Efficient and high-precision surgical action detection system based on YOLOv11 architecture

Abstract

The invention belongs to the technical field of computer vision and medical intelligent assistance, and discloses a high-efficiency and high-precision surgical action detection system based on a YOLOv11 architecture, the system takes YOLOv11 as a basic architecture, and the system is mainly improved by the following steps: deeply embedding an ESPPF module in a backbone network, and cooperatively capturing multi-scale spatial features through multi-branch convolution and pooling; an ST-MCAM module is inserted into the neck network, and collaborative modeling of multi-scale spatial features and inter-frame time sequence information is achieved; and replacing a prediction head with a DDPH prediction head, separating classification and bounding box regression tasks, and performing independent optimization. Through the synergistic effect of the three core modules, the multi-scale feature characterization capability, the space-time information utilization efficiency and the classification and positioning precision of surgical action detection are effectively improved, meanwhile, good real-time performance is kept, the system can be integrated into an intelligent surgical auxiliary system, technical support is provided for real-time monitoring and auxiliary decision making of minimally invasive surgery, and the system is suitable for popularization and application. The important clinical application value is realized.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and medical intelligent assistance technology, specifically relating to a high-efficiency and high-precision surgical action detection system based on the YOLOv11 architecture, which is applicable to intelligent surgical assistance equipment, surgical process recording and analysis, medical training and other scenarios. Background Technology

[0002] Minimally invasive surgery, with its significant advantages such as minimal trauma and rapid recovery, has become the mainstream surgical method in clinical surgery. Accurately recognizing the surgeon's movements during minimally invasive surgery is a core element in developing intelligent surgical assistance systems, which can provide crucial data support for surgical navigation, risk warning, postoperative review, and medical teaching.

[0003] Currently, deep learning-based object detection technology has been widely applied in general scenarios, among which the YOLO series architecture has become the preferred solution for real-time detection tasks due to its combination of detection speed and accuracy. YOLOv11, as the latest iteration of this series, adopts a classic three-segment structure of Backbone, Neck, and Head. The core design revolves around "lightweight and efficient + improved accuracy": In the Backbone, multi-scale feature extraction is achieved through initial convolutional layer downsampling, C3k2 module (dynamically selecting 3×3 / 5×5 convolutional kernels), SPPF module (fast spatial pyramid pooling), and C2PSA module (embedding position-sensitive attention), taking into account both details and semantic information; the Neck uses BiFPN bidirectional feature pyramid and C3k2 module reuse to optimize the efficiency of high-low layer feature fusion; the Head uses depthwise separable convolution, decoupled prediction branches, and multi-scale output to improve the accuracy and speed of object detection while reducing the number of parameters, demonstrating excellent comprehensive performance in general scenarios.

[0004] However, directly applying the YOLOv11 infrastructure to surgical motion detection still faces the following key technical bottlenecks:

[0005] 1. Surgical scenes have complex spatiotemporal characteristics. The visual features of surgical instruments and human tissues are highly similar, and the sequence of surgical actions is continuous and variable. YOLOv11's native SPPF module and feature fusion mechanism are difficult to capture the multi-scale spatial details and temporal dynamic information required by the surgical scene.

[0006] 2. The objective functions of surgical action classification and localization tasks differ. Although the YOLOv11 prediction head design has certain decoupling characteristics, it does not fully adapt to the task-specific requirements of surgical action detection, which can easily lead to inter-task interference and reduce detection accuracy and localization accuracy.

[0007] 3. Minimally invasive surgery places extremely high demands on the real-time performance of auxiliary systems. Existing solutions that simply enhance feature extraction capabilities often come with increased model complexity. YOLOv11's lightweight design is geared towards general-purpose scenarios, making it difficult to achieve a balance between "high-precision capture of complex surgical features" and "meeting real-time response requirements" when applied directly.

