Intelligent robot target detection method in complex lawn scene

Abstract

The invention discloses an intelligent robot target detection method in a complex lawn scene. The method comprises the following steps: carrying out data preprocessing on an input image; performing feature extraction on the preprocessed image; constructing a feature pyramid; and step 4, converting information in the convolutional network into a final target detection result and outputting the final target detection result. The step 1 specifically comprises the following steps: firstly, carrying out image size adjustment, pixel value normalization and color space conversion on an input image; and then correcting an overexposure and low-illumination region in the image to a middle brightness range through a local gamma transformation technology. According to the method, high calculation efficiency is ensured, and interference of environmental noise on a detection result is effectively reduced.

Description

Technical Field

[0001] The invention relates to an intelligent robot target detection method in a complex lawn scene, belonging to the technical field of robots. Background Art

[0002] With the continuous advancement of intelligent surveillance technology, the demand for object detection in outdoor environments, particularly natural scenes like lawns, has grown significantly. However, current methods face challenges when dealing with complex backgrounds, such as lighting variations, grass occlusion, and environmental noise, which can affect accurate object recognition and tracking. Therefore, developing an algorithm that can efficiently and accurately detect objects in complex environments like lawns is crucial.

[0003] YOLOv5, the object detection algorithm proposed by Glenn Jocher and his team, uses CSPDarkNet53 as its backbone network and introduces a Focus module, which improves model performance by periodically extracting pixels from high-resolution images and reconstructing them into low-resolution images. It also uses Mosaic data augmentation technology to improve model performance. YOLOv5 optimizes anchor box selection by calculating the aspect ratio between the bounding box and the anchor point of the current layer. Although YOLOv5 performs well in object detection, it still has room for improvement in resource consumption, small object detection accuracy, and model size. In complex scenes, the algorithm is easily affected by the surrounding environment, increasing the false positive and false detection rates. Achieving high-precision recognition typically requires a more complex computational model, which not only increases the computational burden but also makes real-time processing more difficult when computing power is limited.

[0004] Existing target detection technologies mainly face the following challenges: high computational complexity requires the deployed equipment to have powerful computing capabilities; the larger model size increases the demand for storage space and memory resources; the model is highly dependent on specific conditions and lacks robustness in complex environments; and recognition accuracy needs to be improved in the presence of occlusion or uneven lighting. Summary of the Invention

[0005] The purpose of the present invention is to overcome the problems existing in the prior art and provide an intelligent robot target detection method in complex lawn scenes, which integrates advanced lightweight noise reduction algorithms and backbone networks to effectively reduce the interference of environmental noise on detection results while ensuring high computational efficiency.

[0006] To solve the above technical problems, the present invention provides an intelligent robot target detection method in a complex lawn scene, comprising the following steps: Step 1: Preprocess the input image data; Step 2: Extract features from the preprocessed image; Step 3: Feature pyramid construction; Step 4: Convert the information in the convolutional network into the final target detection result and output it.

[0007] Furthermore, step 1 specifically includes the following steps: first, the input image is resized, pixel values are normalized, and color space is converted; then, overexposed and low-light areas in the image are corrected to a medium brightness range through a local gamma transform technique.

[0008] Furthermore, step 2 specifically includes the following steps: Step 2.1: Perform 1×1 convolution on the image to increase the dimension, and then use 3×3 convolution for feature extraction; Step 2.2: Process the image obtained in step 2.1 by setting multiple depth-wise separable convolutions and output the feature matrix; Step 2.3: Introduce the SE module and perform pooling processing on each channel of the feature matrix output in step 2.2 to obtain a weight vector. The number of weight vectors is equal to the number of channels. Step 2.4: The weight vector passes through two fully connected layers to obtain the final weight vector. The number of vectors in the first fully connected layer is one-quarter the number of channels, and the output of the second fully connected layer is the same as the number of channels. The first fully connected layer uses the ReLu activation function, and the second fully connected layer uses the h-swish activation function. Step 2.5: Perform 1×1 convolution to achieve dimensionality reduction, and use a linear bottleneck structure to avoid nonlinear destruction of information in the low-dimensional space.

