Shadowless lamp control system and method based on behavior recognition and prediction

Abstract

The invention relates to the field of intelligent control, and particularly discloses a shadowless lamp control method based on behavior recognition and prediction, which comprises the following steps: S1, establishing a three-dimensional rectangular coordinate system by taking an initial mounting position of a shadowless lamp as an original point, and determining coordinates as follows by adopting image data acquired by at least two groups of directional image acquisition equipment with different visual angles; s2, capturing the real-time position of the scalpel in real time through an image acquisition device, outputting coordinates, acquiring data of a plurality of continuous durations to form a historical trajectory data set, calling an operation scene template library preset with a plurality of types of typical operation trajectory features by the control terminal for the acquisition times, matching a trajectory feature template corresponding to the current operation type, and outputting the historical trajectory data set; performing feature alignment processing on the data and the input data by a fusion module to obtain adaptive trajectory data; according to the technical scheme, pre-judgment can be carried out in advance, the scene adaptability is high, and sterile fine adjustment is facilitated.

Description

Technical Field

[0001] This invention relates to the field of intelligent control, and in particular to a shadowless lamp control method based on behavior recognition and prediction. Background Technology

[0002] In surgical procedures, the shadowless lamp, as a core auxiliary device, directly affects the clarity of the surgical field of vision due to the precision and real-time nature of its light focusing, thus impacting the safety and efficiency of the surgical operation. With the development of medical technology, existing shadowless lamp control methods have evolved from early manual adjustment to automated tracking. For example, some technical solutions establish a three-dimensional coordinate axis to capture the position of the scalpel in real time and calculate its spatial distance and angle with the shadowless lamp, thereby driving the lamp to adjust its angle for tracking (as described in Chinese Patent Publication No. CN112767483A). While such technologies reduce manual intervention and improve the automation level of tracking to some extent, they still have significant shortcomings in adapting to the complex needs of surgical scenarios, making it difficult to meet the requirements of high-precision, low-latency, and sterile clinical operations.

[0003] The aforementioned automated tracking technology relies on real-time scalpel position feedback for passive adjustment, lacking the ability to anticipate the surgeon's intentions and exhibiting significant adjustment lag. During surgery, the surgeon moves the scalpel rapidly and along varied trajectories (e.g., the movement patterns differ significantly for different operations such as cutting, suturing, and dissection). When the scalpel's position changes, current technology requires a complete process of position capture, data calculation, and drive adjustment, which typically involves a certain delay. This lag can cause temporary light shifts in the surgical field, and brief visual blurring may increase the risk of operational errors, affecting surgical safety.

[0004] Furthermore, due to the simplistic tracking logic, the lighting requirements for shadowless lamps vary across different types of surgeries (such as abdominal surgery and orthopedic surgery) and different stages of the same surgery (such as incising the skin and handling internal organs). Using only a uniform coordinate calculation logic results in insufficient tracking accuracy and adaptability, still requiring frequent manual intervention.

[0005] In addition, if the surgeon needs to unlock the position of the light and scalpel in real time during the operation, that is, if the angle of the light needs to be finely adjusted during the operation (such as focusing and strengthening the light on local tissues), the surgeon still needs to rely on the assistance of an assistant. This not only increases the cost of personnel coordination, but may also lead to insufficient adjustment precision due to communication misunderstandings.

[0006] In summary, existing shadowless lamp control methods suffer from problems such as strong lag, inflexible adjustment of tracking logic, and complex personalized fine-tuning operations, which cannot fully meet the light control needs of surgical procedures. There is an urgent need for a control method that can achieve advance prediction, strong scene adaptability, and facilitate aseptic fine-tuning. Summary of the Invention

[0007] This invention provides a shadowless lamp control method based on behavior recognition and prediction, which can predict in advance, has strong scene adaptability, and is easy to fine-tune aseptically.

[0008] To solve the above-mentioned technical problems, this application provides the following technical solution: The shadowless lamp control method based on behavior recognition and prediction includes the following steps: S1. Initial installation position of the shadowless lamp Establish a three-dimensional Cartesian coordinate system with the origin, and acquire images using at least two sets of directional image acquisition devices with different viewpoints. Image data, determine Coordinates are ; S2. Real-time capture of the scalpel's position using an image acquisition device. and output coordinates Collect data for a continuous period of time. Data forms historical trajectory dataset , To determine the number of data collection attempts, the control terminal calls a pre-set surgical scenario template library containing typical surgical trajectory features for several types of surgeries, and matches the trajectory feature template corresponding to the current surgical type. ,Will and The input data fusion module, after feature alignment processing, obtains the adapted trajectory data. ; S3, will An LSTM behavior prediction model trained on trajectory data from multiple surgical scenarios is input, filtered, and then time series analysis is used to predict the target position of the scalpel within a preset time period. Output coordinates ; S4. The control terminal calculates the current position of the shadowless lamp. and lateral distance Longitudinal distance Altitude and distance ,according to Calculate the real-time angle Similarly, calculate and Target angle First adjust the driver of the shadowless lamp to ,like If the angle exceeds the preset value, then move. to Make the angle between the two less than or equal to the preset angle; S5. The image acquisition device captures images of the doctor's hands, extracts contour features after preprocessing, and matches them with a preset gesture library. If the match is successful, it outputs a fine-tuning command to drive the shadowless lamp to perform fine-tuning and monitors in real time that the included angle is less than the preset value. S6, Repeat S2-S5 updates at preset intervals. And optimize the LSTM behavior prediction model, if If the deviation from the actual position of the scalpel exceeds a preset value, the trajectory feature template will be rematched. .

