, a position verification module, a vision adjustment module, and a position adjustment module. Data acquisition module: Receives multi-dimensional data collected by the spherical robot, performs preprocessing and parsing, and obtains the initial dataset; Target recognition module: Uses target recognition algorithms to perform target recognition on the data in the initial dataset. After the target is recognized, it performs target locking and parameter analysis to obtain the analysis dataset. Location verification module: After the target is locked, a quantitative evaluation algorithm is used to perform a concealment assessment and shooting feasibility assessment of the current location from multiple dimensions based on the initial dataset and the analysis dataset. The concealment assessment result and shooting assessment result are obtained. If the concealment assessment result is unqualified, the location is adjusted; otherwise, the shooting assessment result is judged to be qualified. If the shooting assessment result is qualified, no action is taken. Visual adjustment module: If the shooting evaluation result is unqualified, environmental analysis is performed on the initial dataset, and the operation of the vision device is adjusted based on the environmental analysis result to obtain an adjustment and lock report. The adjustment and lock result is evaluated. If the adjustment and lock result is qualified, no action is taken; otherwise, position adjustment is performed. Position Adjustment Module: Before performing position adjustment, a deep reinforcement learning model is used, combined with multi-dimensional evaluation indicators, to select and evaluate hidden points, screen hidden points, plan candidate paths based on the screened hidden points, evaluate the optimal hidden path and convert it into control commands to drive the spherical robot to perform position adjustment.

[0016] This invention provides a control method and system for a spherical robot based on artificial intelligence, which has the following beneficial effects: This invention combines target recognition algorithms with multi-dimensional data acquisition technology to automatically identify and lock onto targets. It effectively solves the problems of existing spherical robot control systems that require manual target locking after throwing and have delayed response. It achieves automated and precise target locking, ensures rapid tracking after throwing, prevents target loss due to manual operation, creates conditions for subsequent covert tracking, and improves the efficiency and timeliness of reconnaissance.

[0017] This invention combines segmented control technology with concealment judgment technology. It adjusts different parameters of the vision device according to different lighting environments and monitors the orientation of the vision device lens in real time to prevent it from pointing towards reflective surfaces or densely populated areas. This solves the problem that existing spherical robot control systems require manual adjustment of vision device parameters and are prone to exposure or blurry images due to improper parameter settings. It enables the vision device parameters to adapt autonomously to environmental changes, ensuring the quality of evidence collection, avoiding exposure risks caused by reflections or excessive supplementary lighting of the vision device, and enhancing the concealment of the spherical robot.

[0018] This invention combines reinforcement learning and quantitative evaluation techniques to achieve autonomous assessment of concealed locations and optimal concealed path planning. It eliminates the need for manual designation of concealed locations, solving the shortcomings of existing spherical robot control systems, which require manual designation of concealed locations and suffer from unreasonable and easily exposed path planning. It enables autonomous decision-making on concealed locations and paths, ensuring that the spherical robot can quickly move to a suitable concealed area while maintaining the optimal tracking distance to the target, thus improving the safety and efficiency of concealed tracking.

[0019] This invention combines a hierarchical closed-loop control architecture with adaptive optimization technology, integrating the entire process modules of perception, decision-making, execution, and feedback. It achieves fully autonomous concealment and stable tracking after deployment, solving the problems of insufficient intelligence and the need for manual intervention in existing spherical robot control systems. It enables unmanned operation of the entire process from target locking and concealment control to stable tracking after deployment, with high overall tracking accuracy. This significantly improves the autonomous combat capability of spherical robots, expands the application scope in complex reconnaissance scenarios, reduces manual operation costs and exposure risks, and has good application prospects. Attached Figure Description

[0020] Figure 1 This is a flowchart of a control method for a spherical robot based on artificial intelligence according to the present invention; Figure 2 This is a framework diagram of a control system for a spherical robot based on artificial intelligence according to the present invention; Figure 3 This diagram illustrates the analysis of concealed locations in the control system of an artificial intelligence-based spherical robot according to the present invention. Detailed Implementation

[0021] 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.

[0022] Example 1: Please see Figure 1 This embodiment provides a control method for a spherical robot based on artificial intelligence, including: S1. Receive multi-dimensional data collected by the spherical robot, and perform preprocessing and parsing to obtain the initial dataset; The multi-dimensional data collected by the spherical robot includes: environmental images and potential target area images collected by the vision device on the spherical robot; and light intensity, environmental noise, ground slope and obstacle height collected by several sensors on the spherical robot.

[0023] S1 is activated after the spherical robot is thrown and stabilized. When collecting data, the default 940nm no-red-gloom night vision mode is enabled. This no-red-gloom night vision mode has no visible light leakage and can avoid exposure. The initial parameters of the vision device are set as follows: field of view is 60°, shooting resolution is 1080P, and frame rate is 25fps. The vision device collects environmental images and potential target area data in real time. Each frame of image data is denoised and grayscaled.

[0024] Image denoising and grayscale conversion are common visual inspection techniques, and those skilled in the field are familiar with how they are implemented, so they will not be described in detail.

[0025] The preprocessing and parsing process includes: denoising environmental images and potential target area images; analyzing ground slope and obstacle height; generating a raster map of passable areas; and marking potential concealed locations, open areas, and areas easily covered by human sight. Further, the denoised image data, illumination intensity, and environmental noise are combined to form a preprocessing set. A classification parsing algorithm is used to perform hierarchical processing on the preprocessing set to obtain an environmental parameter report and an image analysis report. The preprocessing set, environmental parameter report, and image analysis report are then summarized to form the initial dataset.

[0026] The spherical robot has several sensors, including an ambient light sensor, an acoustic sensor, and a ground flatness sensor. An ambient light sensor detects light intensity in real time, with a data acquisition range of 0-100,000 Lux, outputting data every 0.1 seconds for subsequent adaptive adjustment of visual device parameters. An acoustic sensor collects ambient noise, with a detection range of 30-80 dB, and uses feature extraction algorithms to filter characteristic signals such as footsteps and conversations, which are then simultaneously uploaded to the data processing unit. A ground flatness sensor works in conjunction with the inertial measurement unit on the spherical robot to detect ground slope and obstacle height, generating a grid map of passable areas, marking potential hiding places such as shadow areas, corners, and the bottom of furniture, and marking open areas and areas easily covered by human vision.

[0027] The environmental parameter report covers illumination level, interference level, coordinates of concealed locations, and obstacle distribution; the image analysis report includes information on potential target areas and background interference.

