A method for field road segmentation of agricultural machinery operation trajectory based on self-supervised representation model

Abstract

The present invention provides a method for segmenting farm machinery operation trajectories based on a self-supervised representation model, which belongs to the field of computer technology, and includes: inputting the operation trajectory data of the farm machinery into a trained trajectory segmentation model to obtain the field and road segmentation result corresponding to the operation trajectory data; the trajectory segmentation model is composed of a point-level encoder based on point density classification, a trajectory-level encoder based on trajectory similarity learning, and a bidirectional long short-term memory network layer, and is trained according to the following methods: based on the original trajectory data set, the trajectory representation model is pre-trained by self-supervised learning; based on the network weights of the pre-trained trajectory representation model, the labeled trajectory data set, and the cross entropy loss function, the trajectory segmentation model is further trained. The method can fully and effectively learn a wide range of driving modes, more accurately identify farmland trajectory points and road trajectory points in the trajectory sequence, and has high robustness and reliability in the face of dynamic environmental changes.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method for segmenting a field road of an agricultural machinery operation trajectory based on a self-supervised representation model. Background Art

[0002] During the operation of agricultural machinery, a complete agricultural machinery driving trajectory map can be drawn. These trajectory data provide important data support for analyzing the operation area and efficiency of agricultural machinery, which is crucial for agricultural machinery operation subsidies, social services and operation planning.

[0003] The first step in analyzing agricultural machinery operations is to accurately segment the trajectories of fields and roads, that is, to determine whether each trajectory point is located in the field or on the road. After determining the categories of all points in the trajectory data, the effective operation trajectory length of the agricultural machinery can be calculated, and then key indicators such as the operation area of ​​the agricultural machinery (i.e., the length of the track in the field multiplied by the operation width), operation time and efficiency can be analyzed. The quality of the trajectory segmentation algorithm directly affects the accuracy of subsequent analysis.

[0004] Progress has been made in this field in recent years, but the robustness of trajectory segmentation methods shows some variability when applied to trajectories of different qualities. Agricultural machinery operation trajectories vary in terms of Global Navigation Satellite System (GNSS) equipment accuracy and recording frequency. The source of this variability can be attributed to two main challenges: First, the scale of available training data is insufficient to represent the diverse driving behaviors of agricultural vehicles. Commonly used trajectory segmentation models are trained using supervised learning methods that rely on manually labeled data. However, due to the high cost of data labeling, the datasets used in previous studies are of relatively small data size, and the models cannot fully and effectively learn a wide range of driving patterns. Second, the structure of existing field-road segmentation models makes it difficult to accurately capture the complex dynamics, long-term spatial and temporal relationships of vehicle motion. Summary of the invention

[0005] The present invention provides a method for segmenting farmland and roads of agricultural machinery operation trajectories based on a self-supervised representation model, so as to solve the defects that the field and road segmentation model in the prior art cannot fully and effectively learn a wide range of driving modes and has difficulty in capturing the complex dynamic, long-term spatial and temporal relationships of vehicle motion, so as to achieve full and effective learning of a wide range of driving modes, more accurately identify farmland trajectory points and road trajectory points in trajectory sequences, and have high robustness and reliability in the face of dynamic environmental changes.

[0006] The present invention provides a method for segmenting a field road of an agricultural machinery operation trajectory based on a self-supervised representation model, comprising the following steps:

[0007] Inputting the operation trajectory data of the agricultural machinery into the trained trajectory segmentation model to obtain the field road segmentation result corresponding to the operation trajectory data;

[0008] The trajectory segmentation model consists of a point-level encoder based on point density classification, a trajectory-level encoder based on trajectory similarity learning, and a bidirectional long short-term memory network layer;

[0009] The trajectory segmentation model is trained according to the following method:

[0010] Based on the original trajectory data set, the trajectory representation model is pre-trained by self-supervised learning; the trajectory representation model consists of a point density classifier, a trajectory similarity learner and a joint model training module;

[0011] The trajectory segmentation model is further trained based on the network weights of the pre-trained trajectory representation model, the labeled trajectory dataset and the cross entropy loss function.

[0012] According to a method for segmenting agricultural machinery operation trajectories using a self-supervised representation model, the point-level encoder is composed of a convolutional neural network and a Transformer model; the trajectory-level encoder is composed of a skip-gram model and a long short-term memory layer.

[0013] According to a method for segmenting a field road of an agricultural machinery operation trajectory of a self-supervised representation model provided by the present invention, the trajectory representation model is pre-trained by self-supervised learning based on an original trajectory data set, specifically comprising:

[0014] Perform data cleaning on the original trajectory dataset;

[0015] Perform segment sampling on the original trajectory data set after data cleaning to obtain a sampled segment data set;

[0016] Based on the sampling segment dataset, the trajectory representation model is pre-trained in a self-supervised manner.

[0017] According to a method for segmenting a field road of an agricultural machinery operation trajectory of a self-supervised representation model provided by the present invention, the original trajectory data set after data cleaning is subjected to segment sampling to obtain a sampled segment data set, specifically comprising:

[0018] The similarity between any two trajectories in the original trajectory data set after data cleaning is calculated based on the dynamic time warping algorithm, and the trajectories are selected in order from low to high similarity to obtain a set of sampled trajectories;

[0019] Segmenting the trajectories in the sampled trajectory set according to fixed lengths to obtain a segmented segment set;

[0020] The similarity between any two trajectory segments in the segmented segment set is calculated based on a dynamic time warping algorithm, and trajectory segments are selected in order of similarity from low to high to obtain a sampling segment data set.

