Marine organism detection method and system based on two-stage multi-target tracking

Abstract

The invention relates to the technical field of marine organism detection, in particular to a marine organism detection method and system based on two-stage multi-target tracking. The method comprises the following steps: acquiring marine organism data sets including a category label data set and a target frame selection image data set; feature extraction is carried out on a marine organism data set based on an improved YOLOv8 model, the problem of information loss in the down-sampling process is solved by introducing an ADown module into the YOLOv8 model, calculation complexity is reduced by introducing a DRC2f mechanism into the YOLOv8 model, and interference of background information is reduced by introducing a CA attention mechanism into the YOLOv8 model; through combination of the YOLOv8 detector and the SMC-PHD tracker, the method can effectively improve the target detection precision in a complex underwater environment.

Description

Technical Field

[0001] The present invention relates to the technical field of marine life detection, and in particular to a marine life detection method and system based on two-stage multi-target tracking. Background Art

[0002] Marine biological target detection is an important application of computer vision and artificial intelligence technology in marine ecological monitoring, especially in marine biodiversity protection, fishery resource management, and marine environmental research. With the continuous development of deep learning, computer vision, and image processing technology, the research and application of marine biological target detection have made significant progress. Efficient multi-target tracking is the key to achieving green, accurate, large-scale aquaculture, marine exploration, and marine aquaculture, but it still faces some challenges. Traditional multi-target tracking methods based on Bayesian estimation usually rely on the probability model of the target and infer the state of each target through Bayesian estimation. However, such methods face some significant problems in practical applications. First, the detection probability is unknown, that is, it is impossible to determine whether the target exists in the observation data at certain moments, which makes the detection process uncertain, increases the possibility of false positives and false negatives, and affects the accuracy of the target state. Second, the target is born randomly, which means that new targets will appear randomly during the tracking process. Traditional methods usually rely on a preset number of targets or assume that the targets are known, which is very unsuitable for dynamically changing scenes. Finally, data association is complex. In multi-target tracking, the motion trajectories of different targets may intersect or occlude, making it difficult to accurately associate the observation data and the target state. Traditional Bayesian methods often cannot effectively handle such complex association problems, resulting in low tracking efficiency, especially when multiple targets are densely packed or severely occluded.

[0003] At present, there are mainly the following defects in the task of target detection for marine life detection: 1. The complexity of the underwater environment, especially in deep sea areas, where light can hardly penetrate, the image quality is extremely poor, and problems such as low contrast, blur, and noise appear, which brings great challenges to target detection. 2. There are many kinds of marine life with different forms, and the appearance of the same species will also change in different waters, seasons, or at different ages. For example, marine life such as small fish and whales have very different body shapes, colors, and behavioral characteristics, which makes the task of appearance-based target detection and classification more complicated. 3. The training of deep learning models depends on a large amount of high-quality labeled data, but the labeled data of marine life is generally insufficient. 4. Deep learning models usually require high computing resources and long inference time, which conflicts with real-time requirements. Therefore, how to improve the processing speed and efficiency of the algorithm while ensuring detection accuracy is a key issue. Summary of the invention

[0004] In order to solve the above-mentioned problems, the present invention provides a marine life detection method and system based on two-stage multi-target tracking.

[0005] In a first aspect, the present invention provides a method for detecting marine organisms based on two-stage multi-target tracking, which adopts the following technical solution: A marine life detection method based on two-stage multi-target tracking, comprising: Obtain marine biological datasets, including category label datasets and target box selection image datasets; The feature extraction of marine biological dataset is performed based on the improved YOLOv8 model. The Adown module is introduced into the YOLOv8 model to solve the information loss problem in the downsampling process, the DRC2f mechanism is introduced into the YOLOv8 model to reduce the computational complexity, and the CA attention mechanism is introduced into the YOLOv8 model to reduce the interference of background information. The extracted features are used for multi-target tracking using SMC-PHD filters; The data association process is optimized based on the Hungarian algorithm, and continuous tracking of multiple targets is achieved by matching the detection results in the dataset image with the target trajectory.

[0006] Furthermore, the problem of information loss in the downsampling process is solved by introducing the Adown module in the YOLOv8 model, including reducing the feature map size through average pooling operation and maintaining the smoothness of the image. Assume that the size of the input feature map is , after average pooling, the size of the output feature map becomes , in, is the size of the pooling window, is the pixel value in the pooling window. The Adown module divides the pooled feature map into two parts. The first part is Convolution extracts finer-grained features and reduces the dimension; the second part uses maximum pooling and Point-by-point convolution further enhances the extraction of nonlinear features. in, is the convolution kernel, is the input feature map, is the bias term. Finally, the two parts of the feature map after convolution are merged to form the final Adown module output.