[0028] S2. Use a target recognition algorithm to perform target recognition on the data in the initial dataset. After the target is recognized, perform target locking and parameter analysis to obtain the analysis dataset. The target recognition algorithm is used to perform target recognition on the data in the initial dataset. After target recognition, target locking and parameter analysis are performed, including: real-time frame parsing of the image data in the initial dataset using the target recognition algorithm to extract human contour features and motion trajectory features; the target recognition algorithm is the target recognition algorithm mounted on an existing spherical robot, and a threshold judgment method is used to calculate the matching score of human contour features. The threshold judgment method can exclude furniture, clutter, and other items, reducing the probability of recognition errors; when calculating the matching score of human contour features, the matching degree of human features is calculated, and then the matching degrees of all human features are weighted and summed to obtain the matching score. Human features include contour, head, torso, limbs, movement, gait, skin color, and clothing features; the matching score is compared with a preset score threshold. If the matching score is lower than the preset score threshold, it is determined to be a non-target and a non-target label is executed; if the matching score is not lower than the preset score threshold, it is determined to be a target and target label locking is executed.

[0029] Both the target and non-target are people. The preset score threshold is usually set to 0.88. The higher the preset score threshold, the more accurate the result. However, the target will move and be subject to various interferences, which will reduce the matching score. Therefore, it is set to 0.88.

[0030] When the target is locked, a coordinate system is established with the current position of the spherical robot as the origin, and the real-time coordinates of the target are locked based on the coordinate positioning algorithm. Specifically, a two-dimensional coordinate system is constructed with the current position of the spherical robot as the origin. The movement trajectory characteristics of non-target and target personnel are analyzed, and the movement speed, target orientation, line-of-sight coverage area and movement direction are marked. The data are then summarized to form an analysis dataset.

[0031] The movement speed can be calculated by dividing the distance between the target and the spherical robot by the time at different time points. This reflects the relative movement speed compared to the spherical robot, not the actual movement speed.

[0032] In practical applications, when the target is briefly occluded (the number of occluded frames is less than or equal to 2 frames), a linear prediction algorithm is used to predict the target's next movement position based on the target's historical motion trajectory in order to maintain the locked state. When the target is lost, an instruction is sent to the vision device to expand the shooting range and re-search for the target. The target is lost when no target feature is detected for 3 consecutive frames.

[0033] For non-target personnel, calculate the line-of-sight coverage area for subsequent analysis. Specifically, the line-of-sight coverage angle is a = arctan(H / d) × 2, where H is the height of the non-target personnel and d is the straight-line distance between the non-target personnel and the spherical robot.

[0034] This invention combines target recognition algorithms with multi-dimensional data acquisition technology to automatically identify and lock onto targets. It effectively solves the problems of existing spherical robot control systems that require manual target locking after throwing and have delayed response. It achieves automated and precise target locking, ensures rapid tracking after throwing, prevents target loss due to manual operation, creates conditions for subsequent covert tracking, and improves the efficiency and timeliness of reconnaissance.

[0035] S3. After the target is locked, a quantitative evaluation algorithm is used to perform a concealment assessment and shooting feasibility assessment of the current position from multiple dimensions based on the initial dataset and the analysis dataset. The concealment assessment results and shooting assessment results are obtained. If the concealment assessment result is unqualified, the position is adjusted; otherwise, the shooting assessment result is judged to be qualified. If the shooting assessment result is qualified, no action is taken. The quantitative evaluation algorithm is used to perform a concealment assessment of the current position from multiple dimensions based on the initial dataset and the analysis dataset. This includes: in the target line-of-sight avoidance dimension, calculating the angle between the target orientation and the current position, and performing a target line-of-sight avoidance score based on the angle; when the angle is greater than or equal to 90°, it is judged as complete avoidance, and the target line-of-sight avoidance score is 100 points; if this condition is not met, the score is 0 points.

[0036] In the non-target visibility dimension, the average overlap between the coverage area of ​​all non-target lines of sight and the current position is calculated, and a non-target visibility score is performed based on the average overlap. The specific calculation formula is Scp = 100 - 1000 × arctan(h / d) × 2 / 360, where Scp is the non-target visibility score, 100 is the full score, h is the height of the non-target line of sight, d is the straight-line distance between the non-target and the current position, 2 is the conversion factor for obtaining the complete line of sight coverage angle by switching from one side to both sides, 360 is the total circumferential angle, and 1000 is a fixed scoring conversion factor set to unify the scoring scale and strengthen the weight of overlap. In the occlusion recognition dimension, the occlusion is determined based on the current position's environmental color and the spherical robot's colorimeter. The contrast score is calculated using the formula: Sy = 100 - 100 × |Cjd - Chd|, where Sy is the contrast score, Cjd is the grayscale value of the spherical robot shell color (acquired in real-time by a vision device and then converted to grayscale), and Chd is the grayscale value of the current location area. The volume ratio score of occluded objects within the current location area is calculated using the formula: Sz = 100 × Vz / Vy, where Sz is the total volume of all occluded objects, and Vy is the volume of the preset current location area. The volume of the occluded objects is roughly estimated by multiplying the average cross-sectional area and average height of the occluded objects within the current location area.

[0037] The contrast score and the volume ratio of the occluder are processed by a weighted summation method to obtain the occlusion identification score. The target line-of-sight avoidance score, the non-target visibility score and the occlusion identification score are weighted and summed to obtain the concealment score. The concealment score is compared with the preset evaluation interval to obtain the concealment evaluation result.

[0038] The weights for target line-of-sight avoidance score, non-target visibility score, and occlusion recognition score were 0.4, 0.3, and 0.3, respectively. The preset evaluation range is 0-80 for unqualified and 80-100 for qualified. 0-80 does not include 80, and 80-100 includes 80. For example, if the concealment score is 80, the concealment evaluation result is qualified.

[0039] A quantitative evaluation algorithm is employed, based on the initial dataset and the analysis dataset, to perform a feasibility assessment of shooting at the current location from multiple dimensions. This includes: in terms of shooting range, assessing whether the target is within the shooting range based on the field of view of the visual device, the straight-line distance between the spherical robot and the target, and the target's real-time coordinates; specifically, a two-dimensional coordinate system is constructed with the current location as the origin, comprehensively considering the current field of view of the visual device and its orientation angle to calculate the coordinate range of the visual device's field of view. Simultaneously, the target's real-time coordinates are extracted, and through coordinate intersection operations, it is determined whether the target coordinates are within the shooting field of view. If the target coordinates are completely within the field of view, it is considered a fit; if the target partially exceeds the field of view, and the area exceeding the field of view is less than 10%, it is considered a basic fit; if the area exceeding the field of view is not less than 10%, it is considered an unfit.