[0009] The basic principle and beneficial effects of this invention are as follows: This invention continuously captures scalpel position data using an image acquisition device, constructs a historical trajectory dataset, and combines it with feature matching from a surgical scenario template library to accurately adapt the trajectory data to the operational patterns of the current surgical type. The adapted trajectory data is then input into an LSTM behavior prediction model trained on multiple surgical scenarios. Leveraging the LSTM model's strong ability to capture time-series data, and after filtering to reduce noise interference, the model performs time-series analysis and calculation on the scalpel movement trajectory within a preset timeframe, outputting the target position coordinates in advance.

[0010] Based on the predicted coordinates, the control terminal drives the shadowless lamp to adjust its angle and position in advance, realizing the control logic of trajectory prediction and active adaptation, thus fundamentally solving the problem of adjustment lag in traditional methods.

[0011] This invention establishes a scene template library containing typical trajectory features of several types of surgeries, covering the differences in operational trajectory patterns across different surgical types. By aligning historical trajectory datasets with the trajectory feature templates corresponding to the current surgery, the trajectory data is specifically adapted to the operational characteristics of the current surgery, avoiding adaptation biases caused by uniform calculation logic. Simultaneously, through periodic repetitive data collection, template matching, and model optimization processes, the trajectory dataset and model parameters are dynamically updated to ensure accurate trajectory feature matching at different stages of the same surgery (such as incision, dissection, and suturing), achieving comprehensive adaptation to multiple surgical types and operational stages.

[0012] Leveraging existing multi-view image acquisition equipment, the system directly captures images of the surgeon's hands and extracts their contour features, eliminating the need for additional contact-based operating devices and preventing disruption of the sterile surgical environment. A pre-defined gesture library defines clear mappings for fine-tuning commands, allowing surgeons to trigger adjustments with simple hand gestures, eliminating the need for assistants or verbal communication and simplifying the process. During fine-tuning, the angle between the shadowless lamp and the target position of the scalpel is monitored in real time to ensure that the preset angle requirements are met after adjustment. This guarantees flexibility in fine-tuning without compromising core lighting performance, enabling convenient and precise adjustments within a sterile environment.

[0013] By using LSTM model time-series prediction and active adjustment of the shadowless lamp, the light focusing lag time is significantly shortened, avoiding temporary blurring of the surgical field, reducing the risk of operational errors caused by field deviation, and ensuring surgical safety.

[0014] By using surgical scene template matching and feature alignment, the matching degree between trajectory data and current surgical operation patterns is improved. The accuracy of lighting adaptation for different surgical types and different operation stages is significantly improved, reducing the frequency of manual intervention and improving the continuity of surgical operations.

[0015] Non-contact gesture control eliminates the need to leave sterile areas, avoiding the risk of cross-infection. At the same time, it simplifies the operation process, meets doctors' personalized needs for local tissue lighting, and reduces personnel coordination costs and communication discrepancies.

[0016] The real-time angle monitoring and dynamic adjustment mechanism ensures that the central illumination of the surgical field remains stable, providing continuous and clear visual support for delicate surgical operations, especially suitable for high-precision surgeries such as neurosurgery and orthopedics.

[0017] Periodically updating the dataset and optimizing the model process allows the LSTM model to dynamically adapt to the doctor's personalized operating habits, and the prediction accuracy gradually improves as the surgery progresses.

[0018] Only at least two sets of directional image acquisition devices need to be deployed, without modifying the core hardware structure of the shadowless lamp. It can be adapted to existing mainstream shadowless lamp equipment, reducing the modification cost and technical threshold for clinical applications.

[0019] The combination of automated predictive adjustment and convenient gesture fine-tuning reduces the doctor's attention allocation to lighting adjustments, allowing the doctor to focus on the surgical procedure itself, indirectly improving surgical efficiency and operational accuracy.

[0020] In summary, this invention achieves the goals of enabling advance prediction, strong adaptability to various scenarios, and facilitating aseptic fine-tuning.

[0021] Furthermore, the filtering process in step S3 employs a dynamic weighted adaptive filtering algorithm, and the filtering formula is as follows: , in, The final adapted trajectory data, after filtering, is used as input to the LSTM behavior prediction model to improve prediction accuracy. For the first Dynamic weighting coefficients for each trajectory data point , For the first The scalpel movement speed corresponding to each trajectory data point The average moving speed over 10 consecutive trajectory data points. This is the speed sensitivity coefficient; Adapting trajectory data before filtering The first in One original trajectory data point, .

[0022] By dynamically allocating weights based on speed differences, the reliability of data in fast-moving trajectory segments is enhanced to improve prediction response speed, while noise interference in slow-moving segments is reduced to ensure trajectory smoothness. No additional hardware support is required, achieving both noise suppression and speed adaptation, thus solving the problem of prediction accuracy fluctuations when switching between different speeds.