[0008] Therefore, there is an urgent need for a surgical motion detection solution that optimizes and improves the YOLOv11 architecture to suit the characteristics of the surgical environment, combining high efficiency and high accuracy, in order to solve the aforementioned technical bottlenecks. Summary of the Invention

[0009] The purpose of this invention is to overcome the problems of insufficient multi-scale feature capture, weak temporal context awareness, interference from classification and localization tasks, and difficulty in balancing real-time performance and accuracy in existing technologies for surgical action detection. It provides a high-efficiency, high-precision surgical action detection system based on the YOLOv11 architecture. By integrating three core enhancement modules—ESPPF, ST-MCAM, and DDPH—it specifically strengthens the spatiotemporal feature capture and task decoupling capabilities of surgical scenarios while retaining the lightweight characteristics of YOLOv11, significantly improving the accuracy and robustness of surgical action detection and providing core technical support for the development of intelligent surgical assistance systems.

[0010] To achieve the above objectives, this invention provides a high-efficiency and high-precision surgical action detection system based on the YOLOv11 architecture. The system is an improvement upon the YOLOv11 architecture, comprising three core components: a backbone network, a neck network, and a prediction head. The backbone network performs initial feature extraction and hierarchical feature enhancement on the input surgical video frames. Subsequently, the neck network performs multi-scale feature fusion and spatiotemporal information modeling on the extracted feature maps, integrating low-level detail features with high-level semantic features and inter-frame dynamic information. Finally, the prediction head performs category recognition and location localization of the surgical action based on the fused features, outputting the detection results.

[0011] The backbone network adopts the basic structure of YOLOv11, including standard convolutional layers Conv and C3K2 convolutional modules. It extracts hierarchical features of the input surgical video frame images by progressive downsampling. The deep layers of the backbone network (after the C3K2 module and before the Neck network) embed an enhanced fast spatial pyramid pooling module (ESPPF). Through the synergistic effect of multi-branch convolution and pooling operations, it captures fine-grained local features and wide-area multi-scale spatial features of surgical actions.

[0012] The neck network is based on the YOLOv11 multi-scale feature fusion path, which includes a C2PSA attention module. It achieves the initial fusion of low-level and high-level features through upsampling, skip connections and C3K2 convolution module reuse. A spatiotemporal multi-scale context aggregation module (ST-MCAM) is inserted into the feature fusion path to achieve the spatiotemporal information collaborative representation of surgical actions through multi-scale spatial feature extraction, inter-frame temporal dynamic modeling and cross-scale fusion.

[0013] The prediction head is replaced by the original shared branch prediction head of YOLOv11, which is a decoupled dual-branch prediction head (DDPH), including a parallel classification branch and a bounding box regression branch. These branches perform independent feature learning for the category recognition and location localization of the surgical action, respectively, thereby achieving decoupling optimization of the classification and localization tasks.

[0014] As a further description of the above technical solution, the ESPPF module includes two parallel branches: a convolutional branch and a pooling branch. The outputs of the two branches are fused through element-wise addition, and then refined by an additional convolutional block before being output. Each branch performs feature processing based on a convolutional block (CBL). The convolutional block (CBL) consists of a convolutional layer (Conv), batch normalization (BN), and a leaky rectified linear unit (L-ReLU) activation function, mathematically defined as:

[0015]

[0016] Where X is the input feature map.

[0017] As a further description of the above technical solution, the convolution branch adopts a serialized convolution operation, and the specific process is as follows:

[0018]

[0019]

[0020]

[0021]

[0022] Where k1 and k3 represent kernel sizes of 1×1 and 3×3 respectively, and s1 represents a stride of 1.

[0023] As a further description of the above technical solution, the pooling branch adopts a cascaded max pooling operation, and the specific process is as follows:

[0024]

[0025]

[0026] .