[0040] In terms of image sharpness, a comprehensive evaluation is performed based on light intensity, visual device parameters, and image analysis reports. Extract the image's resolution, noise density, and contrast ratio. Resolution must be ≥1080P; images below this standard are considered blurry. Noise density must be ≤5 noise particles per 100 pixels (calculated as the total number of noise particles divided by the total number of pixels); images exceeding this standard are considered blurry. In low-light (0.001-10 Lux) and no-light (less than 0.001 Lux) environments, noise density can be relaxed to 8 noise particles per 100 pixels, and contrast ratio to 20:1 to avoid misjudgments due to environmental limitations. Simultaneously, combine image... The preprocessing noise reduction algorithm corrects the noise density assessment results, improving the accuracy of the assessment. The image contrast ratio is greater than or equal to 30:1 (which can be relaxed to 20:1 in strong light (greater than 10000 Lux) or low light environments; the contrast ratio is calculated by dividing the gray value of the brightest area of ​​the image by the gray value of the darkest area). If it is lower than this standard, the image is considered blurry. When all three sub-indicators meet the requirements, the image is considered clear. When two are met and one is not, the image is considered basically clear. When only one is met or all three are not met, the image is considered blurry.

[0041] The target-to-image proportion dimension is evaluated by extracting the ratio of the target's pixel area to the total pixel area of ​​the image based on image segmentation algorithms. A ratio of 30%-50% is considered acceptable, 20%-30% and 50%-60% are considered basically acceptable, and other ratios are considered unacceptable. 30% and 50% are acceptable, and 20% and 60% are considered basically acceptable.

[0042] In the target occlusion dimension, the ratio of the pixel area of ​​the occluded part of the target to the pixel area of ​​the target in the image is extracted based on the target recognition algorithm. Specifically, the occluded and unoccluded areas of the target are segmented by using target contour extraction and obstacle recognition algorithms, and the total number of pixels in the occluded area is counted. When the ratio is between 0-10%, it is considered qualified; when the ratio is between 10%-20%, it is considered basically qualified; and the rest are considered unqualified. 10% is qualified, and 20% is considered basically qualified.

[0043] In practical applications, if the occluded area is a non-core part of the target, such as the edge of the limbs, and the occlusion rate is less than or equal to 15%, then the occlusion can be judged to meet the requirements, thus enhancing the flexibility of the assessment.

[0044] The evaluation results from the four dimensions are summarized and judged to obtain the shooting evaluation result. If all four items are qualified, the shooting evaluation result is qualified. If no more than one of the four items is basically qualified and the rest are qualified, there is no need to adjust the position of the spherical robot. Only the parameters of the vision device need to be fine-tuned to make the shooting meet the qualified standard. If two or more of the four indicators are basically qualified, or if any one of them is unqualified, the vision device adjustment process needs to be triggered to optimize the shooting effect by adjusting the parameters and orientation of the vision device. If the shooting still cannot meet the qualified standard after adjustment, the position adjustment process is triggered to move it to a better shooting position.

[0045] This invention combines segmented control technology with concealment judgment technology. It adjusts different parameters of the vision device according to different lighting environments and monitors the orientation of the vision device lens in real time to prevent it from pointing towards reflective surfaces or densely populated areas. This solves the problem that existing spherical robot control systems require manual adjustment of vision device parameters and are prone to exposure or blurry images due to improper parameter settings. It enables the vision device parameters to adapt autonomously to environmental changes, ensuring the quality of evidence collection, avoiding exposure risks caused by reflections or excessive supplementary lighting of the vision device, and enhancing the concealment of the spherical robot.

[0046] S4. If the shooting evaluation result is unqualified, perform environmental analysis on the initial dataset, perform visual equipment operation control based on the environmental analysis result, obtain control lock report, evaluate the control lock result, if the control lock result is qualified, no action is taken; otherwise, perform position adjustment. Environmental analysis was performed on the initial dataset. Based on the analysis results, the operation of the vision device was adjusted, including: extracting light intensity and determining the lighting environment; adjusting the corresponding operating mode of the vision device according to the lighting environment; in a strong light environment, i.e., when the light intensity is greater than 10000 Lux: the spherical robot's control system issued a command to activate the HDR wide dynamic range mode, reduce the exposure time to 0.1s, adjust the white balance parameter to 4500K-5500K, suppress overexposure, ensure clear target details, and prevent reflections from causing exposure; in a low light environment, when the light intensity is between 0.001 and 1... At 0 Lux: Maintaining the no-red-storm night vision mode, the spherical robot control system adjusts the sensitivity of the visual equipment to ISO 3200 and optimizes the exposure time to 0.3s-0.5s to reduce image noise and ensure that the image contrast is not less than 30:1. In the absence of light, i.e., when the light intensity is less than 0.001 Lux: Activate infrared supplementary lighting with no visible light, with a wavelength of 940nm. The intensity of the infrared supplementary lighting is automatically adjusted according to the target distance. The supplementary lighting intensity is 30% at a distance of 3-5 meters and 50% at a distance of 5-8 meters to avoid exposing the position of the spherical robot due to excessive supplementary lighting, while ensuring that the target is captured clearly.

[0047] Infrared supplementary light intensity = maximum supplementary light intensity × straight-line distance between the spherical robot and the target / 8.

[0048] Based on the distance between the current position and the target, calculate the adjustment value of the field of view of the vision device. When the target distance is 3 to 5 meters, this is the best situation for tracking at close range: the spherical robot control system issues a command to adjust the field of view to 30°, focusing on the target's face and movements, ensuring that the target occupies 30%-40% of the frame, in order to facilitate accurate evidence collection. When the target distance is 5 to 8 meters, this is the best situation for tracking at long range: adjust the field of view to 60°, expand the shooting range, cover the target's movement range, prevent the target from being lost due to exceeding the shooting range, and at the same time ensure that the target occupies 40%-50% of the frame. The field of view of the vision device = 12 × the straight-line distance between the spherical robot and the target - 6.

