Intelligent monitoring method for wading shoreside fishing behavior based on unmanned ship

Abstract

The invention discloses an intelligent monitoring method for wading shoreside fishing behaviors based on an unmanned ship. The intelligent monitoring method comprises the following steps: step 1, acquiring an image of a wading river and lakeshore zone based on the unmanned ship; step 2, preprocessing the acquired image; step 3, collecting the preprocessed image data set, and labeling each image to form a labeled image data set; step 4, constructing a YOLOv5 image recognition model; 5, taking the preprocessed image data set and the labeled image data set as training data to train the model; step 6, evaluating and optimizing the model; and step 7, performing wading shoreside fishing behavior monitoring on a newly acquired unknown image by using the trained and optimized model. According to the method, the actual wading shoreside fishing monitoring requirement is effectively combined with the unmanned ship image acquisition capability, the wading shoreside fishing behavior is monitored through the YOLOv5 image recognition model, the accuracy is high, and the method has great significance in maintaining the ecological health of rivers and lakes.

Description

Technical Field

[0001] The present invention belongs to the technical field of river and lake ecological protection, and in particular relates to an intelligent monitoring method for wading shore fishing behavior based on an unmanned boat. Background Art

[0002] Controlling fishing behavior near the shore is an integral part of the River and Lake Chief System. By restricting fishing behavior near the shore, it can provide protection for people's life safety and the ecological health of rivers and lakes. Unmanned boats have received extensive attention and adaptation in supporting water conservancy work such as the River and Lake Chief System. They have the capabilities of hydrological monitoring, water quality monitoring, hydrodynamic monitoring, image acquisition, and autonomous action. They can carry out necessary routine monitoring tasks in areas with fewer water conservancy workers, and obtain necessary monitoring data in a timely and effective manner, providing important support for water conservancy work.

[0003] Traditional control of wading shore fishing behavior requires a large number of river patrol personnel to conduct field surveys. On the one hand, traditional methods have high requirements for human resources. Some river sections, especially in wild areas, have weak manpower and are difficult to support high-intensity river patrols. On the other hand, the monitoring frequency of traditional methods is limited. Considering the round-trip distance and time cost of river patrol personnel, the monitoring frequency of most river sections is less than once a day, which is difficult to meet the normal needs of wading shore monitoring. In summary, the traditional wading shore fishing behavior monitoring method has a low degree of intelligence, high labor cost requirements, and limited monitoring efforts. Therefore, there is an urgent need for an intelligent monitoring method for wading shore fishing behavior based on unmanned boats. Summary of the invention

[0004] The purpose of the present invention is to provide an intelligent monitoring method for wading shore fishing behavior based on an unmanned boat to solve the above-mentioned technical problems.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The present invention discloses an intelligent monitoring method for wading shore fishing behavior based on an unmanned boat, the method comprising the following steps:

[0007] Step 1: Image acquisition: Based on the high-definition imaging equipment carried by the unmanned boat, the image of the shore zone of the wading river and lake is acquired by controlling the angle and width of the equipment;

[0008] Step 2: Image preprocessing: preprocess the acquired image, including adjusting the image size, normalizing the image pixel value, and enhancing the data;

[0009] Step 3, image annotation: collect the preprocessed image data set, annotate each image, and the annotation content includes the location and category of wading shore fishing behavior, including the location and type of prohibited actions of personnel in the management area, to form an annotated image data set;

[0010] Step 4: Build a YOLOv5 image recognition model: Build a YOLOv5 image recognition model based on YOLOv5;

[0011] Step 5, model training: The preprocessed image dataset and the annotated image dataset are used as training data to train the YOLOv5 image recognition model. The model input is the preprocessed image dataset, and the output is the corresponding annotated image dataset. The training data is divided into a training set and a test set in proportion.

[0012] Step 6: Model evaluation and optimization: Use the test set to evaluate the trained YOLOv5 image recognition model, calculate the accuracy, precision, and recall indicators of wading shore fishing behavior monitoring; and optimize and adjust the model based on the model evaluation results;

[0013] Step 7: Monitoring fishing behavior by wading on the shore: Use the trained and optimized YOLOv5 image recognition model to monitor the newly acquired unknown images for fishing behavior by wading on the shore. If the image contains fishing behavior, the model will identify the location and category of the fishing behavior by wading on the shore in the image, draw a bounding box to mark the fishing behavior by wading on the shore, and output the marked image.