[0007] Furthermore, the computational complexity is reduced by introducing the DRC2f mechanism in the YOLOv8 model, including fusing multi-scale information through convolution operations to extract features, which is expressed as: in, and are convolution kernels of different scales, is the input feature map. The DR module uses dilated convolution to expand the receptive field without increasing the amount of calculation. The operation of dilated convolution is expressed as: in, is the expansion factor, is the convolution kernel, Represents a larger range of pixels considered during the dilated convolution.

[0008] Furthermore, the CA attention mechanism is introduced into the YOLOv8 model to reduce the interference of background information, including assigning a weight coefficient to each channel based on the channel information of the input feature map, indicating the importance of the channel in the feature map, which is expressed as: in, is a weight matrix, is the input feature map, is the sigmoid function, is the activation function.

[0009] Furthermore, the SMC-PHD filter is used to perform multi-target tracking on the extracted features, including generating multiple particles to represent the position and velocity of the target. The weight of the particle is initialized to be associated with the confidence of the YOLOv8 output. The target particle with high confidence will obtain a large weight. At each time step, SMC-PHD predicts the state of the particle according to the target's motion model and updates the particle weight in combination with the latest YOLOv8 detection result. Suppose the output of YOLOv8 at a certain moment is ,in is the location of the target, is the detection confidence of the target, and the particle weight update process is expressed as: in, It is a particle At the moment The weight of It is a particle The detection probability of the corresponding position indicates the matching degree between the particle and the current detection result. Through the resampling step, the particles with higher weights are selected and the target state at the next moment is generated.

[0010] Furthermore, the method of performing multi-target tracking on the extracted features using the SMC-PHD filter also includes maintaining and updating the tracking status of multiple targets through particle filtering using the SMC-PHD filter, wherein the SMC-PHD filter estimates the number and status of targets based on the probability hypothesis density PHD.

[0011] Furthermore, the data association process is optimized based on the Hungarian algorithm, and continuous tracking of multiple targets is achieved by matching the detection results in the data set image with the target trajectory, including constructing a cost matrix, calculating the matching cost between different targets, and minimizing the cost through an optimization algorithm to achieve optimal matching. In multi-target tracking, the construction of the cost matrix is ​​based on a similarity measure between the target detection results and the tracking particle estimation.

[0012] In the second aspect, a marine life detection system based on two-stage multi-target tracking includes: The data acquisition module is configured to acquire a marine biological dataset, including a category label dataset and a target frame selection image dataset; The feature extraction module is configured to extract features from a marine biological data set based on an improved YOLOv8 model, wherein an Adown module is introduced into the YOLOv8 model to solve the information loss problem in the downsampling process, a DRC2f mechanism is introduced into the YOLOv8 model to reduce computational complexity, and a CA attention mechanism is introduced into the YOLOv8 model to reduce interference from background information; The tracking module is configured to perform multi-target tracking on the extracted features using the SMC-PHD filter; The multi-target module is configured to optimize the data association process based on the Hungarian algorithm and achieve continuous tracking of multiple targets by matching the detection results in the dataset images with the target trajectories.

[0013] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for example, a method for detecting marine life based on two-stage multi-target tracking.

[0014] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded by the processor and executing the method for detecting marine life based on two-stage multi-target tracking.

[0015] In summary, the present invention has the following beneficial technical effects: 1. Improve detection accuracy and robustness: By combining the YOLOv8 detector and the SMC-PHD tracker, this method can effectively improve the accuracy of target detection in complex underwater environments. The multiple improved modules of YOLOv8 improve the processing capabilities of low-contrast, blurred and noisy images, so that even in deep-sea environments with poor lighting, it can accurately identify and extract target information. The SMC-PHD tracker uses particle filtering technology to enhance the tracking capability of dynamic targets and reduce missed detections and false detections.

[0016] 2. Improve real-time performance and processing efficiency: In response to the real-time requirements of marine biological detection, this method introduces a lightweight Hungarian algorithm for data association, which effectively optimizes the use of computing resources and reduces the inference time caused by high computing requirements in traditional deep learning models. Through reasonable optimization, the algorithm can complete target detection and multi-target tracking in a shorter time while ensuring detection accuracy, providing technical support for real-time monitoring.