[0049] Based on the target's real-time coordinates and the current orientation of the vision device, the horizontal and vertical offsets are calculated. The rotation speed and direction of the gimbal are then calculated based on the offsets. The calculated field-of-view adjustment values ​​of the vision device, the rotation speed and direction of the gimbal are converted into adjustment commands and sent to the spherical robot to perform field-of-view adjustment operations and gimbal adjustments. The gimbal can perform self-stabilized rotation in both horizontal and vertical directions. Its horizontal rotation range is 0-360°, and its vertical rotation range is -30°-90°. The rotation speed is adaptively adjusted according to the target offset. When the offset is less than or equal to 5°, the rotation speed is 5° / s; when the offset is greater than 5°, the rotation speed is 10° / s, ensuring that the target can enter the shooting range and be centered in the image. The alignment accuracy is less than or equal to ±1°. When the target is obstructed by an obstacle, the spherical robot control system uses an image recognition algorithm to identify the position of the obstruction and fine-tunes the pitch angle of the vision device within ±5° to avoid the obstruction and maintain the feasibility of shooting. During the fine-tuning process, the vision device should be kept away from reflective surfaces or densely populated areas.

[0050] When the field of view of the vision device is less than 5°, the gimbal rotation speed is 5° per second; when the field of view of the vision device is not less than 5°, the gimbal rotation speed is 10° per second.

[0051] After adjustment, check whether the vision device is pointing towards a reflective surface or the direction of the person's line of sight. If the vision device is pointing towards a reflective surface or the direction of the person's line of sight, then perform a fine adjustment of the vision device's direction and re-detect; otherwise, do not make any adjustment.

[0052] The orientation of the vision device is monitored in real time, and a reflection detection algorithm is used to identify reflective surfaces such as light sources and glass. When the vision device is detected to be facing a reflective area, an adjustment command is immediately sent to adjust the orientation of the vision device. The adjustment response time is less than or equal to 0.2 seconds to avoid revealing the position of the spherical robot due to reflection. At the same time, the vision device is prohibited from pointing towards densely populated areas. A personnel detection algorithm is used to determine the number of people within the field of view of the vision device in real time. When the number of non-target personnel is greater than or equal to 3, the orientation is immediately fine-tuned to ensure that the concealment is not affected.

[0053] During the adjustment of the vision device, data such as vision device parameters, image clarity, and target alignment accuracy are collected in real time and transmitted synchronously to the spherical robot control system. The spherical robot control system compares the collected data with preset standards. If the standards are not met, the parameters are fine-tuned until the shooting requirements are met.

[0054] S5. Before performing position adjustment, a deep reinforcement learning model is used, combined with multi-dimensional evaluation indicators, to select and evaluate hidden points, screen hidden points, plan candidate paths based on the screened hidden points, evaluate the optimal hidden path and convert it into control commands to drive the spherical robot to perform position adjustment.

[0055] The selection and evaluation of concealment points include: classifying and organizing the data in the initial dataset and analysis dataset, then inputting them into the deep reinforcement learning model to perform format standardization and outlier removal. Specifically, the input includes environmental grid map data and target and point parameter data. The environmental grid map data contains four channels: an obstacle distribution grid that marks the location and outline of obstacles, a shadow coverage grid that marks the range and density of shadow areas, an open area grid that marks unobstructed and easily exposed areas, and a non-target line-of-sight coverage grid that marks areas that can be covered by the line of sight of non-target personnel. The target and point parameter data includes the coordinates of candidate concealment points, the real-time coordinates of the target, the target's movement speed, the target's orientation, the coordinates of non-target personnel, the light intensity, the current position coordinates, the distance between the spherical robot and the target, the current attitude angle of the spherical robot, the average height of obstacles, the shadow coverage ratio, and the overlap between the non-target line of sight and the current position of the spherical robot.

[0056] When inputting data, environmental raster map data is input in a channel-independent manner, while target and point parameter data are input in vector form. Data format standardization and validation are then performed, removing outlier data that exceeds reasonable limits. This yields raster map data and parameter vectors. Normalization is then applied to the raster map data and parameter vectors to obtain the normalized first raster map spatial features and the first parameter feature vector. The min-max normalization algorithm is used for standardization; for the raster map data, standardization is performed channel-by-channel; for the parameter vector data, standardization is performed for each dimension. Simultaneously, a random deactivation layer is introduced to randomly discard some redundant data. Spatial correlation features of obstacle distribution, shadow coverage, and non-target line-of-sight coverage areas are extracted from the first raster map spatial features and compressed. These features are then fused with the first parameter feature vector to obtain the first fused feature vector, specifically using three convolutional layers and global average pooling. In the stacked layers, the first layer uses the ReLU activation function after batch normalization to extract basic spatial features of individual obstacles and local shadow contours. The second layer uses the ReLU activation function after batch normalization to extract mid-level spatial features of the relationship between shadows and obstacles, as well as the concealment potential of local areas. The third layer uses the ReLU activation function after batch normalization to extract high-level spatial features of the spatial distribution patterns of concealed areas and the coverage of non-target line-of-sight areas. Global average pooling is used to compress the feature map, resulting in a grid map spatial feature vector. The ReLU activation function is used to initially fuse the grid map spatial feature vector with the first parameter feature vector, highlighting the impact of spatial features on concealment and path planning with weights of 0.7 and 0.3. The ReLU activation function is then used to further refine the initially fused features, removing redundant features irrelevant to decision-making. L2 regularization is introduced to obtain the first fused feature vector.

[0057] Combining a pre-defined decoupled reward function, each concealed location is scored across four dimensions: concealment, shooting effect, distance, and environmental adaptability, resulting in a comprehensive score. Specifically, the raster map spatial feature vector includes two dimensions: concealment and environmental adaptability. The scores of these two dimensions are weighted and summed. The first parameter feature vector includes three dimensions: shooting effect, concealment, and distance. These three dimensions are also weighted and summed. Finally, the comprehensive score of the concealed location is obtained by weighting the scores with 0.7 and 0.3 respectively. The comprehensive score of the concealed location is compared with a preset screening threshold. Concealed locations with comprehensive scores below the preset threshold are removed, resulting in a set of concealed locations. The preset screening threshold is 88 points. However, if the number of concealed locations with scores above 85 points is less than the set value, concealed locations with higher scores between 75 and 88 points are adopted as supplementary locations to avoid the inability to analyze paths later.