[0027] As a further description of the above technical solution, the outputs of the convolutional branch and the pooling branch are fused through element-wise addition to achieve the integration of spatial features and channel features:

[0028] ;

[0029] The fused features are then further processed and refined using additional convolutional blocks to obtain the final output features:

[0030]

[0031] in, For branch fusion features, This is the final output feature of the module.

[0032] As a further description of the above technical solution, the process of generating multi-scale spatial features in the ST-MCAM module is as follows:

[0033] Let the input feature map be Multi-scale representations are generated using 3×3 depthwise separable convolutions with a stride of s:

[0034] ,

[0035] Where C is the number of channels, H and W are the height and width of the feature map, respectively; and s is the scale factor.

[0036] As a further description of the above technical solution, the process of inter-frame temporal dynamic modeling of the ST-MCAM module is as follows:

[0037] Expand the feature maps specific to each scale into spatial tokens ,in H s W s These represent the height and width of the feature map at scale s, respectively.

[0038] Motion tokens are obtained by calculating inter-frame feature differences to capture temporal dynamic characteristics. ;

[0039] For each scale of spatial token, a query vector Q, a key vector K, and a value vector V are generated through a 1×1 convolution. Attention weights are then calculated. ,in, ;

[0040] The value vector is fused with the motion token to calculate the motion-aware value, thus integrating temporal motion information. ;

[0041] Calculate temporal context features using attention weights: ;

[0042] in, , Let represent the spatial tokens of frame t and frame t-1 at scale s, respectively, where d is the vector dimension.

[0043] As a further description of the above technical solution, the cross-scale fusion and gating optimization process of the ST-MCAM module is as follows:

[0044] Cross-scale context fusion is performed by using softmax-normalized learned gating parameters to control the weights of features at each scale: ,in , The gating parameters corresponding to scale s are obtained through training.

[0045] The merged tokens Restored to the spatial dimension, we obtain ;

[0046] Second-order covariance pooling is used to obtain discriminative channel correlation:

[0047]

[0048] in, , Representation of feature map The i-th spatial vector, μ is the mean vector calculated from all spatial locations;

[0049] The covariance is compressed into a gated vector through two learnable 1×1 convolutional layers and a nonlinear function:

[0050]

[0051] Where ρ is the ReLU activation function, σ is the Sigmoid activation function, and W1 and W2 are 1×1 convolution kernels;

[0052] By employing residual connections and using a learnable parameter β to ensure optimization stability and enhance feature representation capabilities, the final output features of the module are obtained:

[0053]

[0054] Here, ⊙ represents element-wise multiplication.

[0055] As a further description of the above technical solution, the specific structure and working principle of the DDPH prediction head are as follows:

[0056] Given an input feature map X and a scale parameter S, input them into the classification branch (Cls) and the bounding box regression branch (Box), respectively, and process the two branches in parallel and independently.

[0057] The classification branch employs a combination of convolution (Conv) and depthwise convolution (DWConv) to enhance spatial encoding performance. The specific process is as follows:

[0058]

[0059]

[0060]

[0061]

[0062]

[0063] The acquired feature maps are then connected via residual joins. To merge:

[0064]

[0065] The fused feature maps are then processed by additional convolutional layers to generate the final classification prediction result:

[0066]

[0067] Where Conv2d is the convolutional layer that outputs the classification results, and the output dimension corresponds to the number of categories of surgical actions; Y cls Output by category.

[0068] The bounding box regression branch directly applies three consecutive convolutional layers (Conv) to obtain accurate spatial localization and bounding box prediction results. The specific process is as follows:

[0069]

[0070]

[0071]

[0072] The final bounding box coordinates are generated using a dedicated convolutional layer:

[0073]

[0074]

[0075] Where Conv2d is the convolutional layer that outputs the bounding box coordinates, Y box This is the bounding box regression output.