[0021] According to a method for segmenting a field road of an agricultural machinery operation trajectory of a self-supervised representation model provided by the present invention, the trajectory representation model is pre-trained in a self-supervised manner based on the sampling segment data set, specifically comprising:

[0022] Generating a pseudo label of the sampling segment data set based on the local density of the trajectory points in the sampling segment data set;

[0023] Based on a plurality of statistical motion parameters, encoding the sampling segment data set to obtain a sampling segment code;

[0024] Based on the pseudo labels of the sampling segment data set and the sampling segment encoding, the trajectory representation model is pre-trained in a self-supervised manner.

[0025] According to a method for segmenting agricultural machinery operation trajectory field roads of a self-supervised representation model provided by the present invention, data cleaning of the original trajectory data set is performed by one or more of the following methods:

[0026] Cleaning the resampled trajectory points in the original trajectory data set;

[0027] Cleaning the trajectory points belonging to the stationary trajectory in the original trajectory data set;

[0028] Cleaning the trajectory points of static drift of the original trajectory data set;

[0029] The trajectory points with abnormal longitude and latitude in the original trajectory data set are cleaned.

[0030] The present invention also provides a field road segmentation device for agricultural machinery operation trajectory based on a self-supervised representation model, comprising the following modules:

[0031] An acquisition module is used to input the operation trajectory data of the agricultural machinery into the trained trajectory segmentation model to obtain the field road segmentation result corresponding to the operation trajectory data;

[0032] The trajectory segmentation model consists of a point-level encoder based on point density classification, a trajectory-level encoder based on trajectory similarity learning, and a bidirectional long short-term memory network layer;

[0033] The trajectory segmentation model is trained according to the following method:

[0034] Based on the original trajectory data set, the trajectory representation model is pre-trained by self-supervised learning; the trajectory representation model consists of a point density classifier, a trajectory similarity learner and a joint model training module;

[0035] The trajectory segmentation model is further trained based on the network weights of the pre-trained trajectory representation model, the labeled trajectory dataset and the cross entropy loss function.

[0036] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a method for segmenting a field road of an agricultural machinery operation trajectory based on a self-supervised representation model as described in any one of the above is implemented.

[0037] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for segmenting a field road of an agricultural machinery operation trajectory based on a self-supervised representation model as described in any one of the above is implemented.

[0038] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-mentioned methods for segmenting agricultural machinery operation trajectories and fields based on the self-supervised representation model.

[0039] The present invention provides a method for segmenting field and road of agricultural machinery operation trajectory based on a self-supervised representation model. The trajectory segmentation model is composed of a point-level encoder based on point density classification, a trajectory-level encoder based on trajectory similarity learning, and a bidirectional long short-term memory network layer. The trajectory representation model is pre-trained in a self-supervised learning manner based on an original trajectory data set. The self-supervised learning method can overcome the limitations of small-scale manually labeled data sets. The trajectory representation model can learn the motion and spatial characteristics of agricultural machinery operation trajectories from a wider range of driving behaviors without the need for a large amount of labeled data. Then, the trajectory segmentation model can be fine-tuned based on the network weights of the pre-trained trajectory representation model, the labeled trajectory data set, and the cross-entropy loss function, so as to capture the complex dynamic, long-term spatial and temporal relationships of vehicle motion. The trained trajectory segmentation model can more accurately identify the farmland trajectory points and road trajectory points in the agricultural machinery operation trajectory sequence, and has high robustness and reliability in the face of dynamic environmental changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0041] Figure 1 A schematic flow chart of a method for segmenting a field road of an agricultural machinery operation trajectory based on a self-supervised representation model provided by the present invention.

[0042] Figure 2 A schematic diagram of the technical process of trajectory point classification provided by the present invention.

[0043] Figure 3 A schematic diagram of the architecture of the trajectory representation model provided by the present invention.

[0044] Figure 4 This is a schematic diagram for comparing the actual effects of the field road segmentation model provided by the present invention.

[0045] Figure 5 A schematic diagram of the structure of a field road segmentation device for agricultural machinery operation trajectories based on a self-supervised representation model provided by the present invention.

[0046] Figure 6 This is a schematic structural diagram of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] Figure 1 A schematic diagram of the process of the field road segmentation method of the agricultural machinery operation trajectory based on the self-supervised representation model provided by the present invention is shown in FIG. Figure 1 As shown, the method comprises the following steps:

[0049] Step 100: Input the operation trajectory data of the agricultural machinery into the trained trajectory segmentation model to obtain the field road segmentation result corresponding to the operation trajectory data.

[0050] The trajectory segmentation model consists of a point-level encoder based on point density classification, a trajectory-level encoder based on trajectory similarity learning, and a bidirectional long short-term memory network layer;

[0051] The trajectory segmentation model is trained as follows:

[0052] Based on the original trajectory dataset, the trajectory representation model is pre-trained by self-supervised learning; the trajectory representation model consists of a point density classifier, a trajectory similarity learner, and a joint model training module;

[0053] The trajectory segmentation model is further trained based on the network weights of the pre-trained trajectory representation model, the labeled trajectory dataset and the cross entropy loss function.