[0023] Furthermore, in step S5, gesture recognition introduces a multi-dimensional matching degree calculation model, and the matching degree formula is: , in, This represents the overall matching degree between the current hand image and the target gesture, with a value ranging from 0 to 1. The match is considered successful at that time. The area of ​​the overlapping region between the current hand image contour and the target gesture contour is obtained through image contour intersection operation; The area of ​​the union region between the current hand image contour and the target gesture contour is obtained through image contour union operation; The shape similarity coefficient, , This is the difference in perimeter between the current hand outline and the target gesture outline, used to correct outline shape deviations; For position adaptation coefficient, , The straight-line distance between the current hand center and the center of the monitoring area of ​​the image acquisition device is preset to exclude invalid gestures outside the surgical area; The time stability coefficient, , This is the difference between the current gesture holding time and the target gesture standard holding time, used to filter out momentary erroneous actions.

[0024] By comprehensively evaluating multiple parameters, the system solves the problem of misidentification caused by slight hand tremors or positional shifts. It also avoids accidental adjustments triggered by random movements through time stability screening, eliminating the need to increase the number of image acquisition devices and controlling hardware costs.

[0025] Furthermore, in step S6, the deviation hierarchical feedback adjustment algorithm is adopted for optimization, and the adjustment formula is as follows: , in, These are the updated weight parameters for the LSTM behavior prediction model, used to optimize the next prediction calculation. These are the original weight parameters before the LSTM behavior prediction model is updated, and these are the parameter values ​​used in the previous prediction. The learning rate is tiered and dynamically adjusted based on the deviation level. ,when hour, ;when hour, ;when hour, ; The actual position coordinates of the scalpel The image is acquired in real time by an image acquisition device and then calibrated by coordinates. Predicted position coordinates of the scalpel , which is the output of the LSTM behavior prediction model; For scene adaptation coefficients, , The trajectory complexity coefficient corresponding to the current surgical type is preset according to different surgical types and is used to match the trajectory features of different surgeries.

[0026] By using graded learning rates, the model can be differentiated to avoid model oscillations caused by over-adjustment when the deviation is small, and the parameter correction speed can be accelerated when the deviation is large. The model can also be adapted to different surgical needs by combining scene complexity coefficients, thereby improving the model's versatility and prediction accuracy.

[0027] Furthermore, the feature alignment processing of the data fusion module in step S2 adopts trajectory feature point priority matching logic, including: firstly, from the historical trajectory dataset... Trajectory Feature Template Core and non-core feature points are extracted separately. Core feature points include the scalpel tip position and key turning points during the surgical operation, while non-core feature points include the scalpel handle position and transition points during the operation. Point-by-point precise matching is used for core feature points, requiring a matching error ≤ 0.8 mm. Interval matching is used for non-core feature points, allowing a matching error ≤ 2 mm. If the matching success rate of core feature points is ≥ 90% and the matching success rate of non-core feature points is ≥ 75%, the feature alignment is considered valid, and the adapted trajectory data is output. If the success rate of matching core feature points is less than 90%, then the feature points are extracted again and the matching process is executed again.

[0028] Differentiated matching is achieved by distinguishing the importance of feature points, prioritizing the alignment accuracy of the core surgical operation trajectory, reducing the interference of non-core feature point deviations on the fusion results, and improving scene adaptability.

[0029] Furthermore, in step S4, when adjusting the angle of the shadowless lamp, a dynamic illuminance compensation strategy is adopted, including: pre-setting a baseline illuminance according to the type of surgery. When the target angle is calculated Afterwards, if According to The illuminance increases proportionally to the difference from 45°; for every 10° decrease in the difference, the illuminance increases by [missing value]. Increase by 5%-8% on the basis; if Then, when moving the shadowless lamp to After ensuring the included angle is ≤45°, adjust the illumination to match the type of surgery. After adjustment, the illuminance at the center of the surgical field is monitored in real time using an illuminance sensor. If the illuminance is lower than the threshold, the illuminance will be detected. Then, fine-tune the illuminance again until the requirements are met.

[0030] By employing the above methods, we can avoid the decrease in surgical field illumination caused by changes in the angle, ensure stable field brightness under different surgical scenarios and angles, and improve the clarity of operation.

[0031] Furthermore, in step S3, the input layer of the LSTM behavior prediction model employs a trajectory temporal feature enhancement strategy, including: from adapted trajectory data... Extract the time interval between every two adjacent trajectory data points , Movement speed Calculate the average time interval of 10 consecutive data points. and average movement speed ;right and The absolute value of the difference exceeds 20%. and Trajectory segments with an absolute difference exceeding 30% are marked as critical time segments, and 1.2-1.5 times the feature weights are assigned to these trajectory segments during model training; trajectory segments with an absolute difference not exceeding the above threshold are marked as stationary time segments, and 1.0 times the base weights are assigned.

[0032] By enhancing the model's ability to capture key temporal features such as sudden changes in scalpel speed and abnormal time intervals through the above methods, the model can more accurately identify the trajectory patterns of different surgical operations and further improve the accuracy of predictions.