[0014] Furthermore, the acquisition of the image of the shore zone of the wading river and lake described in step 1 follows the following principles:

[0015] (1) Number of images: During the operation of the unmanned ship, at least 300-500 images are collected at each monitoring point;

[0016] (2) Image quality: The image resolution is no less than 1920x1080 pixels, and the images are collected under sunny, cloudy, and rainy weather conditions, and the image color depth reaches 24-bit true color;

[0017] (3) Collection angle: Collect images from multiple angles, including the front and side, to obtain images with a comprehensive viewing angle range;

[0018] (4) Lighting conditions: images are collected under different lighting conditions, including sunny days, cloudy days, heavy fog, and dusk;

[0019] (5) Data format: The acquired images are saved in a lossless format to prevent information loss caused by data compression; the image metadata is fully preserved, including shooting time, GPS location, and device parameters.

[0020] Furthermore, the image resizing described in step 2 is specifically as follows: resizing all images, the size standard of the input model is 640x640 pixels, and when resizing, a bilinear interpolation method is used to maintain the clarity of the image, and gray pixels are added to fill the edges of the image to maintain the original aspect ratio;

[0021] The normalized image pixel value is specifically: calculating the mean and standard deviation of the data set collected by the unmanned ship, using these as standard parameters, and then subtracting the mean from each pixel value of the image and dividing it by the standard deviation;

[0022] The enhanced data specifically includes: performing random rotation, translation, scaling, flipping and color dithering operations on the images collected by the unmanned ship to increase the diversity of the data.

[0023] Furthermore, the step 3 of marking the location of the wading shore fishing behavior is specifically as follows: using a bounding box to mark the location of the wading shore fishing behavior, and converting the bounding box of the wading shore fishing behavior into a standard format.

[0024] Furthermore, the specific process of constructing the YOLOv5 image recognition model in step 4 is: using the YOLOv5 feature extraction network to obtain feature information of the target as an appearance descriptor;

[0025] Feature extraction module of the YOLOv5 backbone network: The Conv convolution layer first extracts features, then sends them to the BN batch normalization layer for unified processing to speed up network learning, and then the SiLU activation function layer retains and maps the features and removes redundant features;

[0026] Feature fusion module: The neck uses FPN and PAN structures to fuse features at multiple scales. The FPN structure samples from top to bottom, and the PAN structure samples from bottom to top, so that feature maps of different sizes contain strong semantic information and strong feature information, thereby enhancing the ability to extract target features of different sizes.

[0027] After the main part and neck of the YOLOv5 detection network extract and fuse the features, three enhanced effective feature layers are obtained. The feature map sizes of these three enhanced features are (20, 20, 512), (40, 40, 256), and (80, 80, 128), respectively. At this time, the feature map is regarded as a set of feature points with channel number features. Considering that the targets to be tracked in the unmanned ship video sequence are small and dense, the enhanced feature map of size (80, 80, 128) is selected as the feature extraction result in the DeepSORT target tracking network. YOLOv5 is responsible for detecting objects in images or videos, while DeepSORT performs subsequent target tracking based on the detection results. The position of the object is predicted and matched by Kalman filtering and Hungarian algorithm. The feature map is followed by a fully connected layer and an L2 normalization layer to refit the appearance feature information of the target to obtain a 128-dimensional feature vector; then the obtained appearance feature vector is measured by cosine similarity and linearly weighted with the Mahalanobis distance metric, and the obtained cost formula is used as the metric of the Hungarian algorithm to obtain the best match:

[0028] d^(*(Cos))(ij)=min{|1-r_j^Tr_k^((i))|r_k^((i))∈R_i^*} (1)

[0029] c_(i,j)^*=λd^((Mah))(i,j)+(1-λ)d^(*(Cos))(ij) (2)

[0030] Where d^(*(Cos))(ij) represents the cosine distance metric used as the appearance feature vector of R_i^*; c_(I,j)^* represents the Hungarian algorithm cost formula obtained by combining the cosine distance metric with the Mahalanobis distance metric; r_j represents the appearance descriptor corresponding to the j-th detection box d_j; r_k^((i)) represents the appearance descriptor corresponding to the i-th trajectory; d^((Mah))(i,j) represents the Mahalanobis distance between the j-th detection box and the predicted bounding box of the i-th trajectory.

[0031] Furthermore, the step 5 of dividing the training data into a training set and a test set in proportion is specifically: dividing the training data into a training set and a test set in a ratio of 3:1.