[0076] Compared with the prior art, the present invention has the following beneficial effects:

[0077] This invention embeds an Enhanced Fast Spatial Pyramid Pooling (ESPPF) module into the backbone network, leveraging the synergistic effect of multi-branch convolution and pooling operations to simultaneously capture fine-grained local features and wide-area multi-scale spatial features of surgical actions. This effectively adapts to the scale variations of instruments and tissues in surgical scenarios, providing rich feature support for subsequent detection tasks. By inserting a Spatiotemporal Multi-Scale Context Aggregation (ST-MCAM) module into the neck network, it not only achieves cross-scale fusion of multi-scale spatial features but also fully utilizes the continuity and dynamic correlation of surgical actions through inter-frame motion token computation and temporal context modeling. This significantly improves the representation ability of complex temporal surgical actions and reduces the impact of interference factors such as occlusion and deformation. The invention employs a Decoupled Dual-Branch Prediction Head (DDPH) to separate the classification task and bounding box regression task into independent branches. Targeted feature extraction structures are used to optimize the two tasks separately, avoiding optimization conflicts between tasks caused by shared branches. At the same time, the combination of convolution and depthwise convolution in the classification branch and the continuous convolution design in the regression branch improve the accuracy of category recognition and the precision of location positioning, respectively, significantly improving the overall detection performance.

[0078] This invention is based on the YOLOv11 architecture with lightweight improvements. The computational complexity of each core module is controllable and does not significantly increase the model inference overhead. It can meet the real-time detection requirements in minimally invasive surgery scenarios and can be directly integrated into intelligent surgical assistance systems. It provides technical support for practical applications such as surgical navigation and operation standard monitoring, and has broad clinical application prospects. Attached Figure Description

[0079] Figure 1 This is an overall architecture diagram of the high-efficiency and high-precision surgical motion detection system based on the YOLOv11 architecture of this invention.

[0080] Figure 2 This is an architecture diagram of the Enhanced Fast Spatial Pyramid Pooling Module (ESPPF) in this invention.

[0081] Figure 3 This is an architecture diagram of the spatiotemporal multi-scale context aggregation module (ST-MCAM) in this invention.

[0082] Figure 4 This is an architecture diagram of the decoupled dual-branch prediction head (DDPH) in this invention. Detailed Implementation

[0083] The claims of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but this does not constitute any limitation on the present invention. Any limited modifications made by any person within the scope of protection of the claims of the present invention shall still be within the scope of protection of the claims of the present invention.

[0084] This high-efficiency, high-precision surgical action detection system is based on the YOLOv11 architecture. Improvements are made to the YOLOv11 architecture by adding three core enhancement modules: Enhanced Fast Spatial Pyramid Pooling Module (ESPPF), Spatiotemporal Multi-Scale Context Aggregation Module (ST-MCAM), and Decoupled Dual-Branch Prediction Head (DDPH), and optimizing the network flow. This achieves efficient and accurate detection of surgical actions. Its overall architecture is as follows: Figure 1 As shown, it mainly consists of three parts: the backbone network, the neck network, and the predicted head. Each part works together to complete the feature extraction, fusion, and detection of surgical actions.

[0085] The backbone network adopts the basic structure of YOLOv11, including standard convolutional layers (Conv) and C3K2 convolutional modules. An enhanced fast spatial pyramid pooling module (ESPPF) is embedded in the deep layers of the backbone network (after the C3K2 module and before the Neck network) to enhance multi-scale feature extraction. The Neck network contains a C2PSA attention module, which achieves multi-scale feature fusion through upsampling and skip connections. A spatiotemporal multi-scale context aggregation module (ST-MCAM) is inserted in the feature fusion path to achieve spatiotemporal information modeling. The prediction head adopts a decoupled dual-branch prediction head (DDPH) to classify and locate the fused features.