[0058] The concealment capability of candidate locations is assessed using the concealment score from the preceding steps. The feasibility of shooting at candidate locations is assessed using the same assessment method as the previous shooting assessment results. A score of 60 is considered "passing" and 0 is considered "basically passing". The evaluation assesses the reasonableness of the distance between the candidate point and the current position and the target, combining two sub-indicators: safe distance and path length. Specifically, the safe and optimal tracking distances are as follows: 100 points for a candidate point within 3-8 meters of the target; 60 points for 2-3 meters or 8-10 meters; and 0 points for <2 meters or >10 meters. The path length is as follows: 100 points for the shortest path from the candidate point to the spherical robot's current position; 60 points for a medium path length; and 20 points for the longest path length. The accessibility and anti-interference capabilities of candidate locations were evaluated. The accessibility evaluation criteria were as follows: if there were no obstacles obstructing the candidate location and the ground slope was less than or equal to 30°, 100 points were awarded; if there were minor obstacles that could be bypassed and the slope was between 20° and 30°, 60 points were awarded; if there were significant obstacles that could not be bypassed and the slope was greater than 30°, 0 points were awarded. The anti-interference capability evaluation criteria were as follows: if there was no strong light reflection and no significant environmental noise at the candidate location, 4 points were awarded; if there was minor interference, 2 points were awarded; if there was severe interference that affected the shooting or concealment effect, 0 points were awarded.

[0059] The output of a deep reinforcement learning model is a Q-value; the Q-value includes the comprehensive score of the hidden points and the overall value of the path. For ease of understanding, the hidden points and the path are described separately.

[0060] The deep reinforcement learning model also includes a priority experience replay layer. The function of this layer is to store the experience gained during the training process of the deep reinforcement learning model and perform weighted random sampling based on the priority of the experience. This improves the utilization efficiency of high-value experience, increases the convergence speed of the deep reinforcement learning model, and thus optimizes the accuracy of decision-making. A priority queue is used to store the experience data, calculate the priority of each experience, and employ a weighted random sampling algorithm to balance priority sampling and uniform sampling, reducing sampling bias. The sampled experience is then used to update the parameters of the deep reinforcement learning model. Gradient descent is used in conjunction with the loss function for optimization, and the optimized experience data is then re-stored into the experience replay pool, replacing the old low-value experience. This ensures that high-value samples are always preserved in the experience replay pool, forming a closed loop for the training of the deep reinforcement learning model.

[0061] Based on the candidate paths planned from the selected concealed locations, the evaluation of the optimal concealed path and its transformation into control commands includes: classifying and organizing the data from the initial dataset, analysis dataset, and concealed location set, and then inputting them into the deep reinforcement learning model to perform format standardization and outlier removal, obtaining normalized second grid map spatial features and second parameter feature vectors. In addition to the data from the concealed location evaluation, the input data also includes the core parameters of the selected concealed locations, covering location coordinates, comprehensive score, and reward scores for each dimension, serving as the endpoint benchmark for path planning and clarifying the core objective of path planning; it also includes the current coordinates, current attitude angle, maximum movement speed, diameter of the spherical robot, and the straight-line distance between the selected concealed location and the current position of the spherical robot, as well as the real-time coordinates of the target; it also includes ground slope distribution data, real-time light intensity data, real-time coordinates and movement trends of non-target personnel, with normalization in the same way as for the preceding concealed locations.

[0062] The algorithm extracts spatial features from the second grid map, including obstacle distribution, shadow area continuity, and spatial range correlation features related to the non-target line-of-sight coverage. These features are then compressed and fused with the second parameter feature vector to obtain the second fused feature vector. Similarly, a three-layer convolutional layer and global average pooling are used. Specifically, the ReLU activation function is applied to extract basic spatial features within the path planning range, such as the outline of a single obstacle, the range of local shadows, and the distribution details of ground slope, providing basic features for path traversability judgment. The ReLU activation function is also used to extract mid-level spatial correlation features, such as the positional correlation between the shadow area and the obstacle when determining whether a path can be planned along the shadow behind the obstacle, and the overlap relationship between the non-target line-of-sight coverage area and potential path nodes when predicting path exposure risk. Finally, the ReLU activation function is used to extract high-level spatial regularity features, such as the spatial distribution of shadows and obstacle occlusion, the passage gaps in densely obstructed areas, and blind spots not covered by the non-target line-of-sight coverage, providing core spatial basis for initial path planning and optimization. Vector compression reduces the number of parameters, improving computational efficiency while retaining core spatial features related to path planning.

[0063] The ReLU activation function is used for initial fusion. In this process, the spatial features of the second grid map mainly support the judgment of path concealment and accessibility, while the feature vector of the second parameter mainly supports the judgment of path distance and speed planning, achieving initial synergy between the two types of features. The ReLU activation function is then used to further refine the features of the initial fusion, eliminating redundant features that are not related to path planning and retaining core features, such as the spatial distribution of concealed areas, the gaps between obstacles, the distance features between the spherical robot and concealed points, the features of third-party blind spots, and the features of ground slope, to ensure the accuracy of subsequent Q-value decisions. L2 regularization is introduced to limit the case of excessively large parameters and avoid overfitting of the deep reinforcement learning model during path planning.

[0064] Candidate paths are planned, and the second fusion feature vectors of the candidate paths are converted into Q-values. By comparing the Q-values, the three candidate paths with the highest Q-values ​​are selected for quantization verification. When planning candidate paths, an initial shortest path is planned using the A* algorithm, followed by 20 candidate paths. Specifically, a single fully connected layer is used, whose output dimension corresponds to the number of candidate actions for the path. This ensures that the number of candidate paths initially generated by the A* algorithm matches the 20 candidate action dimensions initialized in the deep reinforcement learning model, outputting a 1×20 Q-value vector. A target network is then introduced to calculate the target Q-value and to calculate the Q-value for each candidate path. The Q-value of the path selection action is calculated by fully incorporating four constraints of path planning: First, concealment weight, with 40% of the Q-value corresponding to path concealment; the higher the concealment, the higher the Q-value. Second, accessibility weight, with 20% of the Q-value corresponding to path accessibility; if the path has no inescapable obstacles and the slope is ≤30°, the Q-value is higher. Third, distance efficiency weight, with 30% of the Q-value corresponding to path distance; the shorter the total path length, the higher the Q-value. Fourth, no-exposure weight, with 10% of the Q-value corresponding to no-exposure constraints; if the path does not pass through areas with strong light or dense crowds, the Q-value is higher.

[0065] like Figure 3 As shown, Figure 3 This is the display interface for the selected concealed locations. The black circle represents the furthest observation distance. The larger and darker the dot in the image, the better the concealment effect.

[0066] The three candidate paths with the highest Q-values ​​are selected as the initial optimization targets. Based on these three candidate paths, the path length is optimized using the A* algorithm to obtain three alternative paths with short distances and good concealment. At the same time, each alternative path is divided into nodes every 0.1m, and the node features are re-input into the deep reinforcement learning model to verify the concealment score and constraint satisfaction of each node. The candidate optimal concealment path that meets all constraints is then selected.