[0054] Specifically, in order to achieve accurate trajectory segmentation, the trajectory segmentation model used in the embodiment of the present invention uses a dual encoder structure: a point-level encoder based on point density classification and a trajectory-level encoder based on trajectory similarity learning. Among them, the point-level encoder focuses on extracting detailed features of a single point and performing point density classification, and the trajectory-level encoder enriches these features by calculating trajectory similarity and captures multi-level spatiotemporal features. The trajectory segmentation model also includes a bidirectional long short-term memory network (Bidirectional Long Short-Term Memory, BiLSTM) layer. The features obtained by the point-level encoder and the features obtained by the trajectory encoder are jointly input into the BiLSTM layer to obtain the final feature vector, and then the trajectory points are classified by the linear decoder.

[0055] In some implementations, the point-level encoder is composed of a convolutional neural network and a Transformer model; the trajectory-level encoder is composed of a skip-gram model and a long short-term memory layer.

[0056] Specifically, the point-level encoder in the trajectory segmentation model provided by the embodiment of the present invention may be composed of two model structures: a convolutional neural network (CNN) and a transformer.

[0057] Among them, CNN is used to learn the spatial pattern between adjacent trajectory points, and Transformer is used to model the temporal dependencies in short-term and long-term time periods. This dual-branch network architecture can effectively handle short-term and long-term spatiotemporal dependencies, helping the model to understand trajectory dynamics more comprehensively and significantly improving the effect of trajectory segmentation. Combining CNN and Transformer networks to process temporal and spatial dependencies in trajectory data can ensure that the model has high robustness and reliability in the face of dynamic environmental changes.

[0058] The trajectory-level encoder in the trajectory representation model provided by the embodiment of the present invention is composed of a skip-gram model and a long short-term memory (LSTM) layer. By adopting this trajectory-level representation architecture, the trajectory-level representation can be captured to provide background information for each trajectory point.

[0059] In the process of training the trajectory segmentation model, an embodiment of the present invention proposes a pre-training method for the trajectory segmentation model. First, based on the original trajectory dataset, the trajectory representation model is pre-trained by self-supervised learning.

[0060] Among them, the original trajectory dataset is data that has not been manually labeled.

[0061] Alternatively, the raw trajectory dataset can be obtained from the precision agriculture application project data service platform. For example, the raw trajectory dataset includes GNSS activity trajectories of agricultural vehicles in China, from which a sample of 50,000 raw trajectories is selected to form the raw trajectory dataset. Each trajectory point in the raw trajectory dataset contains 5 parameters recorded by the GNSS device: timestamp (YYYY:MM:DD-hh:mm:ss), geographic coordinates (longitude and latitude in WGS84 coordinate system), speed (m / s) and direction (°).

[0062] As mentioned above, the trajectory segmentation model in the embodiment of the present invention includes a point-level encoder based on point density classification and a trajectory-level encoder based on trajectory similarity learning. Accordingly, the pre-trained trajectory representation model is composed of a point density classifier, a trajectory similarity learner and a joint model training module.

[0063] Among them, the point density classifier is used to predict whether the neighborhood of a point has high point density. Specifically, for points with matrix input features, the point-level encoder extracts new feature representations, and then uses a linear layer through a linear decoder to perform point density prediction.

[0064] The trajectory similarity learner is used to analyze and compare time series or spatial trajectory data to determine the similarity between different trajectories. Specifically, a trajectory-level encoder is first used to obtain the trajectory-level feature representation of the input segment. Then, a similarity calculation loss function is used to calculate the similarity between pairs of segments based on the trajectory-level feature representation.

[0065] The joint model training module is used to improve model performance and generalization ability through multi-task learning. In order to consider the interaction between point-level feature representation and trajectory-level feature representation in the segment, the point density classifier and trajectory similarity learner are jointly learned through a self-supervised learning method. The loss function of the model training is The calculation is done by the following formula:

[0066]

[0067] in is the cross entropy loss for point density classification, is the loss of trajectory similarity calculation, and is the ratio of each loss function.

[0068] The limitation of small-scale manually labeled datasets can be overcome by using self-supervised learning methods. This method enables the trajectory representation model to learn the motion and spatial characteristics of agricultural machinery operation trajectories from a wider range of driving behaviors without the need for a large amount of labeled data.

[0069] After completing the pre-training process of the trajectory representation model, the network weights obtained by pre-training in the trajectory representation model can be used as the initialization weights of the trajectory segmentation model, and then the trajectory segmentation model can be further fine-tuned according to the labeled trajectory dataset and the cross entropy loss function.

[0070] Optionally, the labeled trajectory dataset may include two types, namely a high-quality trajectory dataset and a low-quality trajectory dataset. The quality of the trajectory is defined according to the acquisition frequency and the regularity of the trajectory shape. High-quality trajectories have a high sampling frequency and a more standardized shape, while low-quality trajectories have a low sampling acquisition frequency and a more disordered shape. For example, a high-quality trajectory dataset may include trajectory points recorded approximately every 5 seconds, and a low-quality trajectory dataset may include trajectory points recorded approximately every 30 seconds.

[0071] The encoder in the trajectory segmentation model uses the weights loaded from the trajectory representation model and the BiLSTM layer to extract feature representations, and the linear decoder classifies each point as a field or road based on these feature representations. The trajectory segmentation model is optimized using the cross entropy loss function to improve the model's performance on the trajectory segmentation task. This training process achieves the transition from self-supervised learning to supervised learning, effectively improving the accuracy and robustness of trajectory segmentation.