[0086] The workflow is as follows: Surgical video frame images enter the backbone network through the input layer, and initial features are extracted through the standard convolutional layer Conv. Subsequently, the C3K2 convolutional module progressively downsamples and extracts hierarchical features, and the ESPPF module captures multi-scale spatial features. The multi-scale features output by the backbone network enter the neck network, and the weights of key features are enhanced through the C2PSA attention module. The initial fusion of low-level and high-level features is achieved through upsampling and skip connections. The ST-MCAM module performs spatiotemporal information collaborative modeling on the fused features to further improve the feature representation capability. Finally, the spatiotemporally fused features are fed into the DDPH prediction head. The classification branch outputs the surgical action category, and the regression branch outputs the action location bounding box to complete the detection task.

[0087] The enhanced fast spatial pyramid pooling module (ESPPF) optimizes feature extraction by integrating multi-branch convolution and pooling operations. It can simultaneously capture fine-grained local features and wide-area contextual information, effectively improving the feature representation capabilities of downstream tasks such as object detection and semantic segmentation. Its specific architecture is as follows: Figure 2 As shown, it contains two main parallel branches: a convolution branch and a pooling branch, which are ultimately fused together through element-wise addition. Each convolutional block (CBL) consists of a convolutional layer (Conv), batch normalization (BN), and a leaky rectified linear unit (L-ReLU) activation function, mathematically defined as follows:

[0088]

[0089] Where X is the input feature map;

[0090] The convolution branch uses sequential convolution operations, passing through sequentially... , , The input feature map X is processed by a three-layer convolutional block to obtain refined local features, which are then fused with the original input features through residual connections.

[0091]

[0092]

[0093]

[0094] ;

[0095] The pooling branch employs cascaded max pooling operations, obtaining wide-area context features through three two-dimensional max pooling operations (MaxPool2d), namely:

[0096]

[0097]

[0098] .

[0099] Output the convolution branch With pooling branch output , , Multi-scale spatial features are obtained through element-wise additive fusion, thereby integrating spatial features with channel features:

[0100]

[0101] The fused features are then processed by additional convolutional blocks. Output after information integration and refinement:

[0102] .

[0103] The ESPPF module aggregates multi-scale features through convolutional blocks and max pooling operations, thereby enhancing the spatial context representation capability.

[0104] The aforementioned Spatiotemporal Multi-Scale Context Aggregation Module (ST-MCAM) captures multi-scale features and motion-aware features through parallel attention branches. These features are then processed through cross-scale fusion and covariance gate to achieve enhanced spatiotemporal representation capabilities, ultimately outputting optimized features. Its architecture is as follows: Figure 3 As shown, the specific workflow is as follows:

[0105] Multi-scale feature generation: Representing the input feature map as... Convolutions with a stride of s and a depth of 3×3 can be separable. Generate multi-scale representations:

[0106] , ;

[0107] Then, through a tokenization operation (Patchify), the feature maps for each scale are... Expand into a space token , ( H s W s (These represent the height and width of the feature map at scale s, respectively). Simultaneously, inter-frame feature differences are calculated to generate a simple motion token, thereby capturing temporal dynamic characteristics.

[0108] ;

[0109] Fusion of attention mechanisms and motion perception: spatial tokens at each scale The query vector Q, key vector K, and value vector V are generated through 1×1 convolution, and the attention weights are calculated:

[0110] , ;

[0111] The value vector is fused with the motion token to obtain the motion-aware value:

[0112] ;

[0113] Calculate temporal context features using attention weights:

[0114] ;

[0115] Cross-scale fusion: using softmax-normalized learned gating parameters Temporal context features at different scales Perform cross-scale context fusion:

[0116] ,in, ;

[0117] Channel-level covariance gating: This gates the merged tokens. Restored to the spatial dimension feature map, represented as .

[0118] Subsequently, a second-order covariance pooling method was used to obtain the discriminative channel correlation:

[0119] , ;

[0120] in, Representation of feature map The i-th spatial vector, μ is the mean vector calculated from all spatial locations.

[0121] The covariance is compressed into a gated vector through two learnable 1×1 convolutional layers and a nonlinear function:

[0122] ;

[0123] Where ρ is the ReLU activation function, σ is the Sigmoid activation function, and W1 and W2 are 1×1 convolution kernels.