[0032] Furthermore, the accuracy described in step 6 is: reflects the ability of the model to correctly judge the overall sample, and the calculation formula is:

[0033] Accuracy =(TP+TN) / (TP+TN+FP+FN) (3)

[0034] Where Accuracy is the accuracy; TP is the number of samples correctly predicted as positive; TN is the number of samples correctly predicted as negative; FP is the number of samples incorrectly predicted as positive; FN is the number of samples incorrectly predicted as negative.

[0035] The precision rate is: a measure of how many of the samples predicted as positive are truly positive, that is, it reflects the purity of the model in identifying positive samples. The larger the value, the better. The calculation formula is:

[0036] Precision = TP / (TP+FP) (4)

[0037] In the formula, Precision is the accuracy;

[0038] The recall rate is: a measure of how many samples are correctly identified as positive among all samples that are actually positive, that is, it reflects the coverage of the model for positive samples. The larger the value, the better. The calculation formula is:

[0039] Recall = TP / (TP+FN) (5)

[0040] In the formula, Recall is the recall rate.

[0041] The beneficial effect of the present invention is that the intelligent monitoring method of wading shore fishing behavior based on an unmanned boat described in the present invention effectively combines the actual needs of wading shore fishing monitoring with the image acquisition capability of the unmanned boat, and monitors wading shore fishing behavior through the YOLOv5 image recognition model with high accuracy and strong practicality. It can support the development of related work of the river and lake chief system, and is of great significance for protecting people's lives and maintaining the ecological health of rivers and lakes.

[0042] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of the method of the present invention;

[0044] Figure 2 Annotate images with schematics;

[0045] Figure 3 This is a schematic diagram of the overall structure of YOLOv5;

[0046] Figure 4 Schematic diagram of the YOLOv5 network structure. DETAILED DESCRIPTION

[0047] The present invention discloses an intelligent monitoring method for wading shore fishing behavior based on an unmanned boat, such as Figure 1 As shown, the method comprises the following steps:

[0048] Step 1: Image acquisition: Based on the high-definition imaging equipment carried by the unmanned boat, the image of the river and lake shore zone is acquired by controlling the angle and width of the equipment. Image acquisition should follow the following principles:

[0049] (1) Number of images: During the operation of the unmanned boat, at least 300 to 500 images must be collected at each monitoring point to cover the perspectives and behavioral characteristics of different images and ensure the diversity and representativeness of the data.

[0050] (2) Image quality: The image resolution should be no less than 1920x1080 pixels to ensure image clarity. To adapt to different lighting conditions and environmental backgrounds, images should be collected under various weather conditions such as sunny, cloudy, and rainy days. The color depth of the image should reach 24-bit true color to ensure the accuracy of color information and help the model recognize different objects and behaviors.

[0051] (3) Acquisition angle: Acquire images from multiple angles (such as front, side, etc.) to obtain images with a comprehensive viewing angle range.

[0052] (4) Lighting conditions: Collect images under different lighting conditions (such as sunny, cloudy, foggy, dusk, etc.) to increase data diversity and enhance the generalization ability of the model.

[0053] (5) Data format: The acquired images need to be saved in a lossless format (such as PNG, TIFF) to prevent information loss caused by data compression. The image metadata (such as shooting time, GPS location, equipment parameters, etc.) should be fully preserved to facilitate subsequent data analysis and traceability.

[0054] Step 2: Image preprocessing: Preprocess the acquired river and lake shore images to meet the requirements of image recognition. Preprocessing mainly includes adjusting image size, normalizing image pixel values, and enhancing data.

[0055] Specifically, adjust the image size: all images are resized, and the standard size of the input model is 640x640 pixels. When adjusting, the bilinear interpolation method is used to maintain the clarity of the image, and the original aspect ratio is maintained by adding gray pixels to fill the edges of the image.

[0056] Normalized image pixel values: used to accelerate the convergence speed of the model, reduce the brightness contrast difference between different images, improve training stability and ensure the input consistency during model training. The specific processing flow is: calculate the mean ([0.485, 0.456, 0.406]) and standard deviation ([0.229, 0.224, 0.225]) of the data set collected by the unmanned ship, use these as standard parameters, and then subtract the mean from each pixel value of the image and divide it by the standard deviation.

[0057] Enhanced data: In order to enhance the generalization ability of the model and cope with the complexity of different environments, data enhancement technology was used in this training process. The images collected by the unmanned ship were randomly rotated, translated, scaled, flipped, and color jittered, thereby effectively increasing the diversity of the data and improving the robustness of the model to various interference factors.