[0072] After the trajectory segmentation model training is completed, the operation trajectory data of the agricultural machinery can be input into the trained trajectory segmentation model to obtain the field road segmentation result corresponding to the operation trajectory data.

[0073] The present invention provides a method for segmenting field and road of agricultural machinery operation trajectory based on a self-supervised representation model. The trajectory segmentation model is composed of a point-level encoder based on point density classification, a trajectory-level encoder based on trajectory similarity learning, and a bidirectional long short-term memory network layer. The trajectory representation model is pre-trained in a self-supervised learning manner based on an original trajectory data set. The self-supervised learning method can overcome the limitations of small-scale manually labeled data sets. The trajectory representation model can learn the motion and spatial characteristics of agricultural machinery operation trajectories from a wider range of driving behaviors without the need for a large amount of labeled data. Then, the trajectory segmentation model can be fine-tuned based on the network weights of the pre-trained trajectory representation model, the labeled trajectory data set, and the cross-entropy loss function, so as to capture the complex dynamic, long-term spatial and temporal relationships of vehicle motion. The trained trajectory segmentation model can more accurately identify the farmland trajectory points and road trajectory points in the agricultural machinery operation trajectory sequence, and has high robustness and reliability in the face of dynamic environmental changes.

[0074] According to a method for segmenting agricultural machinery operation trajectory field road of a self-supervised representation model provided by the present invention, based on the original trajectory data set, the trajectory representation model is pre-trained by self-supervised learning, specifically comprising:

[0075] Perform data cleaning on the original trajectory dataset;

[0076] Perform segment sampling on the original trajectory data set after data cleaning to obtain a sampled segment data set;

[0077] Based on the sampling segment dataset, the trajectory representation model is pre-trained in a self-supervised manner.

[0078] Specifically, in the process of pre-training the trajectory representation model by self-supervised learning based on the original trajectory dataset, the original trajectory dataset can be firstly cleaned.

[0079] Since agricultural vehicles usually generate noise and repeated track points when they stop and wait during field work, the purpose of data cleaning the original track data set is to eliminate noise and repeated points in the track. Noise points refer to track points with deviated positioning, while repeated points are track points with the same longitude and latitude coordinates. Data cleaning of the original track data set can reduce the impact of noise and repeated points on the track segmentation task. In this process, noise points are corrected as much as possible and repeated points are eliminated.

[0080] Optionally, the original trajectory dataset is cleaned by one or more of the following methods:

[0081] (1) Clean the resampled trajectory points in the original trajectory data set. For example, clean the trajectory points with the same time in consecutive moments in the original trajectory data set and retain the first trajectory point in the abnormal trajectory segment.

[0082] (2) Clean the trajectory points that belong to the stationary trajectory in the original trajectory data set. For example, clean the trajectory points with the same longitude and latitude coordinates in consecutive moments in the original trajectory data set, and retain the first trajectory point in the abnormal trajectory segment.

[0083] (3) Clean the trajectory points with static drift in the original trajectory data set. For example, clean the trajectory points in the original trajectory data set whose longitude and latitude coordinates are different at consecutive moments and whose speed is 0, and retain the first trajectory point in the abnormal trajectory segment.

[0084] (4) Clean the trajectory points with abnormal longitude and latitude in the original trajectory dataset. For example, the data collection area of ​​the points in the original trajectory dataset covers many provinces and cities in China. Therefore, when the longitude and latitude range exceeds the scope of China, the collection point is abnormal in longitude and latitude and needs to be cleaned. The cleaning method is to delete it directly.

[0085] In order to meet the input requirements of some neural networks (such as LSTM, Transformer), the trajectory data needs to be divided into trajectory segments and then input into these networks. Therefore, it is necessary to perform segment sampling on the original trajectory dataset after data cleaning to obtain a sampled segment dataset. The process of segment sampling is to select some trajectory segments from the original trajectory dataset after data cleaning as training data and input them into the trajectory representation model for model training. It can be understood that a trajectory segment refers to a series of predetermined number of continuous points in the trajectory dataset.

[0086] According to a method for segmenting a field road of an agricultural machinery operation trajectory of a self-supervised representation model provided by the present invention, segment sampling is performed on the original trajectory data set after data cleaning to obtain a sampled segment data set, specifically including:

[0087] The similarity between any two trajectories in the original trajectory data set after data cleaning is calculated based on the dynamic time warping algorithm, and the trajectories are selected in order from low to high similarity to obtain a set of sampled trajectories;

[0088] The trajectories in the sampled trajectory set are segmented according to fixed lengths to obtain a segmented segment set;

[0089] The similarity between any two trajectory segments in the segment set is calculated based on the dynamic time warping algorithm, and the trajectory segments are selected in order from low to high similarity to obtain the sampling segment data set.

[0090] Specifically, in the embodiment of the present invention, the most dissimilar trajectory segments can be selected to obtain a sampling segment data set, so as to cover more driving modes. In this process, the sampling segment data set can be obtained through two steps of trajectory selection and trajectory segment selection according to the dynamic time warping (DTW) algorithm.

[0091] First, the similarity between any two trajectories in the original trajectory data set after data cleaning can be calculated according to the DTW algorithm, and the trajectories can be sorted from low to high according to the similarity to obtain a set of sampled trajectories. Then, the trajectories in the sampled trajectory set can be segmented according to a fixed length to obtain a set of segmented segments.