[0033] Furthermore, after the fine-tuning instruction is executed in step S5, a dual-precision verification mechanism is adopted, including: the first verification is angle verification, which captures the angle sensor data of the shadowless lamp through the image acquisition device and compares it with the actual angle after fine-tuning. From the perspective of the target If the angle deviation is ≤0.5°, the angle verification passes; if the deviation is >0.5°, a secondary fine-tuning command is output; the second verification is the illumination verification, which uses an illumination sensor to collect the actual illumination at the center of the surgical field of view. Compare the target illuminance after fine-tuning If the illuminance deviation rate ( / If the deviation rate (×100%) ≤ 5%, the illuminance verification passes; if the deviation rate > 5%, an illuminance compensation command is output; fine-tuning is considered complete only when both verifications pass.

[0034] By using the above methods, we can avoid abnormalities in the surgical field of view or illumination caused by fine-tuning deviations, ensure the precision of aseptic fine-tuning operations, and reduce interference with the surgical procedure.

[0035] Furthermore, in step S6, updating the historical trajectory dataset employs a sliding window combined with an anomaly correction strategy, including: the sliding window duration is fixed at 10 seconds, sliding forward once every 200ms, deleting the earliest 200ms data within the window, and adding the most recently collected 200ms data to form the updated dataset. ;right Each data point in , Calculate its relationship with the previous adjacent data point. The next data point If the distance is greater than 3mm, then it is judged. These are outlier data points; for outlier data points, linear interpolation is used to... and Replace the average coordinates After the correction is completed, if If the proportion of abnormal data points is ≤3%, they are directly used for subsequent fusion and prediction; if the proportion is >3%, 10 seconds of data are collected again to construct the dataset.

[0036] By employing the above methods, we can ensure the timeliness and accuracy of the dataset, providing reliable input data for the LSTM behavior prediction model and further optimizing the prediction effect and scenario adaptability. Attached Figure Description

[0037] Figure 1 This is a flowchart of a shadowless lamp control method based on behavior recognition and prediction. Detailed Implementation

[0038] The following detailed description illustrates the specific implementation method: Shadowless lamp control methods based on behavior recognition and prediction (e.g.) Figure 1 (As shown), including the following steps: S1. Initial installation position of the shadowless lamp Establish a three-dimensional Cartesian coordinate system with the origin, and acquire images using at least two sets of directional image acquisition devices with different viewpoints. Image data, determine Coordinates are ; S2. Real-time capture of the scalpel's position using an image acquisition device. and output coordinates Collect data for a continuous period of time. Data forms historical trajectory dataset , To determine the number of data collection attempts, the control terminal calls a pre-set surgical scenario template library containing typical surgical trajectory features for several types of surgeries, and matches the trajectory feature template corresponding to the current surgical type. ,Will and The input data fusion module, after feature alignment processing, obtains the adapted trajectory data. ; S3, will An LSTM behavior prediction model trained on trajectory data from multiple surgical scenarios is input, filtered, and then time series analysis is used to predict the target position of the scalpel within a preset time period. Output coordinates ; S4. The control terminal calculates the current position of the shadowless lamp. and lateral distance Longitudinal distance Altitude and distance ,according to Calculate the real-time angle Similarly, calculate and Target angle First adjust the driver of the shadowless lamp to ,like If the angle exceeds the preset value, then move. to Make the angle between the two less than or equal to the preset angle; S5. The image acquisition device captures images of the doctor's hands, extracts contour features after preprocessing, and matches them with a preset gesture library. If the match is successful, it outputs a fine-tuning command to drive the shadowless lamp to perform fine-tuning and monitors in real time that the included angle is less than the preset value. S6, Repeat S2-S5 updates at preset intervals. And optimize the LSTM behavior prediction model, if If the deviation from the actual position of the scalpel exceeds a preset value, the trajectory feature template will be rematched. .

[0039] In practical use: Deploy 3 sets of directional image acquisition devices (at least 2 sets are required), and install them directly above the operating table (acquiring the top view), to the left of the operating table (acquiring the front view), and to the right of the operating table (acquiring the side view). Set the device resolution to 1920×1080, the acquisition frequency to 50 frames / second, and calibrate the coordinate recognition accuracy to ≤0.5mm. A control terminal with data processing and model calculation capabilities is configured to communicate with the image acquisition equipment via a wired network with a latency of ≤20ms. The shadowless lamp is a medical shadowless lamp with electrically adjustable angle and position. Its angle adjustment range is 0-90°, and its position adjustment range covers an area of ​​1.5-2.5m above the operating table. It is equipped with an angle sensor and an illuminance sensor for real-time feedback of angle data and illuminance data, respectively.

[0040] Following step S1, the shadowless lamp is initially installed in its designated position. A three-dimensional Cartesian coordinate system was established with the origin as the origin, where the X-axis is the horizontal axis (parallel to the width of the operating table), the Y-axis is the horizontal axis (parallel to the length of the operating table), and the Z-axis is the vertical axis (perpendicular to the plane of the operating table). Images were simultaneously captured using three sets of image acquisition devices. The installation reference point is determined by the control terminal after performing calibration processing such as grayscale correction and distortion correction on the three sets of image data collected. The coordinates are .