[0017] 3. Strong adaptability and support for multiple target types: Due to the wide variety of marine life and different appearance characteristics in different environments, this method can handle target types with high variability through the multi-target detection capabilities of YOLOv8 and the adaptability of the SMC-PHD tracker. Whether it is small fish or large whales, the method can accurately identify and track, has strong adaptability, and can cope with a variety of marine life target detection tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of a marine life detection method based on two-stage multi-target tracking according to Example 1 of the present invention. DETAILED DESCRIPTION

[0019] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0020] Example 1 Reference Figure 1 , a marine life detection method based on two-stage multi-target tracking of this embodiment includes: Obtain marine biological datasets, including category label datasets and target box selection image datasets; The feature extraction of marine biological dataset is performed based on the improved YOLOv8 model. The Adown module is introduced into the YOLOv8 model to solve the information loss problem in the downsampling process, the DRC2f mechanism is introduced into the YOLOv8 model to reduce the computational complexity, and the CA attention mechanism is introduced into the YOLOv8 model to reduce the interference of background information. The extracted features are used for multi-target tracking using SMC-PHD filters; The data association process is optimized based on the Hungarian algorithm, and continuous tracking of multiple targets is achieved by matching the detection results in the dataset image with the target trajectory.

[0021] Among them, include: S1: Data acquisition: Obtain marine ecological biological data sets, collect image information of different fish or marine plants at different stages, then perform feature calibration and classification annotation on the collected image information of fish and marine plants at different stages, and finally preprocess the data set.

[0022] S2: Multi-module improved YOLOv8: By introducing the Adown module, DRC2f module and CA attention mechanism into YOLOv8, the working efficiency and detection accuracy of the YOLOv8 model are improved. Then the processed marine ecological biological dataset is input into the multi-module improved YOLOv8 model to train the model and extract data features.

[0023] S3: SMC-PHD filter: The target detection results provided by YOLOv8 are used as the initialization particle information of the SMC-PHD filter, and then the SMC-PHD filter is used to further maintain and update the tracking status of multiple targets through particle filtering to achieve efficient multi-target tracking.

[0024] S4: Establish data correlation: Introduce the efficient Hungarian algorithm to optimize the data association process, accurately match the detection results and target trajectories in each frame, and achieve continuous tracking of multiple targets.

[0025] S5: Real-time detection of marine life: Use the trained detection model to perform real-time detection of different fish or marine plants to obtain information on the health status of marine life.

[0026] Preferably, a further technical solution of the present invention is: The marine life dataset obtained in the data acquisition of S1 includes a variety of marine life category label datasets and a marine life target frame selection image dataset. The data preprocessing includes random scaling, cropping, rotation and image enhancement operations on the data.

[0027] S2's multi-module improvement in YOLOv8 introduces the Adown module, DRC2f module, and CA attention mechanism in YOLOv8. The introduction of the Adown module is mainly to solve the problem of information loss that may be caused by the traditional downsampling process. Downsampling operations (such as pooling) will reduce the size of the feature map, resulting in the loss of some key information, especially in the expression of low-level features. The Adown module reduces the size of the feature map and maintains the smoothness of the image through the average pooling operation. The size of the input feature map is , after average pooling, the size of the output feature map becomes .

[0028] in, is the size of the pooling window, is the pixel value in the pooling window. The Adown module divides the pooled feature map into two parts. The first part is Convolution extracts finer-grained features and reduces the dimension; the second part uses maximum pooling and Point-wise convolution further enhances the extraction of nonlinear features.

[0029] in, is the convolution kernel, is the input feature map, is the bias term. Finally, the two parts of the feature map after convolution are merged to form the final Adown module output.

[0030] The design goal of the DRC2f module is to enhance feature extraction capabilities while maintaining lightweight computational characteristics. This module combines the C2f module in YOLOv8 and uses the Dilated Reparam Block (DR) to reduce computational complexity and expand the receptive field. C2f extracts features by fusing multi-scale information through convolution operations. Its operation can be expressed as: in, and are convolution kernels of different scales, is the input feature map. The DR module uses dilated convolution to expand the receptive field without increasing the amount of computation. The dilated convolution operation can be expressed as: in, is the expansion factor, is the convolution kernel, Represents a larger range of pixels considered during the dilated convolution process. The DRC2f module expands the receptive field by dilated convolution and reduces the dimension of the feature map by point-by-point convolution, thereby improving the efficiency and effect of feature extraction. DRC2f is implemented by the following operations: The DRC2f module enhances feature extraction capabilities, reduces computational complexity, and avoids computational redundancy, making it particularly suitable for use in resource-constrained environments.