[0067] If any node in the candidate optimal concealed path does not meet the constraints (i.e., concealment score < 70 points or unavoidable obstacles), then a new candidate node is selected based on the Q value, and the path direction is adjusted. At the same time, the path length is optimized using the A* algorithm, and the verification process is repeated until the path meets all constraints, at which point it is determined as the optimal concealed path. Based on the Q value and path parameters of the optimal concealed path, the node coordinates, total path length, planned movement speed, and estimated transfer time of the optimal concealed path are output, and the spherical robot's position adjustment is initiated.

[0068] This invention combines reinforcement learning and quantitative evaluation techniques to achieve autonomous assessment of concealed locations and optimal concealed path planning. It eliminates the need for manual designation of concealed locations, solving the shortcomings of existing spherical robot control systems, which require manual designation of concealed locations and suffer from unreasonable and easily exposed path planning. It enables autonomous decision-making on concealed locations and paths, ensuring that the spherical robot can quickly move to a suitable concealed area while maintaining the optimal tracking distance to the target, thus improving the safety and efficiency of concealed tracking.

[0069] As the spherical robot moves along the optimal concealed path, it collects real-time environmental change data, such as the appearance of dynamic obstacles, movement of third-party personnel, and changes in concealment caused by target movement. This real-time data is synchronously transmitted to a deep reinforcement learning model. The deep reinforcement learning model performs real-time evaluation and decision-making. If the following situations occur, dynamic path adjustment is immediately initiated to ensure that the path always meets the requirements of concealment and feasibility. The entire adjustment process is led by the deep reinforcement learning model, with a response latency of no more than 0.3 seconds: 1. Dynamic obstacles appear on the path, such as moving debris or personnel, making it impossible to pass smoothly; 2. The coordinates and size data of the dynamic obstacles are input into the deep reinforcement learning model. After standardization and feature extraction, the model outputs the Q-value changes of the original path nodes, and the deep reinforcement learning model immediately triggers path optimization, reselects alternative concealed points, and plans a local obstacle avoidance path; 3. Movement of non-target personnel. This can cause some nodes on the path to enter areas outside the target's line of sight. The coordinate data of non-target personnel is updated, and the deep reinforcement learning model re-extracts the spatial features of the non-target line of sight coverage through convolution. After fusion, exposed nodes are identified, and path adjustments are initiated to bypass the line of sight coverage. Target movement reduces the concealment or shooting feasibility of the original optimal location. The deep reinforcement learning model re-evaluates the original optimal location; if the location does not meet the requirements, it generates 20 candidate locations, repeats the quantitative evaluation process, selects new optimal concealment locations, and then re-plans the optimal concealment path. Sudden changes in ambient light cause the shadow coverage of some nodes on the path to disappear, reducing concealment. The light intensity data is updated, and after receiving the data, the deep reinforcement learning model standardizes it, re-extracts shadow coverage features, identifies concealment locations with reduced concealment, initiates path fine-tuning, and selects new concealment locations in shadow areas.

[0070] This invention combines a hierarchical closed-loop control architecture with adaptive optimization technology, integrating the entire process modules of perception, decision-making, execution, and feedback. It achieves fully autonomous concealment and stable tracking after deployment, solving the problems of insufficient intelligence and the need for manual intervention in existing spherical robot control systems. It enables unmanned operation of the entire process from target locking and concealment control to stable tracking after deployment, with high overall tracking accuracy. This significantly improves the autonomous combat capability of spherical robots, expands the application scope in complex reconnaissance scenarios, reduces manual operation costs and exposure risks, and has good application prospects.

[0071] Example 2: The solution differs from Example 1 in that it also includes a subsequent specific adjustment process, as follows: Based on the complexity of the optimal path, the spherical robot's movement speed is adaptively adjusted to prevent instability and excessive noise due to excessive speed, or transfer delay due to insufficient speed. Specifically: on flat and obstacle-free paths, the movement speed is controlled at 0.2-0.3 m / s; on paths with dense obstacles, the movement speed is controlled at 0.1-0.2 m / s to ensure accurate obstacle avoidance and prevent collision noise; on turning sections, the movement speed is reduced to 0.1 m / s for slow turning to avoid posture deviations caused by centrifugal force and reduce motor noise; when approaching the optimal position, when the distance to the endpoint is less than 1 m, the movement speed is reduced to 0.05-0.1 m / s for slow approach to avoid collisions with obstacles around the point and ensure accurate arrival.

[0072] Specifically, a PID control algorithm is adopted to collect the deviation between the spherical robot's moving speed and the preset speed in real time. Through proportional, integral, and derivative adjustments, the speed deviation is controlled within ±0.02m / s to ensure speed stability. At the same time, the motor operating noise is controlled to not exceed 56dB, making it lower than the ambient noise level and avoiding the spherical robot's position being exposed by noise.

[0073] Based on the node coordinates of the optimal path, combined with the real-time position and attitude data of the spherical robot collected by the inertial measurement unit, a path tracking algorithm is used to ensure that the spherical robot moves accurately along the optimal path without deviating from it. Specifically, this includes: position deviation detection, which calculates the deviation between the spherical robot's current position and the path node coordinates in real time. If the deviation does not exceed 0.05m, the current direction of movement is maintained; if the deviation is greater than 0.05m, direction adjustment is immediately initiated. Precise direction adjustment uses a PID control algorithm to adjust the spherical robot's rolling direction. The direction adjustment speed is adaptively adjusted according to the magnitude of the deviation to ensure rapid correction of the deviation and return to the optimal path. Obstacle avoidance is supplemented by the vision device scanning the path ahead in real time. If a dynamic obstacle is detected, movement is immediately paused, and a local obstacle avoidance path is replanned through the path planning module. After obstacle avoidance is completed, the robot returns to the original optimal path.

[0074] When the spherical robot moves to a distance of no more than 0.05m from the destination, the point arrival confirmation process is triggered. The specific steps are as follows: Using a visual sensor and coordinate positioning algorithm, it is confirmed that the deviation between the current position of the spherical robot and the coordinates of the optimal point does not exceed 0.05m, and it is determined that the target point has been reached; The inertial measurement unit is activated to collect attitude data such as body pitch angle and roll angle in real time. A PID control algorithm is used to control the attitude angle error within ±0.5°, and the body is adjusted to a low attitude, close to the ground or the edge of an obstruction, to maximize concealment; After reaching the point, the concealment is re-evaluated. If the concealment score is not lower than 80 points, the position adjustment is completed; If the concealment score is lower than 80 points, the path fine-tuning is triggered, and the robot moves to a better position near the point, and the attitude is recalibrated until the concealment is qualified.