[0009] Furthermore, the C3 model in YOLOv5 is introduced into CA attention to obtain the improved C3_CA model. Step 3 specifically includes the following two steps: 1) Spatial information embedding: The input feature map is average pooled channel by channel in the height direction and width direction to obtain the height feature map and width feature maps , the output formula can be expressed as: Where W represents the width of the feature map, H represents the height of the feature map, w represents the width of the feature pixel, h represents the height of the feature pixel, i represents the column number of the feature pixel, and j represents the row number of the feature pixel; Spatial Attention Generation: 1) Obtaining Height Feature Map and width feature maps After that, the feature maps are spliced according to the spatial dimension. 2) The spliced feature maps are subjected to convolution transformation and nonlinear activation function to generate intermediate feature maps. ,3) The intermediate feature map Split into wide tensors by spatial dimension and high tensors And respectively through the convolution transformation function and Perform dimensionality increase operation and combine it with Sigmoid activation function to get the attention vector and , and finally multiply the sum by the original input to get an output of the same shape as the original input. The final calculation formula is as follows: Where: i represents the column number of the feature pixel, j represents the row number of the feature pixel, and the generated spatial attention is generated.

[0010] Furthermore, step 4 specifically includes the following steps: Step 4.1: The extracted feature map is subjected to a convolution to reduce the number of channels and scale the feature map; Step 4.2: Fuse the feature maps at different levels, use the anchor frame generation mechanism to preset a series of anchor points on the feature maps at different scales, and use these anchor points as the initial reference frame for object recognition; Step 4.3: Fine-tune each anchor point to adapt to the size and position of the real object through bounding box regression analysis, and perform category prediction on the anchor point to assign an object category to each anchor point; Step 4.4: Remove redundant detection boxes through non-maximum suppression to ensure that the final output is the best detection result; Step 4.5: Present the detected objects in the form of bounding boxes, each of which contains the object category and confidence information.

[0011] Furthermore, the convolution described in step 4.1 includes two types: The first one: GConv: First, the ordinary convolution layer Conv is combined with the depth-wise separable convolution layer, and then the channel random mixing strategy is used to infiltrate the feature information generated by Conv into each module generated by the depth-wise separable convolution; The second type: OConv: 1) shrinks the input features into a feature vector of length through global average pooling; 2) uses a fully connected layer and ReLU function to map the feature vector to a low-dimensional space to reduce the complexity of dynamic convolution; 3) constructs four branches, each of which has an output size of , , , The FC layer of the network is constructed, where the first three branches pass through the Sigmoid activation function to map the variable values of the feature vectors of the three dimensions of space, input channel, and output channel to between (0,1). The fourth branch passes through the Softmax activation function to predict the weight of the convolution kernel. The variable values of the eigenvector are mapped to (0,1) so that the sum of the four branch variable values is 1.

[0012] Compared with the existing technology, the present invention has achieved the following beneficial effects: (1) Using the local gamma transform technology, the distribution characteristics of illumination are optimized while reducing the image noise level, thereby improving the quality of the input image; (2) A feature extraction module composed of a large number of depth-wise separable convolutions is proposed, which can significantly reduce the overall computational complexity of the algorithm and greatly improve the algorithm's operational efficiency; (3) Based on the C3 module of YOLOv5, the CA attention mechanism is integrated to further improve the feature extraction capability of the algorithm; (4) A GConv structure in which convolution and depth-wise separable convolution are mutually infiltrated is proposed to improve the algorithm's receptive field size and semantic understanding capability; (5) An OConv structure is proposed to refine the local features of the target, accurately locate the target's characteristic target area, and then extract the significant performance features of different targets, thereby improving the accuracy of the target detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The accompanying drawings are only provided for reference and illustration and are not intended to limit the present invention.