[0058] Step 3, image annotation: collect the preprocessed image dataset, annotate each image, and annotate the location and category of fishing behavior near the shore, including the location, prohibited action type of personnel in the management area, etc., to form an annotated image dataset. The above annotation uses a bounding box to mark the location of fishing behavior near the shore, such as Figure 2 As shown, the bounding box of the wading shore fishing behavior is converted into a standard format, such as normalized coordinates relative to the image size.

[0059] Step 4: Build a YOLOv5 image recognition model: Build a YOLOv5 image recognition model based on YOLOv5. YOLOv5 is an efficient real-time target detection algorithm with excellent performance and speed. It uses the YOLOv5 feature extraction network to obtain the target's feature information as an appearance descriptor, such as Figure 3 The feature extraction module of the YOLOv5 backbone network and the feature fusion module of the neck are shown in Figure 4 shown.

[0060] Feature extraction module: The Conv convolution layer first extracts features, then sends them to the BN batch normalization layer for unified processing to speed up network learning, and then the SiLU activation function layer retains and maps the features and removes redundant features.

[0061] Feature fusion module: The neck uses FPN and PAN structures to fuse features at multiple scales. The FPN structure samples from top to bottom, and the PAN structure samples from bottom to top, so that feature maps of different sizes contain strong semantic information and strong feature information, thereby enhancing the ability to extract target features of different sizes.

[0062] After the main part and neck of the YOLOv5 detection network extract and fuse the features, three enhanced effective feature layers can be obtained. The feature map sizes of these three enhanced features are (20,20,512), (40,40,256), and (80,80,128), respectively. At this time, the feature map can be regarded as a set of feature points with channel number features. Considering that the targets to be tracked in the unmanned ship video sequence are small and dense, the enhanced feature map with a size of (80,80,128) is selected as the feature extraction result in the DeepSORT target tracking network. YOLOv5 is responsible for detecting objects in images or videos, while DeepSORT performs subsequent target tracking based on the detection results, and predicts and matches the position of the object through Kalman filtering and Hungarian algorithm. The feature map is followed by a fully connected layer and an L2 normalization layer to refit the appearance feature information of the target to obtain a 128-dimensional feature vector. The obtained appearance feature vector is then measured by cosine similarity and linearly weighted with the Mahalanobis distance metric, and the resulting cost formula is used as the metric of the Hungarian algorithm to obtain the best match:

[0063] d^(*(Cos))(ij)=min{|1-r_j^Tr_k^((i))|r_k^((i))∈R_i^*} (1)

[0064] c_(i,j)^*=λd^((Mah))(i,j)+(1-λ)d^(*(Cos))(ij) (2)

[0065] Where d^(*(Cos))(ij) represents the cosine distance metric used as the appearance feature vector of R_i^*; c_(I,j)^* represents the Hungarian algorithm cost formula obtained by combining the cosine distance metric with the Mahalanobis distance metric; r_j represents the appearance descriptor corresponding to the j-th detection box d_j; r_k^((i)) represents the appearance descriptor corresponding to the i-th trajectory; d^((Mah))(i,j) represents the Mahalanobis distance between the j-th detection box and the predicted bounding box of the i-th trajectory.

[0066] Step 5, model training: Use the preprocessed image dataset and the annotated image dataset as training data to train the YOLOv5 image recognition model. The model input is the preprocessed image dataset, and the output is the corresponding annotated image dataset. The training data is divided into training set and test set at a ratio of 3:1. Both training and testing are performed on the Ubuntu 18.04 operating system. The specific configuration is shown in the following table:

[0067] Table 1

[0068] Parameter name Configuration operating system Ubuntu 18.04 GPU NVIDIA GeForce GTX 1060 6GB GPU Accelerated Environment CUDA v11.4 Pytorch Version 1.10.2 Python Version 3.7.11

[0069] Since the YOLOv5 image recognition model is relatively concise, has low computational overhead, and is more suitable for loading on a mobile terminal, the present invention uses the Ultralytics 6.0 version of the YOLOv5s model for training and testing, respectively. The specific configuration is shown in the following table:

[0070] Table 2

[0071]

[0072]

[0073] Step 6: Model evaluation and optimization: Use the test set to evaluate the trained YOLOv5 image recognition model and calculate the accuracy, precision, recall and other indicators of wading shore fishing behavior monitoring. The relevant indicator definitions and formulas for model evaluation are as follows:

[0074] (1) Accuracy: reflects the model's ability to correctly judge the overall sample. The calculation formula is:

[0075] Accuracy =(TP+TN) / (TP+TN+FP+FN) (3)

[0076] In the formula, TP (True Positive) is the number of samples correctly predicted as positive, TN (True Negative) is the number of samples correctly predicted as negative, FP (False Positive) is the number of samples incorrectly predicted as positive, and FN (False Negative) is the number of samples incorrectly predicted as negative.