[0041] The surgical scenario template library includes pre-set templates featuring typical trajectory characteristics for five types of surgeries: neurosurgery, orthopedics, general surgery, abdominal surgery, and ophthalmology. Each template contains 100 sets of clinical surgical trajectory data (derived from standardized surgical operation recordings by different doctors), such as orthopedic surgery templates. The system includes trajectory features for operations such as bone cutting, drilling, and repositioning. The trajectory point density is 10 points / ms in the cutting segment where the blade tip moves faster, and 5 points / ms in the repositioning segment where the blade tip moves slower. The preset gesture library defines four core fine-tuning gestures, including left-hand left-right swipe (corresponding to X-axis adjustment of +1°, +2°, +3° or -1°, -2°, -3°), left-hand up-down swipe (corresponding to Z-axis adjustment of +1°, +2°, +3° or -1°, -2°, -3°), right-hand left-right swipe (with backup X-axis adjustment), and right-hand up-down swipe (with backup Z-axis adjustment). The standard holding time for each gesture is set to 0.3 seconds.

[0042] The LSTM behavior prediction model has an input layer dimension of 3 (corresponding to X, Y, and Z three-dimensional coordinates), two hidden layers with 64 neurons per layer, and an output layer dimension of 3 (corresponding to the predicted three-dimensional coordinates). The model training data consisted of 500 sets of clinical trajectory data from the five surgical procedures mentioned above, each set lasting 10 seconds, collected at a frequency of 50 frames per second. After data cleaning (outlier removal), the data was divided into training and validation sets in a 7:3 ratio. The training iterations were 100 times, with an initial learning rate of 0.001. After training, the model's prediction error on the validation set was ≤2mm.

[0043] Taking orthopedic fracture reduction surgery as an example, the implementation process of steps S2-S6 is explained in detail: S2, After the surgery begins, the image acquisition device captures the real-time position of the scalpel at a frequency of 50 frames per second. ( The data is collected using timestamps (in milliseconds), and the coordinates corresponding to each frame are output. For example, suppose in At that time, the data was collected. The coordinates are (Unit: mm, same below); At that time, the data was collected. The coordinates are And so on.

[0044] The control terminal continuously collects data for 10 consecutive seconds. Data, forming a historical trajectory dataset , of which frames per second Each data collection attempt. The control terminal, through the surgery type selection interface, identifies the current surgery type as orthopedic surgery and retrieves the corresponding trajectory feature template from the surgical scene template library. ,Will and The input data fusion module performs feature alignment processing.

[0045] Feature alignment processing employs trajectory feature point priority matching logic: first from... Extract core feature points (scalpel tip position, key turning points during fracture reduction) and non-core feature points (handle position, transition points during movement), from... Extract the corresponding core and non-core feature points; perform precise point-by-point matching on the core feature points, requiring a matching error ≤ 0.8mm, for example... The coordinates of the core point of a certain blade tip are , The corresponding core point coordinates are The error is 0.4mm, which meets the matching requirements; for non-core feature points, interval matching is used, and the matching error is allowed to be ≤2mm, for example... The coordinates of a non-core point of a certain tool holder are: , The corresponding non-core point coordinates are The error is 1.8mm, which meets the matching requirements. If the success rate of matching core feature points is ≥90% and the success rate of matching non-core feature points is ≥75%, then the feature alignment is deemed effective, and the adapted trajectory data is output. In this embodiment, the success rate of matching core feature points is 92%, and the success rate of matching non-core feature points is 80%, resulting in a successful output. .

[0046] S3, will The input LSTM behavior prediction model, trained with trajectory data from multiple surgical scenarios, is first filtered using a dynamic weight adaptive filtering algorithm. The filtering formula is as follows: , in, This is the final adapted trajectory data output after filtering. For the first Dynamic weighting coefficients for each trajectory data point , This is the velocity sensitivity coefficient, with a value of 0.02. For the first The scalpel movement speed corresponding to each trajectory data point The average moving speed of 10 consecutive trajectory data points; Adapting trajectory data before filtering The first in One original trajectory data point ( ).

[0047] If in the 200th-210th trajectory data points, , , ..., Calculated ,but , The above method assigns higher weight to fast-moving trajectory segments and weakens noise interference in slow-moving segments.

[0048] After filtering Inputting data into an LSTM model, the model uses time series analysis to predict the target position of the scalpel within the next 0.5 seconds (preset time). For example, the current timestamp. The model is based on Temporal characteristics of the first 500 trajectory data points for prediction Target position of the scalpel Coordinates are .

[0049] S4, the control terminal first obtains the current position of the shadowless lamp. The coordinates, assuming The coordinates are ,calculate and (Assuming the current situation) Coordinates are ) lateral distance Longitudinal distance Altitude and distance (The negative sign indicates that the scalpel is under the operating light), assuming according to the formula Calculate the real-time angle ,get .

[0050] Similarly, calculate and (coordinate ) target angle Horizontal distance Longitudinal distance altitude distance , , Assuming we get .