[0014] Figure 1 Schematic diagram of the feature extraction module in the present invention; Figure 2 Schematic diagram of the structure of the GConv network in the present invention; Figure 3 Schematic diagram of the algorithm network in the present invention. DETAILED DESCRIPTION

[0015] In the following description of the present invention, the terms "up", "down", "front", "back", "left", "right", "inside", "outside", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not mean that the device must have a specific direction.

[0016] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below with reference to specific illustrations.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0018] This paper aims to provide an intelligent robot target detection method in complex lawn scenes. It integrates advanced lightweight noise reduction algorithms and backbone networks to effectively reduce the interference of environmental noise on detection results while ensuring high computational efficiency. The attention mechanism enhances the contextual information around the target, significantly improving the recognition accuracy of the target. Figure 1 As shown, the following steps are included: Step 1: Preprocess the input image data; Step 1 specifically involves resizing the input image, normalizing pixel values, and converting its color space. Then, using local gamma transformation, it corrects overexposed and underlit areas of the image toward a more neutral brightness range. This process optimizes the lighting distribution of the image, improving overall visual quality and ensuring image quality and subsequent analysis accuracy even in extreme lighting conditions.

[0019] Step 2: Extract features from the pre-processed image. Specifically, a new backbone network is used for feature extraction, which introduces a large number of depth-wise separable convolutions to reduce the number of model parameters and computational complexity, making the network more lightweight. This includes the use of an inverse residual structure to alleviate the degradation problem during deep network training. Considering the search space of network structure and hardware characteristics, the neural architecture search technology is used to optimize model performance. The feature extraction network reference Figure 2 .

[0020] Step 2 specifically includes the following steps: Step 2.1: Perform 1×1 convolution on the image to increase the dimension, and then use 3×3 convolution for feature extraction; Step 2.2: Process the image obtained in step 2.1 by setting multiple depth-wise separable convolutions and output the feature matrix; Step 2.3: Introduce the SE module and perform pooling on each channel of the feature matrix output in step 2.2 to obtain a weight vector. The number of weight vectors obtained is equal to the number of channels. Step 2.4: The weight vector passes through two fully connected layers to obtain the output vector, which is the final weight vector. The number of vectors in the first fully connected layer is one-quarter the number of channels, and the output of the second fully connected layer is the same as the number of channels. The first fully connected layer uses the ReLu activation function, and the second fully connected layer uses the h-swish activation function. Step 2.5: Perform 1×1 convolution to achieve dimensionality reduction, and use a linear bottleneck structure to avoid nonlinear destruction of information in the low-dimensional space.

[0021] Step 3: Feature pyramid construction.

[0022] In order to eliminate the interference of background on model detection and improve feature extraction capability and detection effect, the C3 model in YOLOv5 is introduced into CA attention to obtain the improved C3_CA model. Step 3 specifically includes the following two steps: 1) Spatial information embedding: The input feature map is average pooled channel by channel in the height direction and width direction to obtain the height feature map and width feature maps , the output formula can be expressed as: Where W represents the width of the feature map, H represents the height of the feature map, w represents the width of the feature pixel, h represents the height of the feature pixel, i represents the column number of the feature pixel, and j represents the row number of the feature pixel. The spatial information embedding operation can avoid compressing all spatial information into the channel.

[0023] The resulting pair of method-aware feature maps enables CA attention to capture long-range dependencies within a single channel and also helps preserve precise location information, enabling the network to more accurately locate objects.

[0024] Spatial Attention Generation: 1) Obtaining Height Feature Map and width feature maps After that, the feature maps are spliced according to the spatial dimension. 2) The spliced feature maps are subjected to convolution transformation and nonlinear activation function to generate intermediate feature maps. ,3) The intermediate feature map Split into wide tensors by spatial dimension and high tensors And respectively through the convolution transformation function and Perform dimensionality increase operation and combine it with Sigmoid activation function to get the attention vector and Finally, and Multiplying with the original input, we get an output of the same shape as the original input. The final calculation formula is as follows: Where: i represents the column number of the feature pixel, j represents the row number of the feature pixel, and the generated spatial attention is generated.