[0075] During stable tracking, attitude data is collected in real time through the inertial measurement unit, and the attitude is continuously calibrated using a PID control algorithm to ensure the stability of the robot body and avoid attitude deviations caused by ground vibrations or minor collisions. The attitude angle error is always controlled within ±0.5°. At the same time, the spherical robot is kept in a silent state and the motors are stopped to avoid revealing its position due to noise generated by movement.

[0076] During tracking, if the target's movement speed is greater than 3 m / s, shorten the target coordinate update cycle to 0.02 s, increase the orientation adjustment speed of the vision device, expand the field of view to 70°, and predict the target's movement trajectory in advance to avoid losing the target; when adjusting the position, increase the movement speed to 0.3-0.4 m / s to ensure that the target is quickly followed.

[0077] In scenarios with dense obstacles, the path tracking algorithm is optimized to prioritize movement through gaps between obstacles and in shadow areas, reducing the number of times obstacles need to be bypassed. During the tracking process, the pitch angle of the vision device is adjusted to avoid fixed obstacles and ensure that the target is not obscured.

[0078] In scenarios involving non-target personnel, strengthen the monitoring of non-target personnel, update the line-of-sight coverage area of ​​non-target personnel in real time, and strictly avoid the line-of-sight coverage area during path adjustment and tracking; if a non-target personnel approaches the spherical robot, control the spherical robot to stop all movement, turn off the supplementary light, maintain a low posture until the non-target personnel leave, and then resume tracking.

[0079] In scenarios with sudden changes in lighting, quickly adjust the parameters of the vision device to ensure image clarity; at the same time, adjust the robot's posture to stay in the shadow area as much as possible to avoid exposure to reflections.

[0080] Example 3: like Figure 2 As shown, a control system for a spherical robot based on artificial intelligence includes a data acquisition module, a target recognition module, a position verification module, a vision adjustment module, and a position adjustment module. Data acquisition module: Receives multi-dimensional data collected by the spherical robot, performs preprocessing and parsing, and obtains the initial dataset; Target recognition module: Uses target recognition algorithms to perform target recognition on the data in the initial dataset. After the target is recognized, it performs target locking and parameter analysis to obtain the analysis dataset. Location verification module: After the target is locked, a quantitative evaluation algorithm is used to perform a concealment assessment and shooting feasibility assessment of the current location from multiple dimensions based on the initial dataset and the analysis dataset. The concealment assessment result and shooting assessment result are obtained. If the concealment assessment result is unqualified, the location is adjusted; otherwise, the shooting assessment result is judged to be qualified. If the shooting assessment result is qualified, no action is taken. Visual adjustment module: If the shooting evaluation result is unqualified, environmental analysis is performed on the initial dataset, and the operation of the vision device is adjusted based on the environmental analysis result to obtain an adjustment and lock report. The adjustment and lock result is evaluated. If the adjustment and lock result is qualified, no action is taken; otherwise, position adjustment is performed. Position Adjustment Module: Before performing position adjustment, a deep reinforcement learning model is used, combined with multi-dimensional evaluation indicators, to select and evaluate hidden points, screen hidden points, plan candidate paths based on the screened hidden points, evaluate the optimal hidden path and convert it into control commands to drive the spherical robot to perform position adjustment.

[0081] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0082] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0083] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A control method for a spherical robot based on artificial intelligence, characterized in that, include: Receive multi-dimensional data collected by a spherical robot, and perform preprocessing and parsing to obtain an initial dataset; A target recognition algorithm is used to perform target recognition on the data in the initial dataset. After the target is recognized, target locking and parameter analysis are performed to obtain the analysis dataset. After the target is locked, a quantitative evaluation algorithm is used to perform a concealment assessment and a shooting feasibility assessment of the current location from multiple dimensions based on the initial dataset and the analysis dataset. The concealment assessment results and shooting assessment results are obtained. If the concealment assessment result is unqualified, the location is adjusted; otherwise, the shooting assessment result is judged to be qualified. If the shooting assessment result is qualified, no action is taken. If the shooting evaluation results are unsatisfactory, environmental analysis is performed on the initial dataset, and visual equipment operation is adjusted based on the environmental analysis results to obtain an adjustment and lock report. The adjustment and lock results are evaluated, and if the adjustment and lock results are satisfactory, no action is taken. Conversely, position adjustment will be performed; Before performing position adjustment, a deep reinforcement learning model is used, combined with multi-dimensional evaluation indicators, to select and evaluate hidden points, screen hidden points, plan candidate paths based on the screened hidden points, evaluate the optimal hidden path and convert it into control commands to drive the spherical robot to perform position adjustment.

2. The control method for a spherical robot based on artificial intelligence according to claim 1, characterized in that: The multi-dimensional data collected by the spherical robot includes: environmental images and potential target area images collected by the vision device on the spherical robot; and light intensity, environmental noise, ground slope and obstacle height collected by several sensors on the spherical robot.

3. The control method for a spherical robot based on artificial intelligence according to claim 2, characterized in that: The preprocessing and parsing process includes: denoising environmental images and potential target area images; analyzing ground slope and obstacle height; generating a raster map of passable areas; and marking potential concealed locations, open areas, and areas easily covered by human sight. Further, the denoised image data, illumination intensity, and environmental noise are combined to form a preprocessing set. A classification parsing algorithm is used to perform hierarchical processing on the preprocessing set to obtain an environmental parameter report and an image analysis report. The preprocessing set, environmental parameter report, and image analysis report are then summarized to form the initial dataset.

4. The control method for a spherical robot based on artificial intelligence according to claim 1, characterized in that: The initial dataset is processed using a target recognition algorithm. After target recognition, target locking and parameter analysis are performed, including: real-time frame parsing of the image data in the initial dataset to extract human contour features and motion trajectory features; calculating the matching score of the human contour features using a threshold method; comparing the matching score with a preset score threshold; if the matching score is lower than the preset score threshold, the target is identified as a non-target and marked as such; if the matching score is not lower than the preset score threshold, the target is identified as a target and locked; establishing a coordinate system with the current position of the spherical robot as the origin, and locking the real-time coordinates of the target using a coordinate positioning algorithm; analyzing the motion trajectory features of both non-targets and target personnel, marking the movement speed, target orientation, line-of-sight coverage area, and direction of movement, and summarizing these to form an analysis dataset.