[0051] The preset angle is set to 45°, because The control terminal drives the shadowless lamp to move to ,Adjustment The coordinates are Recalculate and Angle: lateral distance Longitudinal distance altitude distance Assuming we get Still greater than 45°, continue moving to Coordinates are Ultimately, through multiple fine-tunings, it was achieved... and The included angle is ≤45°, assuming that in this embodiment it is finally adjusted to .

[0052] Meanwhile, a dynamic illuminance compensation strategy is adopted when adjusting the angle of the shadowless lamp: a baseline illuminance assumption for orthopedic surgery is preset. ,because , The difference from 45° is 0.2°. Following the rule of increasing illuminance by 6% for every 10° decrease in the difference, the illuminance increase ratio is calculated and adjusted. After adjustment, the illuminance sensor monitors the illuminance at the center of the surgical field in real time. If the detected actual illuminance is lower than 45°, the sensor will detect the illuminance. Then the control terminal outputs an illuminance compensation command to drive the shadowless lamp to increase the illuminance and meet the requirements.

[0053] S5, during fracture reduction surgery, the doctor needs to enhance the illumination of the local bone area and trigger fine-tuning commands by sliding the left hand up and down. After the image acquisition device captures the image of the doctor's left hand, the control terminal preprocesses the image, including grayscale conversion, binarization, edge detection, and extraction of hand contour features.

[0054] Gesture recognition introduces a multi-dimensional matching degree calculation model, with the matching degree formula as follows: ,

[0055] in, This represents the area of ​​the overlapping region between the current hand image outline and the target's up-and-down swiping gesture outline. Let be the area of ​​the union region of the two. The shape similarity coefficient, , Let the difference in perimeter between the current hand contour and the target gesture contour be assumed to be... pixels, then ; For position adaptation coefficient, , Let the straight-line distance between the current hand center and the center of the preset monitoring area of ​​the image acquisition device be assumed to be... pixels, then ; The time stability coefficient, The current gesture is held for 0.32 seconds. Seconds, assuming .

[0056] Assuming Upon successful matching, a fine-tuning command of +2° Z-axis is output. After the shadowless lamp performs the fine-tuning, a dual precision verification mechanism is employed: in angle verification, the deviation between the actual angle and the target angle is 0.2°≤0.5°, and the verification passes; in illuminance verification, assuming the actual illuminance is 120300 lux and the target illuminance is 120144 lux, the illuminance deviation rate verification passes, and the fine-tuning is deemed complete. During the fine-tuning process, the included angle is monitored in real time to ensure it is ≤45°. In this embodiment, the included angle after fine-tuning is 44.6°, which meets the requirements.

[0057] S6 sets the process to repeat steps S2-S5 every 200ms, meaning the historical trajectory dataset is updated every 200ms. (Using a sliding window combined with an anomaly correction strategy): The sliding window duration is fixed at 10 seconds, and it slides forward once every 200ms, deleting the earliest 200ms (10 data points) of data in the window and adding the most recently collected 200ms of data to form an updated dataset. .

[0058] right Each data point in Calculate its relationship with the previous adjacent data point. The next data point The distance of a certain data point The coordinates are , The coordinates are , The coordinates are Assuming and The distance is ,and The distance is All values ​​≤3mm are considered normal data points; if a certain data point The coordinates are , The coordinates are , The coordinates are Calculated and The distance is ,and The distance is Data points identified as outliers were analyzed using linear interpolation. and Replace the average coordinates Assuming .

[0059] After the correction is completed, if If the proportion of outlier data points is ≤3%, they are directly used for subsequent fusion and prediction; if the proportion is >3%, 10 seconds of data are collected again to construct the dataset. Simultaneously, the control terminal optimizes the LSTM behavior prediction model using a bias-level feedback adjustment algorithm, with the following adjustment formula: ,

[0060] in, These are the weight parameters after the model update. These are the original weight parameters before the model update; For graded learning rates, bias ,like ,but ;like ,but ;like ,but ; The actual coordinates of the scalpel's position. The location coordinates predicted by the model; For scene adaptation coefficients, Assuming the trajectory complexity coefficient of orthopedic surgery ,but .

[0061] If the model predicts Actual collection , assuming ,but Substitute these values ​​into the formula to update the model weight parameters, thus achieving dynamic optimization of the model. If... Then rematch the trajectory feature template. For example, when At this time, the control terminal recalls the surgical scene template library again and matches the trajectory feature template corresponding to the orthopedic surgery. Then, feature alignment is re-executed to ensure the accuracy of subsequent predictions.

[0062] In step S3, the input layer of the LSTM behavior prediction model employs a trajectory temporal feature enhancement strategy, from... Extract the time interval between every two adjacent trajectory data points and movement speed Calculate the average time interval of 10 consecutive data points. and average movement speed For example, 10 consecutive data points The values ​​are 20ms, 21ms, 19ms, 22ms, 20ms, 18ms, 23ms, 21ms, 19ms, and 20ms respectively. The calculations yield... ; The values ​​are 0.7 mm / ms, 0.8 mm / ms, 0.9 mm / ms, 0.8 mm / ms, 0.7 mm / ms, 0.9 mm / ms, 1.0 mm / ms, 0.8 mm / ms, 0.9 mm / ms, and 0.7 mm / ms, respectively. The calculated values ​​are... .right and The absolute value of the difference exceeds 20% (i.e., >4.06ms). and Track segments with an absolute difference exceeding 30% (i.e., >0.252 mm / ms) are marked as critical timing segments. In this embodiment, a certain track segment... The absolute value of the difference is 4.7ms > 4.06ms. The absolute difference value of 0.36 mm / ms > 0.252 mm / ms is marked as a key time segment and assigned 1.3 times the feature weight; the trajectory segment whose absolute difference value does not exceed the above threshold is marked as a stationary time segment and assigned 1.0 times the basic weight to enhance the model's ability to capture key time features.