[0025] Step 4: Convert the information in the convolutional network into the final target detection result and output it.

[0026] Step 4 specifically includes the following steps: Step 4.1: The extracted feature map is subjected to a convolution to reduce the number of channels and scale the feature map; Step 4.2: Fuse the feature maps at different levels, use the anchor frame generation mechanism to preset a series of anchor points on the feature maps at different scales, and use these anchor points as the initial reference frame for object recognition; Step 4.3: Through bounding box regression analysis, the entire model fine-tunes each anchor point to adapt to the size and position of the real object, and at the same time predicts the category of the anchor point and assigns an object category to each anchor point; Step 4.4: Remove redundant detection boxes through non-maximum suppression to ensure that the final output is the best detection result; Step 4.5: The entire model presents the detected targets in the form of bounding boxes, each of which contains the category and confidence information of the object.

[0027] like Figure 3 As shown, for dimensionality reduction and decompression, the convolution in step 4.1 includes two types: The first type of GConv: first combines the ordinary convolution layer Conv and the depth-wise separable convolution layer, and then uses the channel random mixing strategy to infiltrate the feature information generated by Conv into each module generated by the depth-wise separable convolution; The second type of OConv: 1) shrinks the input features into a feature vector of length through global average pooling; 2) uses a fully connected layer and ReLU function to map the feature vector to a low-dimensional space to reduce the complexity of dynamic convolution; 3) constructs four branches, each of which has an output size of , , , The FC layer of the network is constructed, where the first three branches pass through the Sigmoid activation function to map the variable values of the feature vectors of the three dimensions of space, input channel, and output channel to between (0,1). The fourth branch passes through the Softmax activation function to predict the weight of the convolution kernel. The variable values of the eigenvector are mapped to (0,1) so that the sum of the four branch variable values is 1.

[0028] The present invention uses local gamma transform technology to reduce the image noise level while optimizing the distribution characteristics of illumination and improving the quality of the input image. It can significantly reduce the overall computational complexity of the algorithm and greatly improve the algorithm's operational efficiency. Based on the C3 module of YOLOv5, the CA attention mechanism is integrated to further improve the algorithm's feature extraction capabilities. A GConv structure that mutually penetrates convolution and depthwise separable convolution is proposed to improve the algorithm's receptive field size and semantic understanding capabilities. An OConv structure is proposed to refine the local features of the target, accurately locate the target's characteristic target area, and then extract the significant performance features of different targets, thereby improving the accuracy of the target detection model.

[0029] The above is only a preferred embodiment of the present invention, which shows and describes the basic principles and main features of the present invention and the advantages of the present invention, but does not limit the scope of patent protection of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. In addition to the above embodiments, the present invention may also have other implementation modes without departing from the spirit and scope of the present invention. The present invention may also have various changes and improvements, and all technical solutions formed by equivalent replacement or equivalent transformation fall within the scope of protection required by the present invention. The scope of protection required by the present invention is defined by the attached claims and their equivalents. Technical features not described in the present invention can be achieved by or using existing technologies, and will not be described here.

Claims

1. A method for detecting targets using an intelligent robot in a complex lawn scene, characterized in that: The following steps are involved: Step 1: Preprocess the input image data; Step 2: Extract features from the preprocessed image; Step 3 : Feature pyramid construction; Step 4: Convert the information in the convolutional network into the final target detection result and output it.

2. The intelligent robot target detection method in a complex lawn scene according to claim 1 is characterized in that: Step 1 specifically includes the following steps: first, the input image is resized, pixel values are normalized, and color space is converted; then, the overexposed and low-light areas in the image are corrected to the middle brightness range through local gamma transformation technology.