[0124] Residual connections are used before the final output, and the final features are output through a learnable parameter β to ensure optimization stability and enhance feature representation capability:

[0125] .

[0126] The ST-MCAM module captures multi-scale features and motion-aware features through parallel attention branches, and then performs cross-scale fusion and covariance gating processing to ultimately achieve enhanced spatiotemporal representation capabilities.

[0127] The Decoupled Dual-Branch Prediction Head (DDPH) feeds the fused features output from the neck network into the classification branch and the bounding box regression branch, respectively, to independently determine the category and spatially locate the surgical action. This separate optimization improves detection accuracy and robustness. Its architecture is as follows: Figure 4As shown, given an input feature map X and a scale parameter S, the DDPH prediction head processes them independently into two parallel pathways: dedicated to classification (Cls) and bounding box regression (Box) tasks, respectively.

[0128] The classification branch uses alternating convolutional (Conv) and depthwise convolutional (DWConv) operations to enhance spatial encoding performance.

[0129]

[0130]

[0131]

[0132]

[0133]

[0134] The acquired feature maps are then first connected via residuals. To merge:

[0135]

[0136] The fused feature maps are then processed by additional convolutional layers to generate the final classification prediction result:

[0137]

[0138] Among them, Y cls Output by category.

[0139] The bounding box regression branch directly applies three consecutive convolutional layers (Conv) to obtain accurate spatial localization and bounding box prediction results. The relevant mathematical formulas are shown below:

[0140]

[0141]

[0142]

[0143] The final bounding box coordinates are generated using a dedicated convolutional layer:

[0144]

[0145]

[0146] Among them, Y box This is the bounding box regression output.

[0147] DDPH separates the classification and localization tasks into parallel branches by using depthwise convolution and standard convolution, thereby improving detection accuracy and task-specific learning capabilities.

[0148] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the present invention.

Claims

1. A high-efficiency and high-precision surgical action detection system based on the YOLOv11 architecture. This system is an improvement upon the YOLOv11 architecture, comprising three core components: a backbone network, a neck network, and a prediction head. The backbone network performs initial feature extraction and hierarchical feature enhancement on the input surgical video frames. Subsequently, the neck network performs multi-scale feature fusion and spatiotemporal information modeling on the extracted feature maps, integrating low-level detail features with high-level semantic features and inter-frame dynamic information. Finally, the prediction head performs category recognition and location localization of the surgical action based on the fused features, outputting the detection results. Its features are: The backbone network adopts the basic structure of YOLOv11, including standard convolutional layers and C3K2 convolutional modules. It extracts hierarchical features of the input surgical video frame images by progressive downsampling. The deep layers of the backbone network are embedded with enhanced fast spatial pyramid pooling modules. Through the synergistic effect of multi-branch convolution and pooling operations, it captures fine-grained local features and wide-area multi-scale spatial features of surgical actions. The neck network is based on the YOLOv11 multi-scale feature fusion path, which includes a C2PSA attention module. It achieves the initial fusion of low-level and high-level features through upsampling, skip connections and C3K2 convolution module reuse. A spatiotemporal multi-scale context aggregation module is inserted into the feature fusion path. Through multi-scale spatial feature generation, inter-frame temporal dynamic modeling, cross-scale fusion and gating optimization, it achieves the spatiotemporal information collaborative representation of surgical actions. The prediction head is replaced with a decoupled dual-branch prediction head, which includes a parallel classification branch and a bounding box regression branch. These branches perform independent feature learning for the category recognition and location localization of the surgical action, respectively, thereby achieving decoupling optimization of the classification and localization tasks.