[0063] Before acquiring the scalpel position in step S2, an environment-adaptive noise reduction strategy is adopted for the image acquisition device. The ambient light intensity of the surgical area is detected in real time by the light sensor built into the image acquisition device. .like (In well-lit conditions) a light noise reduction process is used, preserving detail pixels with a grayscale value change rate ≥15% and smoothing noise pixels with a grayscale value change rate <15%; if (200 lux ≤ L < 500 lux, moderate lighting), then moderate noise reduction processing is used, retaining detail pixels with a grayscale value change rate ≥ 10% and smoothing noise pixels with a grayscale value change rate < 10%; if (In low light) heavy noise reduction processing is used. First, the image enhancement algorithm is used to improve the details in the dark areas. Then, the detail pixels with a gray value change rate of ≥8% are retained, and the noise pixels with a gray value change rate of <8% are smoothed to ensure the accuracy of the scalpel position capture.

[0064] Through the above specific implementation methods, the present invention realizes the core functions of shadowless lamps, such as advance prediction, strong scene adaptability, and easy aseptic fine adjustment. At the same time, it has technical effects such as fast response speed, high recognition accuracy, and good lighting stability, which fully meets the high precision, low hysteresis, and aseptic requirements of surgical operations for light control.

[0065] The above are merely embodiments of the present invention. The invention is not limited to the fields covered by these embodiments. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are able to access all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A shadowless lamp control method based on behavior recognition and prediction, characterized in that, Includes the following steps: S1. Initial installation position of the shadowless lamp Establish a three-dimensional Cartesian coordinate system with the origin, and acquire images using at least two sets of directional image acquisition devices with different viewpoints. Image data, determine Coordinates are ; S2. Real-time capture of the scalpel's position using an image acquisition device. and output coordinates Collect data for a continuous period of time. Data forms historical trajectory dataset , To determine the number of data collection attempts, the control terminal calls a pre-set surgical scenario template library containing typical surgical trajectory features for several types of surgeries, and matches the trajectory feature template corresponding to the current surgical type. ,Will and The input data fusion module, after feature alignment processing, obtains the adapted trajectory data. ; S3, will An LSTM behavior prediction model trained on trajectory data from multiple surgical scenarios is input, filtered, and then time series analysis is used to predict the target position of the scalpel within a preset time period. Output coordinates ; S4. The control terminal calculates the current position of the shadowless lamp. and lateral distance Longitudinal distance Altitude and distance ,according to Calculate the real-time angle Similarly, calculate and Target angle First adjust the driver of the shadowless lamp to ,like If the angle exceeds the preset value, then move. to Make the angle between the two less than or equal to the preset angle; S5. The image acquisition device captures images of the doctor's hands, extracts contour features after preprocessing, and matches them with a preset gesture library. If the match is successful, it outputs a fine-tuning command to drive the shadowless lamp to perform fine-tuning and monitors in real time that the included angle is less than the preset value. S6, Repeat S2-S5 updates at preset intervals. And optimize the LSTM behavior prediction model, if If the deviation from the actual position of the scalpel exceeds a preset value, the trajectory feature template will be rematched. .

2. The shadowless lamp control method based on behavior recognition and prediction according to claim 1, characterized in that, The filtering process in step S3 employs a dynamic weighted adaptive filtering algorithm, and the filtering formula is as follows: ， in, The final adapted trajectory data, after filtering, is used as input to the LSTM behavior prediction model to improve prediction accuracy. For the first Dynamic weighting coefficients for each trajectory data point , For the first The scalpel movement speed corresponding to each trajectory data point The average moving speed over 10 consecutive trajectory data points. This is the speed sensitivity coefficient; Adapting trajectory data before filtering The first in One original trajectory data point, .

3. The shadowless lamp control method based on behavior recognition and prediction according to claim 2, characterized in that, In step S5, gesture recognition introduces a multi-dimensional matching degree calculation model. The matching degree formula is as follows: ， in, This represents the overall matching degree between the current hand image and the target gesture, with a value ranging from 0 to 1. The match is considered successful at that time. The area of ​​the overlapping region between the current hand image contour and the target gesture contour is obtained through image contour intersection operation; The area of ​​the union region between the current hand image contour and the target gesture contour is obtained through image contour union operation; The shape similarity coefficient, , This is the difference in perimeter between the current hand outline and the target gesture outline, used to correct outline shape deviations; For position adaptation coefficient, , The straight-line distance between the current hand center and the center of the monitoring area of ​​the image acquisition device is preset to exclude invalid gestures outside the surgical area; The time stability coefficient, , This is the difference between the current gesture holding time and the target gesture standard holding time, used to filter out momentary erroneous actions.