[0031] The CA module (Channel Attention) is used to reduce the interference of background information and improve the detection accuracy of the target. It adaptively adjusts the importance of the feature map and suppresses the influence of background noise by considering both spatial information and channel information. Based on the channel information of the input feature map, the CA module assigns a weight coefficient to each channel to indicate the importance of the channel in the feature map. Its operation can be expressed as: in, is a weight matrix, is the input feature map, is the sigmoid function, The CA module enhances the representation capability of the feature map by integrating spatial position information into channel information, and adaptively adjusts the weights of important features in subsequent calculations, thereby reducing background interference and improving detection accuracy and robustness: in, represents element-by-element multiplication, and the final feature map adjusts the weight of each channel in the feature map according to the output of the attention mechanism. Through the CA module, the system can effectively ignore irrelevant background information, thereby improving the detection accuracy of targets in marine aquaculture environments.

[0032] These three improved modules (Adown module, DRC2f module and CA attention mechanism) solve the problems of different target sizes, strong background interference, and unpredictable target motion in marine aquaculture environments by combining pooling, convolution operations, dilated convolutions, and attention mechanisms.

[0033] In S3's SMC-PHD filter, the target detection result provided by YOLOv8 is used as the initialization particle information of the SMC-PHD filter. At the initial moment, YOLOv8 detects the target in the input image and gives the target's bounding box position and detection confidence. The detection box of each target output by YOLOv8 is ,in and are the coordinates of the target center, and is the width and height of the bounding box. Indicates the probability that the target is detected as this category. Based on these detection results, SMC-PHD generates multiple particles to represent the position and velocity of the target. The weight of the particle is initialized to be associated with the confidence of the YOLOv8 output. The target particle with a higher confidence will receive a larger weight. At each time step, SMC-PHD predicts the state of the particle based on the target's motion model and updates the particle weight based on the latest YOLOv8 detection result. Assume that the output of YOLOv8 at a certain moment is ,in is the target location, is the detection confidence of the target. The particle weight update process can be expressed as: in, It is a particle At the moment The weight of It is a particle The detection probability of the corresponding position indicates the matching degree between the particle and the current detection result. Finally, through the resampling step, the particles with higher weights are selected and the target state at the next moment is generated.

[0034] The SMC-PHD filter of S3 further uses particle filtering to maintain and update the tracking status of multiple targets. The SMC-PHD filter estimates the number and state of targets based on probability hypothesis density (PHD). In multi-target tracking, PHD not only considers the state of the target, but also models the number of targets.

[0035] The workflow of SMC-PHD includes the following steps: Generate particles for each target based on the detection results of YOLOv8. Each particle represents the state of a target (such as position, velocity, etc.) and is assigned an initial weight. Predict the state of each particle at the current moment based on the target's motion model (such as uniform motion, acceleration model, etc.). Use PHD to update the weights of the particles by combining new detection data. Here, the detection results of YOLOv8 provide the probability of the existence of the target and the location information of the target. Resample according to the weight of the particles to ensure that particles with higher weights are retained and particles with low weights are discarded. Finally, the state of the target is estimated by taking a weighted average of all particles.

[0036] In the second stage above, SMC-PHD realizes multi-target tracking by combining particle filtering technology with the detection results of YOLOv8. YOLOv8 provides the preliminary position and confidence of the target. SMC-PHD uses this information to initialize particles and estimates the state of the target through continuous updating of particle filtering and detection results. In this way, SMC-PHD can effectively handle the target state estimation problem in multi-target tracking, and gradually improve the tracking accuracy through data resampling and updating process.

[0037] The efficient Hungarian algorithm is introduced in S4 to optimize the data association process in the establishment of data correlation, accurately matching the detection results and target trajectories in each frame. The basic idea of ​​the Hungarian algorithm is to achieve optimal matching by constructing a cost matrix, calculating the matching cost between different targets, and minimizing the cost through an optimization algorithm. In multi-target tracking, the construction of the cost matrix is ​​based on the distance or similarity measurement between the target detection results and the tracking particle estimates.