2. The high-efficiency, high-precision surgical motion detection system based on the YOLOv11 architecture according to claim 1, characterized in that, The enhanced fast spatial pyramid pooling module includes two parallel branches: a convolution branch and a pooling branch. The outputs of the two branches are fused through element-wise addition, and then refined by an additional convolutional block before being output. Each branch performs feature processing based on a convolutional block, which consists of a convolutional layer, batch normalization, and a leaky rectified linear unit activation function, mathematically defined as: Where X is the input feature map.

3. The high-efficiency, high-precision surgical motion detection system based on the YOLOv11 architecture according to claim 2, characterized in that, The convolutional branch employs serialized convolution operations, the specific process of which is as follows: in, , These represent convolution kernel sizes of 1×1 and 3×3, respectively. This indicates a step size of 1.

4. The high-efficiency, high-precision surgical motion detection system based on the YOLOv11 architecture according to claim 3, characterized in that, The pooling branch uses a cascaded max pooling operation, and the specific process is as follows: 。 5. The high-efficiency, high-precision surgical action detection system based on the YOLOv11 architecture according to claim 4, characterized in that, The outputs of the convolutional and pooling branches are fused element-wise through addition to integrate spatial and channel features. The outputs are then refined by additional convolutional blocks. Specifically: in, For branch fusion features, This is the final output feature.

6. The high-efficiency, high-precision surgical motion detection system based on the YOLOv11 architecture according to claim 1, characterized in that, The multi-scale spatial feature generation process of the spatiotemporal multi-scale context aggregation module is as follows: Let the input feature map be Multi-scale representations are generated using 3×3 depthwise separable convolutions with a stride of s: , Where C is the number of channels, H and W are the height and width of the feature map, respectively; and s is the scale factor.

7. The high-efficiency, high-precision surgical motion detection system based on the YOLOv11 architecture according to claim 6, characterized in that, The inter-frame temporal dynamic modeling process of the spatiotemporal multi-scale context aggregation module is as follows: Feature maps specific to each scale Expand into a space token ,in H s W s These represent the height and width of the feature map at scale s, respectively. Motion tokens are obtained by calculating inter-frame feature differences to capture temporal dynamic characteristics. ; For each spatial token at each scale, a query vector Q, a key vector K, and a value vector V are generated through a 1×1 convolution, and attention weights are calculated: ,in ; The value vector is fused with the motion token to calculate the motion-aware value, thus integrating temporal motion information. ; Calculate temporal context features using attention weights: ; in, , Let represent the spatial tokens at scale s for frame t and frame t-1, respectively; d is the vector dimension.

8. The high-efficiency, high-precision surgical motion detection system based on the YOLOv11 architecture according to claim 7, characterized in that, The cross-scale fusion and gating optimization process of the spatiotemporal multi-scale context aggregation module is as follows: Cross-scale context fusion is performed by using softmax-normalized learned gating parameters to control the weights of features at each scale: ， ； The merged tokens Restored to the spatial dimension, we obtain ; Channel correlation is obtained using second-order covariance pooling: ， ； The covariance is compressed into a gated vector through two learnable 1×1 convolutional layers and a nonlinear function: By employing residual connections and using a learnable parameter β to ensure optimization stability and enhance feature representation capabilities, the final output features of the module are obtained: in, The gating parameters corresponding to scale s are obtained through training. Representation of feature map The i-th spatial vector, μ is the mean vector calculated from all spatial locations, ρ is the ReLU activation function, σ is the Sigmoid activation function, and W1 and W2 are 1×1 convolution kernels.

9. The high-efficiency, high-precision surgical motion detection system based on the YOLOv11 architecture according to claim 1, characterized in that: The classification branch of the decoupled dual-branch prediction head uses a combination of convolution and depthwise convolution operations. The specific process is as follows: 。 10. The high-efficiency, high-precision surgical motion detection system based on the YOLOv11 architecture according to claim 1, characterized in that: The bounding box regression branch of the decoupled dual-branch prediction head uses continuous convolutional layers to obtain accurate spatial localization and bounding box prediction results. The specific process is as follows: 。

Images