5. The control method for a spherical robot based on artificial intelligence according to claim 4, characterized in that: A quantitative evaluation algorithm is employed to perform a concealment assessment of the current location from multiple dimensions based on the initial dataset and the analysis dataset. This includes: in the target line-of-sight avoidance dimension, calculating the angle between the target's orientation and the current location, and performing a target line-of-sight avoidance score based on this angle; in the non-target visibility dimension, calculating the average overlap between all non-target line-of-sight coverage areas and the current location, and performing a non-target visibility score based on this average overlap; in the occlusion recognition dimension, calculating a contrast score based on the current location's environmental color and the spherical robot's color, and calculating an occlusion volume ratio score based on the volume of occluders within the current location's range; a weighted summation method is used to process the contrast score and the occlusion volume ratio score to obtain the occlusion recognition score; a weighted summation is then performed on the target line-of-sight avoidance score, the non-target visibility score, and the occlusion recognition score to obtain the concealment score; and the concealment score is compared with a preset evaluation interval to obtain the concealment assessment result.

6. The control method for a spherical robot based on artificial intelligence according to claim 5, characterized in that: A quantitative evaluation algorithm is used to perform a feasibility assessment of shooting at the current location based on the initial dataset and the analysis dataset from multiple dimensions, including: in terms of shooting range, assessing whether the target is within the shooting range based on the field of view of the vision device, the straight-line distance between the spherical robot and the target, and the real-time coordinates of the target; in terms of image clarity, performing a comprehensive assessment based on light intensity, the parameters of the vision device, and the image analysis report; in terms of the proportion of the target in the image, performing an assessment based on the ratio of the pixel area of ​​the target in the image to the total pixel area of ​​the image based on the image segmentation algorithm; and in terms of target occlusion, performing an assessment based on the ratio of the pixel area of ​​the occluded part of the target to the pixel area of ​​the target in the image based on the target recognition algorithm. The evaluation results of the four dimensions are summarized and judged to obtain the shooting evaluation result.

7. The control method for a spherical robot based on artificial intelligence according to claim 6, characterized in that: Environmental analysis is performed on the initial dataset. Based on the results of the environmental analysis, the operation and control of the vision device are performed, including: extracting the light intensity and determining the lighting environment; adjusting the corresponding operating mode of the vision device based on the lighting environment; calculating the field of view adjustment value of the vision device based on the distance between the current position and the target; calculating the horizontal and vertical offsets based on the real-time coordinates of the target and the current orientation of the vision device, and calculating the rotation speed and direction of the gimbal based on the offsets; converting the calculated field of view adjustment value of the vision device, the rotation speed and direction of the gimbal into adjustment commands, and sending them to the spherical robot to perform the field of view control operation and gimbal adjustment; after adjustment, it is detected whether the vision device is pointing towards a reflective surface or the direction of the person's line of sight. If the vision device is pointing towards a reflective surface or the direction of the person's line of sight, the vision device's orientation is fine-tuned and the detection is repeated; otherwise, no adjustment is made.

8. The control method for a spherical robot based on artificial intelligence according to claim 7, characterized in that: The selection and evaluation of concealed locations includes: classifying and organizing the data in the initial dataset and analysis dataset, inputting them into the deep reinforcement learning model for format standardization and outlier removal to obtain raster map data and parameter vectors; performing normalization processing on the raster map data and parameter vectors to obtain normalized first raster map spatial features and first parameter feature vectors; extracting spatial correlation features of obstacle distribution and shadow coverage from the first raster map spatial features, performing compression processing, and fusing them with the first parameter feature vector to obtain the first fused feature vector; combining a pre-set decoupled reward function, performing scoring calculations on each concealed location from four dimensions: concealment, shooting effect, distance, and environmental adaptation to obtain a comprehensive score for the concealed location; comparing the comprehensive score of the concealed location with a preset screening threshold, and removing concealed locations with comprehensive scores lower than the preset screening threshold to obtain a set of concealed locations.

9. The control method for a spherical robot based on artificial intelligence according to claim 8, characterized in that: Based on the selected concealed locations, candidate paths are planned, and the optimal concealed path is evaluated and converted into control instructions. This process includes: classifying and organizing the data from the initial dataset, analysis dataset, and concealed location set, inputting them into a deep reinforcement learning model, performing format standardization and outlier removal to obtain normalized second grid map spatial features and second parameter feature vectors; extracting the spatial features of obstacle distribution, shadow area continuity, and spatial range correlation of non-target line-of-sight coverage from the second grid map spatial features, performing compression processing, and fusing them with the second parameter feature vector to obtain a second fused feature vector; planning candidate paths, converting the second fused feature vectors of the candidate paths into Q-values, comparing the Q-values, selecting the three candidate paths with the highest Q-values ​​for quantization verification, optimizing the nodes of the candidate paths, recalculating the Q-values, comparing the recalculated Q-values, and selecting the path with the highest Q-value as the optimal concealed path.

10. A control system for a spherical robot based on artificial intelligence, characterized in that: include: Data acquisition module: Receives multi-dimensional data collected by the spherical robot, performs preprocessing and parsing, and obtains the initial dataset; Target recognition module: Uses target recognition algorithms to perform target recognition on the data in the initial dataset. After the target is recognized, it performs target locking and parameter analysis to obtain the analysis dataset. Location verification module: After the target is locked, a quantitative evaluation algorithm is used to perform a concealment assessment and shooting feasibility assessment of the current location from multiple dimensions based on the initial dataset and the analysis dataset. The concealment assessment result and shooting assessment result are obtained. If the concealment assessment result is unqualified, the location is adjusted; otherwise, the shooting assessment result is judged to be qualified. If the shooting assessment result is qualified, no action is taken. Visual adjustment module: If the shooting evaluation result is unqualified, environmental analysis is performed on the initial dataset, and the operation of the visual equipment is adjusted based on the environmental analysis result to obtain an adjustment lock report. The adjustment lock result is evaluated. If the adjustment lock result is qualified, no processing is performed. Conversely, position adjustment will be performed; Position Adjustment Module: Before performing position adjustment, a deep reinforcement learning model is used, combined with multi-dimensional evaluation indicators, to select and evaluate hidden points, screen hidden points, plan candidate paths based on the screened hidden points, evaluate the optimal hidden path and convert it into control commands to drive the spherical robot to perform position adjustment.

Images