[0077] (2) Precision: It measures how many samples predicted as positive are actually positive. Precision reflects the purity of the model in identifying positive samples. The larger the value, the better. The calculation formula is:

[0078] Precision = TP / (TP+FP) (4)

[0079] (3) Recall: It measures how many samples are correctly identified as positive among all samples that are actually positive. Recall is also called sensitivity or recall rate. It reflects the coverage of the model for positive samples. The larger the value, the better. The calculation formula is:

[0080] Recall = TP / (TP+FN) (5)

[0081] Through the calculation of these indicators, we can comprehensively evaluate the performance of the model in different situations and continuously optimize the subsequent iterations of the model. According to the evaluation results of the monitoring of fishing behavior near the shore, the model is optimized and adjusted accordingly, such as adjusting hyperparameters.

[0082] Step 7: Monitoring shore fishing behavior: Use the trained and optimized YOLOv5 image recognition model to monitor the newly acquired unknown images for shore fishing behavior. If the image contains shore fishing behavior, the model will identify the location and category of the shore fishing behavior in the image, draw a bounding box to mark the shore fishing behavior, and output the marked image.

[0083] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred arrangement scheme, a person skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. An intelligent monitoring method for wading shore fishing behavior based on an unmanned boat, characterized in that: The method comprises the following steps: Step 1: Image acquisition: Based on the high-definition imaging equipment carried by the unmanned boat, the image of the shore zone of the wading river and lake is acquired by controlling the angle and width of the equipment; Step 2: Image preprocessing: preprocess the acquired image, including adjusting the image size, normalizing the image pixel value, and enhancing the data; Step 3, image annotation: collect the preprocessed image data set, annotate each image, and the annotation content includes the location and category of wading shore fishing behavior, including the location and type of prohibited actions of personnel in the management area, to form an annotated image data set; Step 4: Build a YOLOv5 image recognition model: Build a YOLOv5 image recognition model based on YOLOv5; Step 5, model training: The preprocessed image dataset and the annotated image dataset are used as training data to train the YOLOv5 image recognition model. The model input is the preprocessed image dataset, and the output is the corresponding annotated image dataset. The training data is divided into a training set and a test set in proportion. Step 6: Model evaluation and optimization: Use the test set to evaluate the trained YOLOv5 image recognition model, calculate the accuracy, precision, and recall indicators of wading shore fishing behavior monitoring; and optimize and adjust the model based on the model evaluation results; Step 7: Monitoring fishing behavior by wading on the shore: Use the trained and optimized YOLOv5 image recognition model to monitor the newly acquired unknown images for fishing behavior by wading on the shore. If the image contains fishing behavior, the model will identify the location and category of the fishing behavior by wading on the shore in the image, draw a bounding box to mark the fishing behavior by wading on the shore, and output the marked image.

2. According to claim 1, a method for intelligent monitoring of shore fishing behavior based on an unmanned boat is characterized in that: The acquisition of images of the shoreline of the wading river and lake described in step 1 follows the following principles: (1) Number of images: During the operation of the unmanned ship, at least 300-500 images are collected at each monitoring point; (2) Image quality: The image resolution is no less than 1920x1080 pixels, and the images are collected under sunny, cloudy, and rainy weather conditions, and the image color depth reaches 24-bit true color; (3) Collection angle: Collect images from multiple angles, including the front and side, to obtain images with a comprehensive viewing angle range; (4) Lighting conditions: images are collected under different lighting conditions, including sunny days, cloudy days, heavy fog, and dusk; (5) Data format: The acquired images are saved in a lossless format to prevent information loss caused by data compression; the image metadata is fully preserved, including shooting time, GPS location, and device parameters.

3. The intelligent monitoring method for wading shore fishing behavior based on an unmanned boat according to claim 1 is characterized in that: The image resizing described in step 2 is specifically as follows: resizing all images, the size standard of the input model is 640x640 pixels, and when resizing, the bilinear interpolation method is used to maintain the clarity of the image, and the original aspect ratio is maintained by adding gray pixels to fill the edges of the image; The normalized image pixel value is specifically: calculating the mean and standard deviation of the data set collected by the unmanned ship, using these as standard parameters, and then subtracting the mean from each pixel value of the image and dividing it by the standard deviation; The enhanced data specifically includes: performing random rotation, translation, scaling, flipping and color dithering operations on the images collected by the unmanned ship to increase the diversity of the data.