[0092] Similarly, the most dissimilar segment in the segment set can be selected according to the DTW algorithm, and finally a sampled segment dataset is obtained, which will be used as the input dataset for the pre-training stage of the trajectory representation model.

[0093] Based on the sampling segment dataset, the trajectory representation model is pre-trained in a self-supervised manner.

[0094] According to a method for segmenting agricultural machinery operation trajectory field road of a self-supervised representation model provided by the present invention, based on a sampling segment data set, a trajectory representation model is pre-trained in a self-supervised manner, specifically comprising:

[0095] Generate pseudo labels for the sampling segment data set based on the local density of trajectory points in the sampling segment data set;

[0096] Based on multiple statistical motion parameters, the sampling segment data set is encoded to obtain the sampling segment code;

[0097] Based on the pseudo labels of the sampling segment dataset and the sampling segment encoding, the trajectory representation model is pre-trained in a self-supervised manner.

[0098] Specifically, in the process of pre-training the trajectory representation model in a self-supervised manner based on the sample segment dataset, data preparation for self-supervised learning is first performed. This process is implemented through two main steps: pseudo label generation and input feature preparation.

[0099] Among them, the pseudo labels of the sampling segment dataset are automatically generated based on the local density of the trajectory points in the sampling segment dataset. Specifically, the Euclidean distance between adjacent points in the trajectory segment is first calculated, and then the local area is determined according to the median of the distance between the two points, and the number of points in the area is counted to determine the point density.

[0100] During the input feature preparation process, the sample segment data set is encoded according to multiple statistical motion parameters to obtain the sample segment encoding. For example, 25 statistical motion parameters can be used to encode the temporal dependency of each trajectory point, and the spatial dependency between each trajectory point is represented in matrix form to integrate the temporal and spatial information of the trajectory points.

[0101] After obtaining the pseudo labels and sampling segment codes of the sampling segment data set, the trajectory representation model can be pre-trained in a self-supervised manner based on the pseudo labels and sampling segment codes of the sampling segment data set.

[0102] The steps of pseudo-label generation and input feature preparation jointly prepare the necessary data for self-supervised learning, enabling the model to learn trajectory features from large-scale unlabeled trajectory data and improve robustness in trajectory segmentation tasks.

[0103] The following further illustrates the method for segmenting the agricultural machinery operation trajectory field road based on the self-supervised representation model provided by the present invention through embodiments in specific application scenarios.

[0104] Figure 2 A schematic diagram of the technical process of trajectory point classification provided by the present invention, such as Figure 2 As shown, in this embodiment, all track points in the agricultural machinery driving track can be automatically classified, and the farmland track points and road track points in the track sequence can be effectively identified.

[0105] This embodiment specifically includes the following steps:

[0106] Step 1: Data Acquisition

[0107] The experiment used in this embodiment utilizes two different types of data, namely, original trajectory data for pre-training the trajectory representation model and manually labeled data for fine-tuning the trajectory segmentation model.

[0108] (1) The original trajectory data is obtained from the precision agriculture application project data service platform. This dataset includes the GNSS activity trajectories of agricultural vehicles in China. This embodiment selects a sample of 50,000 original trajectories to form the dataset of this embodiment. Each trajectory point in the original trajectory contains five parameters recorded by the GNSS device: timestamp (YYYY:MM:DD-hh:mm:ss), geographic coordinates (longitude and latitude in the WGS84 coordinate system), speed (m / s) and direction (°).

[0109] (2) The manually labeled dataset contains two trajectory segmentation datasets, high-quality and low-quality, where the quality of the trajectory is defined based on the sampling frequency and the regularity of the trajectory shape. High-quality trajectories have high sampling frequencies and more standardized shapes, while low-quality trajectories have low sampling frequencies and more disordered shapes. Trajectories in the high-quality dataset are track points recorded approximately every 5 seconds, and trajectories in the low-quality dataset are track points recorded approximately every 30 seconds.

[0110] Step 2: Data preprocessing

[0111] 1. Data cleaning

[0112] In the process of collecting agricultural machinery driving trajectory data, GNSS positioning terminals often have sampling errors due to signal loss. In order to avoid the influence of sampling errors on subsequent classification results, it is necessary to clean the abnormal point types in the trajectory data in advance. Data cleaning includes the following types:

[0113] (1) Resampling type: clean the trajectory points with the same time in consecutive moments and retain the first trajectory point in the abnormal trajectory segment.

[0114] (2) Stationary trajectory type: Clean the trajectory points with the same longitude and latitude coordinates in consecutive moments and retain the first trajectory point in the abnormal trajectory segment.

[0115] (3) Static drift type: Clean the trajectory points with different longitude and latitude coordinates and 0 speed in consecutive moments, and retain the first trajectory point in the abnormal trajectory segment.

[0116] (4) Longitude and latitude anomaly type: Since the data collection area covers many provinces and cities in China, when the longitude and latitude range exceeds the scope of China, the collection point is a longitude and latitude anomaly and needs to be cleaned. The cleaning method is direct deletion.