3. The intelligent robot target detection method in a complex lawn scene according to claim 1 is characterized in that: Step 2 specifically includes the following steps: Step 2.1: Perform 1×1 convolution on the image to increase the dimension, and then use 3×3 convolution for feature extraction; Step 2.2: Process the image obtained in step 2.1 by setting multiple depth-wise separable convolutions and output the feature matrix; Step 2.3: Introduce the SE module and perform pooling processing on each channel of the feature matrix output in step 2.2 to obtain a weight vector. The number of weight vectors is equal to the number of channels. Step 2.4: The weight vector passes through two fully connected layers to obtain the final weight vector. The number of vectors in the first fully connected layer is one-quarter the number of channels, and the output of the second fully connected layer is the same as the number of channels. The first fully connected layer uses the ReLu activation function, and the second fully connected layer uses the h-swish activation function. Step 2.5: Perform 1×1 convolution to achieve dimensionality reduction, and use a linear bottleneck structure to avoid nonlinear destruction of information in the low-dimensional space.

4. The intelligent robot target detection method in a complex lawn scene according to claim 2 is characterized in that: The C3 model in YOLOv5 is introduced into the CA attention to obtain the improved C3_CA model. Step 3 specifically includes the following two steps: Step 3.1: Spatial information embedding: The input feature map is average pooled channel by channel in the height direction and width direction to obtain the height feature map and width feature maps , the output formula can be expressed as: Where W represents the width of the feature map, H represents the height of the feature map, w represents the width of the feature pixel, h represents the height of the feature pixel, i represents the column number of the feature pixel, and j represents the row number of the feature pixel; Step 3.2: Spatial Attention Generation: 1) Obtaining Height Feature Map and width feature maps After that, the feature maps are spliced according to the spatial dimension. 2) The spliced feature maps are subjected to convolution transformation and nonlinear activation function to generate intermediate feature maps. ,3) The intermediate feature map Split into wide tensors by spatial dimension and high tensors And respectively through the convolution transformation function and Perform dimensionality increase operation and combine it with Sigmoid activation function to get the attention vector and ; Finally, and Multiplying with the original input, we get an output of the same shape as the original input. The final calculation formula is as follows: Where: i represents the column number of the feature pixel, j represents the row number of the feature pixel, and the generated spatial attention is generated.

5. The intelligent robot target detection method in a complex lawn scene according to claim 4 is characterized in that: Step 4 specifically includes the following steps: Step 4.1: The extracted feature map is subjected to a convolution to reduce the number of channels and scale the feature map; Step 4.2: Fuse the feature maps at different levels, use the anchor frame generation mechanism to preset a series of anchor points on the feature maps at different scales, and use these anchor points as the initial reference frame for object recognition; Step 4.3: Fine-tune each anchor point to adapt to the size and position of the real object through bounding box regression analysis, and perform category prediction on the anchor point to assign an object category to each anchor point; Step 4.4: Remove redundant detection boxes through non-maximum suppression to ensure that the final output is the best detection result; Step 4.5: Present the detected objects in the form of bounding boxes, each of which contains the object category and confidence information.

6. The intelligent robot target detection method in a complex lawn scene according to claim 5 is characterized by: There are two types of convolution described in step 4.1: The first type of GConv: first combines the ordinary convolution layer Conv and the depth-wise separable convolution layer, and then uses the channel random mixing strategy to infiltrate the feature information generated by Conv into each module generated by the depth-wise separable convolution; The second type of OConv: 1) shrinks the input features into a feature vector of length through global average pooling; 2) uses a fully connected layer and ReLU function to map the feature vector to a low-dimensional space to reduce the complexity of dynamic convolution; 3) constructs four branches, each of which has an output size of , , , The FC layer of the network is constructed, where the first three branches pass through the Sigmoid activation function to map the variable values of the feature vectors of the three dimensions of space, input channel, and output channel to between (0,1). The fourth branch passes through the Softmax activation function to predict the weight of the convolution kernel. The variable values of the eigenvector are mapped to (0,1) so that the sum of the four branch variable values is 1.

Images