[0038] Specifically, at time t, the target set detected by YOLOv8 is , where each test result Represents the bounding box coordinates of the object and confidence At the same time, the SMC-PHD tracker generates a set of particle collections , where each particle Represents the state of the target at the last moment. To achieve data association, it is necessary to first calculate the matching cost (usually distance or similarity) between the target detection result and the particle, and then use the Hungarian algorithm for optimal matching.

[0039] Cost Matrix is a The matrix of is the number of target detections (detection count for YOLOv8), is the size of the particle set (the number of particles in SMC-PHD). Each element in the matrix Indicates the test results and particles The common cost measurement method is based on the Euclidean distance of the location and the detection confidence. Specifically: in, is the Euclidean distance between the target detection position and the particle position, indicating the spatial proximity between the targets. and They are the confidence of the YOLOv8 detection results and the weight of the SMC-PHD particles. It represents the difference between the target detection confidence and the particle weight, and measures the reliability of the detection and the matching degree of the particles. is a weight coefficient used to balance the contribution of spatial distance and confidence.

[0040] The goal of the Hungarian algorithm is to find the optimal target and particle matching solution by minimizing the sum of the cost matrix. The specific process is as follows: Constructing the cost matrix: As mentioned above, construct the cost matrix based on the distance and confidence between the object detection and the particle.

[0041] Solving for the best match: Apply the Hungarian algorithm to find the match with the minimum cost in the cost matrix. The goal of the Hungarian algorithm is to match each target with a particle one-to-one so that the total cost is minimized. The steps of the algorithm are as follows: 1. Perform row and column subtraction on the matrix to ensure that the minimum element of each row and column is zero. 2. Mark the rows and columns to find the matching pair with the minimum cost. 3. Adjust the markings through the optimization step to make the minimum cost matching more effective.

[0042] Eventually, the Hungarian algorithm will return a match. ,in Indicates the test results and particles is matched, and This means there is no match. The optimal match found by the Hungarian algorithm can be used to associate the target detection results of YOLOv8 with the particles of the SMC-PHD tracker. The update of each particle will be affected by the new detection results, thereby estimating the state of the target more accurately. The update process of the particle at each time step can be expressed by the following formula: in, is the predicted motion model of the particle (e.g., a uniform velocity or acceleration model). is the noise term of the motion model, representing random motion changes. It is a correction term determined by the matching result of the Hungarian algorithm, which indicates the correction of the particle according to the detection result after update.

[0043] In this way, the detection results of YOLOv8 and the particle estimation of SMC-PHD are combined to generate an accurate trajectory for each target. Finally, the trajectories of multiple targets are gradually updated over time to form a complete multi-target tracking result.

[0044] In the real-time detection of marine life in S5, the trained detection model is used to perform real-time detection of different fish or marine plants to obtain the health status information of marine life. The model uses the trained weight parameter file to perform real-time target detection of different fish or marine plants at different growth stages, and reflects the status information of marine life through the real-time label of the model.

[0045] Example 2 This embodiment provides a marine life detection system based on two-stage multi-target tracking, including: The data acquisition module is configured to acquire a marine biological dataset, including a category label dataset and a target frame selection image dataset; The feature extraction module is configured to extract features from a marine biological data set based on an improved YOLOv8 model, wherein an Adown module is introduced into the YOLOv8 model to solve the information loss problem in the downsampling process, a DRC2f mechanism is introduced into the YOLOv8 model to reduce computational complexity, and a CA attention mechanism is introduced into the YOLOv8 model to reduce interference from background information; The tracking module is configured to perform multi-target tracking on the extracted features using the SMC-PHD filter; The multi-target module is configured to optimize the data association process based on the Hungarian algorithm and achieve continuous tracking of multiple targets by matching the detection results in the dataset images with the target trajectories.

[0046] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a method for detecting marine organisms based on two-stage multi-target tracking.

[0047] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor for a marine life detection method based on two-stage multi-target tracking.

[0048] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting marine organisms based on two-stage multi-target tracking, characterized in that: include: Obtain marine biological datasets, including category label datasets and target box selection image datasets; The feature extraction of marine biological dataset is performed based on the improved YOLOv8 model. The Adown module is introduced into the YOLOv8 model to solve the information loss problem in the downsampling process, the DRC2f mechanism is introduced into the YOLOv8 model to reduce the computational complexity, and the CA attention mechanism is introduced into the YOLOv8 model to reduce the interference of background information. The extracted features are used for multi-target tracking using SMC-PHD filters; The data association process is optimized based on the Hungarian algorithm, and continuous tracking of multiple targets is achieved by matching the detection results in the dataset image with the target trajectory.