[0117] 2. Trajectory Segment Sampling

[0118] This step involves selecting some segments from the original trajectory dataset as training data and inputting them into the trajectory representation model for model training. A trajectory segment refers to a series of a predetermined number of consecutive points in the trajectory. In order to meet the input requirements of some neural networks (such as LSTM, Transformer), the trajectory data needs to be divided into trajectory segments and then input into these networks. In addition, since the sampled segments are the most dissimilar trajectory segments, the trajectory segment dataset obtained after segment sampling can cover more driving modes. Segment sampling includes two steps:

[0119] Trajectory Selection: Select dissimilar trajectories from the original trajectory dataset.

[0120] Trajectory segment selection: Select dissimilar segments from the trajectories selected in the trajectory selection step.

[0121] The DTW algorithm is used to calculate the similarity between any two trajectories in the original trajectory dataset, and the most dissimilar trajectory is selected accordingly. Then, the selected trajectory is segmented according to a fixed length, and the algorithm selects the most dissimilar segments. The final sampled segments will be used as the input dataset for the pre-training stage of the trajectory representation model.

[0122] Step 3: Data preparation for self-supervised learning

[0123] Data preparation for self-supervised learning is achieved through two main steps: pseudo-label generation and input feature preparation. Pseudo-labels are automatically generated based on the local density of trajectory points. First, the Euclidean distance between adjacent points in the trajectory segment is calculated, and then the local area is determined according to the median of the distance between the two points. The number of points in the area is counted to determine the density of the points. Input feature preparation involves using 25 statistical motion parameters to encode the temporal dependencies of each trajectory point, and representing the spatial dependencies between points in matrix form, integrating the spatiotemporal information of the points. In addition, the segment sampling process includes selecting dissimilar trajectories and segments in the original trajectory dataset using the DTW algorithm to capture a wider range of driving patterns. Together, these steps prepare the necessary data for self-supervised learning, enabling the model to learn trajectory features from large-scale unlabeled trajectory data and improve robustness in trajectory segmentation tasks.

[0124] Step 4: Pre-train trajectory representation model

[0125] The trajectory representation model mainly consists of three components: point density classifier, trajectory similarity learner and joint model training module.

[0126] Point Density Classifier: The point density classifier predicts whether the neighborhood of a point has high point density. Specifically, for a point with matrix input features, the point-level encoder extracts a new feature representation, and then passes through a linear decoder to make point density prediction using a linear layer. Figure 3 A schematic diagram of the architecture of the trajectory representation model provided by the present invention, such as Figure 3 As shown in Figure 1, there are two networks in the point-level encoder: CNN and Transformer, where CNN is used to detect the spatial dependencies of trajectory points, and Transformer is used to capture the temporal dependencies of trajectory points.

[0127] Trajectory similarity learner: Trajectory similarity calculation is a traditional task in the field of urban transportation. Usually, the similarity between trajectories is measured by learning trajectory level feature representation. Figure 3As shown in Figure 1. First, a trajectory-level encoder (including a skip-gram model and an LSTM layer) is used to obtain the trajectory-level feature representation of the input segment. Then, the similarity between pairs of segments is calculated based on the trajectory-level feature representation using a similarity calculation loss function.

[0128] Joint model training module: In order to consider the interaction between point-level feature representation and trajectory-level feature representation in segments, these two components are jointly learned through a self-supervised learning method. Model training loss function As shown below, is the cross entropy loss for point density classification, is the loss of trajectory similarity calculation, and is the ratio of each loss function.

[0129]

[0130] Step 5: Trajectory segmentation model training

[0131] In this embodiment, the training process of the trajectory segmentation model is divided into two stages: the first is the self-supervised pre-training stage, and the second is the fine-tuning stage. In the self-supervised pre-training stage, a large-scale unlabeled original trajectory dataset is used to train the trajectory representation (PreTR) model, which consists of a point-level encoder and a trajectory-level encoder. The multi-level spatiotemporal features of the trajectory points are extracted through self-supervised learning tasks (point density classification and trajectory similarity learning). In the fine-tuning stage, the weights of the encoder in the PreTR model are used as the initialization weights of the trajectory segmentation (TS) model using manually labeled high-quality and low-quality trajectory datasets, and then the trajectory segmentation (TS) model is further trained. The TS model consists of three main parts: trajectory segment sampling, encoder, and linear decoder. The trajectory segment sampling is responsible for cutting the trajectory into a fixed-length sequence. The encoder uses the weights loaded from the PreTR model and the BiLSTM layer to extract feature representations. The linear decoder classifies each point as a field or road based on these feature representations. The TS model is optimized by the cross entropy loss function to improve the performance of the model on the trajectory segmentation task. This training process realizes the transition from self-supervised learning to supervised learning, effectively improving the accuracy and robustness of trajectory segmentation.

[0132] Step 6: Experimental results analysis

[0133] In the experimental analysis step, the purpose is to comprehensively evaluate the PreTR-TS model proposed in this embodiment and compare it with the prior art. To this end, this embodiment selects five trajectory segmentation methods as baselines, including Bagging-SVM (Bagging-Support Vector Machine), Density-Based Spatial Clustering of Applications with Noise+Rules (DBSCAN+Rules) and Decision Trees (DT) based on traditional machine learning technology, as well as Graph Convolutional Networks (GCN) and Multi-View Fusion (MVF) based on deep learning technology to ensure that the comparison of this embodiment covers a wide range of technologies. Two manually labeled data sets, high-quality and low-quality trajectory data sets, are used to simulate scenarios with different GNSS equipment accuracy and recording frequency. The selection of these data sets ensures that the evaluation of this embodiment can truly reflect the performance of the model under different data qualities. In order to systematically measure the performance of the model, this embodiment selects precision (Pre), recall (Rec), F1 score (F1) and accuracy (Acc) as the main performance evaluation indicators. These indicators can comprehensively evaluate the classification accuracy and comprehensiveness of the model.