4. The intelligent monitoring method for wading shore fishing behavior based on an unmanned boat according to claim 1 is characterized in that: The specific steps of marking the location of the wading shore fishing behavior in step 3 are: using a bounding box to mark the location of the wading shore fishing behavior, and converting the bounding box of the wading shore fishing behavior into a standard format.

5. The intelligent monitoring method for wading shore fishing behavior based on an unmanned boat according to claim 1 is characterized in that: The specific process of constructing the YOLOv5 image recognition model in step 4 is: using the YOLOv5 feature extraction network to obtain feature information of the target as an appearance descriptor; Feature extraction module of the YOLOv5 backbone network: The Conv convolution layer first extracts features, then sends them to the BN batch normalization layer for unified processing to speed up network learning, and then the SiLU activation function layer retains and maps the features and removes redundant features; Feature fusion module: The neck uses FPN and PAN structures to fuse features at multiple scales. The FPN structure samples from top to bottom, and the PAN structure samples from bottom to top, so that feature maps of different sizes contain strong semantic information and strong feature information, thereby enhancing the ability to extract target features of different sizes. After the main part and neck of the YOLOv5 detection network extract and fuse the features, three enhanced effective feature layers are obtained. The feature map sizes of these three enhanced features are (20, 20, 512), (40, 40, 256), and (80, 80, 128), respectively. At this time, the feature map is regarded as a set of feature points with channel number features. Considering that the targets to be tracked in the unmanned ship video sequence are small and dense, the enhanced feature map of size (80, 80, 128) is selected as the feature extraction result in the DeepSORT target tracking network. YOLOv5 is responsible for detecting objects in images or videos, while DeepSORT performs subsequent target tracking based on the detection results. The position of the object is predicted and matched by Kalman filtering and Hungarian algorithm. The feature map is followed by a fully connected layer and an L2 normalization layer to refit the appearance feature information of the target to obtain a 128-dimensional feature vector; then the obtained appearance feature vector is measured by cosine similarity and linearly weighted with the Mahalanobis distance metric, and the obtained cost formula is used as the metric of the Hungarian algorithm to obtain the best match: d^(*(Cos))(ij)=min{|1-r_j^Tr_k^((i))|r_k^((i))∈R_i^*} (1) c_(i,j)^*=λd^((Mah))(i,j)+(1-λ)d^(o(Cos))(ij) (2)where d^(*(Cos))(ij) represents the cosine distance metric taken as the appearance feature vector of R_i^*; c_(I,j)^* represents the cost formula of the Hungarian algorithm obtained by combining the cosine distance metric with the Mahalanobis distance metric; r_j represents the appearance descriptor corresponding to the j-th detection box d_j; r_k^((i)) represents the appearance descriptor corresponding to the ith trajectory; d^((Mah))(i,j) represents the Mahalanobis distance between the j-th detection box and the predicted bounding box of the ith trajectory.

6. The intelligent monitoring method for wading shore fishing behavior based on an unmanned boat according to claim 1 is characterized in that: The step 5 of dividing the training data into a training set and a test set in proportion is specifically: dividing the training data into a training set and a test set in a ratio of 3:

1.

7. The intelligent monitoring method for wading shore fishing behavior based on an unmanned boat according to claim 1 is characterized in that: The accuracy described in step 6 is: reflects the model's ability to correctly judge the overall sample, and the calculation formula is: Accuracy=(TP+TN) / (TP+TN+FP+FN) (3) Where Accuracy is the accuracy; TP is the number of samples correctly predicted as positive; TN is the number of samples correctly predicted as negative; FP is the number of samples incorrectly predicted as positive; FN is the number of samples incorrectly predicted as negative. The precision rate is: a measure of how many of the samples predicted as positive are truly positive, that is, it reflects the purity of the model in identifying positive samples. The larger the value, the better. The calculation formula is: Precision=TP / (TP+FP) (4) In the formula, Precision is the accuracy; The recall rate is: a measure of how many samples are correctly identified as positive among all samples that are actually positive, that is, it reflects the coverage of the model for positive samples. The larger the value, the better. The calculation formula is: Recall=TP / (TP+FN) (5) In the formula, Recall is the recall rate.

Images