2. A method for detecting marine organisms based on two-stage multi-target tracking according to claim 1, characterized in that: The problem of information loss in the downsampling process is solved by introducing the Adown module in the YOLOv8 model, including reducing the feature map size through average pooling operation and maintaining the smoothness of the image. Suppose the size of the input feature map is , after average pooling, the size of the output feature map becomes , in, is the size of the pooling window, is the pixel value in the pooling window. The Adown module divides the pooled feature map into two parts. The first part is Convolution extracts finer-grained features and reduces the dimension; the second part uses maximum pooling and Point-by-point convolution further enhances the extraction of nonlinear features. in, is the convolution kernel, is the input feature map, is the bias term. Finally, the two parts of the feature map after convolution are merged to form the final Adown module output.

3. A method for detecting marine organisms based on two-stage multi-target tracking according to claim 2, characterized in that: The computational complexity is reduced by introducing the DRC2f mechanism in the YOLOv8 model, including fusing multi-scale information through convolution operations to extract features, which is expressed as: in, and are convolution kernels of different scales, It is the input feature map. The DR module uses dilated convolution to expand the receptive field without increasing the amount of calculation. The operation of dilated convolution is expressed as: in, is the expansion factor, is the convolution kernel, Represents a larger range of pixels considered during the dilated convolution.

4. A method for detecting marine organisms based on two-stage multi-target tracking according to claim 3, characterized in that: The CA attention mechanism is introduced into the YOLOv8 model to reduce the interference of background information, including channel information based on the input feature map, and a weight coefficient is assigned to each channel to indicate the importance of the channel in the feature map, which is expressed as: in, is a weight matrix, is the input feature map, is the sigmoid function, is the activation function.

5. A method for detecting marine organisms based on two-stage multi-target tracking according to claim 4, characterized in that: The method uses the SMC-PHD filter to perform multi-target tracking on the extracted features, including generating multiple particles to represent the position and velocity of the target. The weight of the particle is initialized to be associated with the confidence of the YOLOv8 output. The target particle with high confidence will obtain a large weight. At each time step, SMC-PHD predicts the state of the particle according to the target's motion model and updates the particle weight in combination with the latest YOLOv8 detection result. Suppose the output of YOLOv8 at a certain moment is ,in is the target location, is the detection confidence of the target, and the particle weight update process is expressed as: in, It is a particle At the moment The weight of It is a particle The detection probability of the corresponding position indicates the matching degree between the particle and the current detection result. Through the resampling step, the particles with higher weights are selected and the target state at the next moment is generated.

6. A method for detecting marine organisms based on two-stage multi-target tracking according to claim 5, characterized in that: The method of using the SMC-PHD filter to perform multi-target tracking on the extracted features also includes using the SMC-PHD filter to maintain and update the tracking status of multiple targets through particle filtering, wherein the SMC-PHD filter estimates the number and status of the targets based on the probability hypothesis density PHD.

7. A method for detecting marine organisms based on two-stage multi-target tracking according to claim 6, characterized in that: The data association process is optimized based on the Hungarian algorithm, and continuous tracking of multiple targets is achieved by matching the detection results in the data set image with the target trajectory, including constructing a cost matrix, calculating the matching cost between different targets, and minimizing the cost through an optimization algorithm to achieve optimal matching. In multi-target tracking, the construction of the cost matrix is ​​based on a similarity measure between the target detection results and the tracking particle estimation.

8. A marine life detection system based on two-stage multi-target tracking, characterized in that: include: The data acquisition module is configured to acquire a marine biological dataset, including a category label dataset and a target frame selection image dataset; The feature extraction module is configured to extract features from a marine biological data set based on an improved YOLOv8 model, wherein an Adown module is introduced into the YOLOv8 model to solve the information loss problem in the downsampling process, a DRC2f mechanism is introduced into the YOLOv8 model to reduce computational complexity, and a CA attention mechanism is introduced into the YOLOv8 model to reduce interference from background information; The tracking module is configured to perform multi-target tracking on the extracted features using the SMC-PHD filter; The multi-target module is configured to optimize the data association process based on the Hungarian algorithm and achieve continuous tracking of multiple targets by matching the detection results in the dataset images with the target trajectories.

9. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the method according to claim 1 .

10. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is used to implement each instruction; and the computer-readable storage medium is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the method as claimed in claim 1 .

Images