[0134] First, the performance of the trajectory characterization model is evaluated using various performance evaluation indicators. Table 1 is a schematic diagram of the performance of the trajectory characterization model. In the trajectory characterization model, this embodiment trains a point density classifier in the trajectory characterization model. As shown in Table 1, the F1 score of the point density classifier on the validation set of the original trajectory dataset is 92.25%, and the accuracy is 96.86%.

[0135] Table 1 Performance of trajectory representation model

[0136]

[0137] For the trajectory similarity analysis module, the performance of the trajectory similarity learner was evaluated in the trajectory representation model. Similar to the training process of the point density classifier, this embodiment trained the trajectory similarity analysis model in the original trajectory dataset. As shown in Table 1, the F1 score and accuracy of the trajectory similarity analysis module were 90.34% and 91.37%, respectively.

[0138] Table 2 shows the performance of different trajectory segmentation methods on two datasets. As shown in Table 2, the PreTR-TS model proposed by this method is better than the baseline method.

[0139] Table 2 Performance of different trajectory segmentation methods on two datasets

[0140]

[0141] As shown in Table 2, on high-quality datasets, the PreTR-TS model achieved an F1 score of 86.14% and an accuracy of 93.95%; on low-quality datasets, the F1 score was 83.87% and the accuracy was 89.32%. Compared with the comparative method, the model performance of this embodiment on high-quality datasets is significantly improved. In terms of F1 score indicators, it is 17.75% higher than DB+Rules, 19.35% higher than DT, 13.38% higher than Bagging-SVM, 11% higher than GCN, and 3.02% higher than MVF. On low-quality datasets, the performance improvements of the F1 score indicators are 17.37%, 15.58%, 9.71%, 3.55% and 4.26% respectively. The results show that the model shows good performance on datasets of different qualities. Figure 4 A schematic diagram showing the actual effect comparison of the field road segmentation model provided by the present invention. Figure 4 As shown, through case studies, this embodiment further confirms the effectiveness of PreTR-TS in processing different types of trajectory data.

[0142] In addition, this method also explores the impact of training data scale on model performance. Table 3 shows the performance of PreTR-TS on two datasets under different training data scales. As the training data scale of the pre-trained trajectory representation model increases, the performance of PreTR-TS in field and road segmentation is also improved. The specific experimental results are shown in Table 3.

[0143] Table 3 Performance of PreTR-TS on two datasets with different training data scales

[0144]

[0145] Taking these findings into account, the PreTR-TS method performs well in the agricultural vehicle trajectory field road segmentation task. It is not only technically superior but also shows stability and reliability under different data quality conditions.

[0146] The following is a description of the agricultural machinery operation trajectory field road segmentation device based on the self-supervised representation model provided by the present invention. The agricultural machinery operation trajectory field road segmentation device based on the self-supervised representation model described below and the agricultural machinery operation trajectory field road segmentation method based on the self-supervised representation model described above can be referenced to each other.

[0147] Figure 5The schematic diagram of the structure of the field road segmentation device for agricultural machinery operation trajectory based on the self-supervised representation model provided by the present invention is as follows: Figure 5 As shown, the device includes the following modules:

[0148] An acquisition module 500 is used to input the operation track data of the agricultural machinery into the trained track segmentation model to obtain the field road segmentation result corresponding to the operation track data;

[0149] The trajectory segmentation model consists of a point-level encoder based on point density classification, a trajectory-level encoder based on trajectory similarity learning, and a bidirectional long short-term memory network layer;

[0150] The trajectory segmentation model is trained as follows:

[0151] Based on the original trajectory dataset, the trajectory representation model is pre-trained by self-supervised learning; the trajectory representation model consists of a point density classifier, a trajectory similarity learner, and a joint model training module;

[0152] The trajectory segmentation model is further trained based on the network weights of the pre-trained trajectory representation model, the labeled trajectory dataset and the cross entropy loss function.

[0153] According to a self-supervised representation model for agricultural machinery operation trajectory field road segmentation device provided by the present invention, the point-level encoder is composed of a convolutional neural network and a Transformer; the trajectory-level encoder is composed of a skip-gram model and a long short-term memory layer.

[0154] According to a self-supervised representation model for agricultural machinery operation trajectory field road segmentation provided by the present invention, based on the original trajectory data set, the trajectory representation model is pre-trained by self-supervised learning, specifically including:

[0155] Perform data cleaning on the original trajectory dataset;

[0156] Perform segment sampling on the original trajectory data set after data cleaning to obtain a sampled segment data set;

[0157] Based on the sampling segment dataset, the trajectory representation model is pre-trained in a self-supervised manner.

[0158] According to a self-supervised representation model of an agricultural machinery operation track field road segmentation device provided by the present invention, segment sampling is performed on the original track data set after data cleaning to obtain a sampled segment data set, specifically including:

[0159] The similarity between any two trajectories in the original trajectory data set after data cleaning is calculated based on the dynamic time warping algorithm, and the trajectories are selected in order from low to high similarity to obtain a set of sampled trajectories;

[0160] The trajectories in the sampled trajectory set are segmented according to fixed lengths to obtain a segmented segment set;

[0161] The similarity between any two trajectory segments in the segment set is calculated based on the dynamic time warping algorithm, and the trajectory segments are selected in order from low to high similarity to obtain the sampling segment data set.