4. The shadowless lamp control method based on behavior recognition and prediction according to claim 3, characterized in that, In step S6, the optimization adopts a deviation hierarchical feedback adjustment algorithm, and the adjustment formula is: ， in, These are the updated weight parameters for the LSTM behavior prediction model, used to optimize the next prediction calculation. These are the original weight parameters before the LSTM behavior prediction model is updated, and these are the parameter values ​​used in the previous prediction. The learning rate is tiered and dynamically adjusted based on the deviation level. ,when hour, ;when hour, ;when hour, ; The actual position coordinates of the scalpel The image is acquired in real time by an image acquisition device and then calibrated by coordinates. Predicted position coordinates of the scalpel , which is the output of the LSTM behavior prediction model; For scene adaptation coefficients, , The trajectory complexity coefficient corresponding to the current surgical type is preset according to different surgical types and is used to match the trajectory features of different surgeries.

5. The shadowless lamp control method based on behavior recognition and prediction according to claim 4, characterized in that, In step S2, the feature alignment processing of the data fusion module adopts trajectory feature point priority matching logic, including: firstly, from the historical trajectory dataset... Trajectory Feature Template Core and non-core feature points are extracted separately. Core feature points include the scalpel tip position and key turning points during the surgical operation, while non-core feature points include the scalpel handle position and transition points during the operation. Point-by-point precise matching is used for core feature points, requiring a matching error ≤ 0.8 mm. Interval matching is used for non-core feature points, allowing a matching error ≤ 2 mm. If the matching success rate of core feature points is ≥ 90% and the matching success rate of non-core feature points is ≥ 75%, the feature alignment is considered valid, and the adapted trajectory data is output. If the success rate of matching core feature points is less than 90%, then the feature points are extracted again and the matching process is executed again.

6. The shadowless lamp control method based on behavior recognition and prediction according to claim 5, characterized in that, In step S4, when adjusting the angle of the shadowless lamp, a dynamic illuminance compensation strategy is adopted, including: pre-setting the baseline illuminance according to the type of surgery. When the target angle is calculated Afterwards, if According to The illuminance increases proportionally to the difference from 45°; for every 10° decrease in the difference, the illuminance increases by [missing value]. Increase by 5%-8% on the basis; if Then, when moving the shadowless lamp to After ensuring the included angle is ≤45°, adjust the illumination to match the type of surgery. After adjustment, the illuminance at the center of the surgical field is monitored in real time using an illuminance sensor. If the illuminance is lower than the threshold, the illuminance will be detected. Then, fine-tune the illuminance again until the requirements are met.

7. The shadowless lamp control method based on behavior recognition and prediction according to claim 6, characterized in that, In step S3, the input layer of the LSTM behavior prediction model employs a trajectory temporal feature enhancement strategy, including: from adapted trajectory data... Extract the time interval between every two adjacent trajectory data points , Movement speed Calculate the average time interval of 10 consecutive data points. and average movement speed ;right and The absolute value of the difference exceeds 20%. and Trajectory segments with an absolute difference exceeding 30% are marked as critical time segments, and 1.2-1.5 times the feature weights are assigned to these trajectory segments during model training; trajectory segments with an absolute difference not exceeding the above threshold are marked as stationary time segments, and 1.0 times the base weights are assigned.

8. The shadowless lamp control method based on behavior recognition and prediction according to claim 7, characterized in that, After the fine-tuning command is executed in step S5, a dual-precision verification mechanism is adopted, including: the first verification is angle verification, which captures the angle sensor data of the shadowless lamp through the image acquisition device and compares it with the actual angle after fine-tuning. From the perspective of the target If the angle deviation is ≤0.5°, the angle verification passes; if the deviation is >0.5°, a secondary fine-tuning command is output; the second verification is the illumination verification, which uses an illumination sensor to collect the actual illumination at the center of the surgical field of view. Compare the target illuminance after fine-tuning If the illuminance deviation rate ( / If the deviation rate (×100%) ≤ 5%, the illuminance verification passes; if the deviation rate > 5%, an illuminance compensation command is output; fine-tuning is considered complete only when both verifications pass.

9. The shadowless lamp control method based on behavior recognition and prediction according to claim 8, characterized in that, Step S6 updates the historical trajectory dataset using a sliding window combined with an anomaly correction strategy, including: the sliding window duration is fixed at 10 seconds, sliding forward once every 200ms, deleting the earliest 200ms data within the window, and adding the most recently collected 200ms data to form the updated dataset. ;right Each data point in , Calculate its relationship with the previous adjacent data point. The next data point If the distance is greater than 3mm, then it is judged. These are outlier data points; for outlier data points, linear interpolation is used to... and Replace the average coordinates After the correction is completed, if If the proportion of abnormal data points is ≤3%, they are directly used for subsequent fusion and prediction; if the proportion is >3%, 10 seconds of data are collected again to construct the dataset.

10. A shadowless lamp control system based on behavior recognition and prediction, characterized in that, The method described in any one of claims 1-9 was adopted.

Images