[0162] According to a self-supervised representation model for agricultural machinery operation trajectory field road segmentation provided by the present invention, based on a sampling segment data set, a trajectory representation model is pre-trained in a self-supervised manner, specifically comprising:

[0163] Generate pseudo labels for the sampling segment data set based on the local density of trajectory points in the sampling segment data set;

[0164] Based on multiple statistical motion parameters, the sampling segment data set is encoded to obtain the sampling segment code;

[0165] Based on the pseudo labels of the sampling segment dataset and the sampling segment encoding, the trajectory representation model is pre-trained in a self-supervised manner.

[0166] According to a self-supervised representation model of an agricultural machinery operation track field road segmentation device provided by the present invention, data cleaning is performed on the original track data set in one or more of the following ways:

[0167] Clean the resampled trajectory points in the original trajectory dataset;

[0168] Clean the trajectory points belonging to the stationary trajectory in the original trajectory data set;

[0169] Clean the trajectory points of static drift in the original trajectory data set;

[0170] Clean the trajectory points with abnormal longitude and latitude in the original trajectory dataset.

[0171] Figure 6 A schematic diagram of the structure of an electronic device provided by the present invention, such as Figure 6 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630 and a communication bus 640, wherein the processor 610, the communication interface 620 and the memory 630 communicate with each other through the communication bus 640. The processor 610 may call the logic instructions in the memory 630 to execute the field road segmentation method of the agricultural machinery operation trajectory based on the self-supervised representation model, and the method includes:

[0172] Input the operation trajectory data of the agricultural machinery into the trained trajectory segmentation model to obtain the field road segmentation result corresponding to the operation trajectory data;

[0173] The trajectory segmentation model consists of a point-level encoder based on point density classification, a trajectory-level encoder based on trajectory similarity learning, and a bidirectional long short-term memory network layer;

[0174] The trajectory segmentation model is trained as follows:

[0175] Based on the original trajectory dataset, the trajectory representation model is pre-trained by self-supervised learning; the trajectory representation model consists of a point density classifier, a trajectory similarity learner, and a joint model training module;

[0176] The trajectory segmentation model is further trained based on the network weights of the pre-trained trajectory representation model, the labeled trajectory dataset and the cross entropy loss function.

[0177] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0178] On the other hand, the present invention further provides a computer program product, the computer program product includes a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, when the computer program is executed by a processor, the computer can execute the agricultural machinery operation trajectory field road segmentation method based on the self-supervised representation model provided by the above methods, the method includes:

[0179] Input the operation trajectory data of the agricultural machinery into the trained trajectory segmentation model to obtain the field road segmentation result corresponding to the operation trajectory data;

[0180] The trajectory segmentation model consists of a point-level encoder based on point density classification, a trajectory-level encoder based on trajectory similarity learning, and a bidirectional long short-term memory network layer;

[0181] The trajectory segmentation model is trained as follows:

[0182] Based on the original trajectory dataset, the trajectory representation model is pre-trained by self-supervised learning; the trajectory representation model consists of a point density classifier, a trajectory similarity learner, and a joint model training module;

[0183] The trajectory segmentation model is further trained based on the network weights of the pre-trained trajectory representation model, the labeled trajectory dataset and the cross entropy loss function.

[0184] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to execute the agricultural machinery operation trajectory field road segmentation method based on the self-supervised representation model provided by the above methods, the method comprising:

[0185] Input the operation trajectory data of the agricultural machinery into the trained trajectory segmentation model to obtain the field road segmentation result corresponding to the operation trajectory data;

[0186] The trajectory segmentation model consists of a point-level encoder based on point density classification, a trajectory-level encoder based on trajectory similarity learning, and a bidirectional long short-term memory network layer;

[0187] The trajectory segmentation model is trained as follows:

[0188] Based on the original trajectory dataset, the trajectory representation model is pre-trained by self-supervised learning; the trajectory representation model consists of a point density classifier, a trajectory similarity learner, and a joint model training module;

[0189] The trajectory segmentation model is further trained based on the network weights of the pre-trained trajectory representation model, the labeled trajectory dataset and the cross entropy loss function.

[0190] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0191] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for segmenting agricultural machinery operation trajectories based on a self-supervised representation model, characterized in that: include: Inputting the operation trajectory data of the agricultural machinery into the trained trajectory segmentation model to obtain the field road segmentation result corresponding to the operation trajectory data; The trajectory segmentation model consists of a point-level encoder based on point density classification, a trajectory-level encoder based on trajectory similarity learning, and a bidirectional long short-term memory network layer; The trajectory segmentation model is trained according to the following method: Based on the original trajectory data set, the trajectory representation model is pre-trained by self-supervised learning; the trajectory representation model consists of a point density classifier, a trajectory similarity learner and a joint model training module; Further training the trajectory segmentation model based on the network weights of the pre-trained trajectory representation model, the labeled trajectory dataset, and the cross entropy loss function; The point-level encoder is composed of a convolutional neural network and a Transformer model; the trajectory-level encoder is composed of a skip-gram model and a long short-term memory layer; The point density classifier is used to predict whether the neighborhood of a point has a high point density. For points with matrix input features, the point density classifier extracts new feature representations through a point-level encoder, and then uses a linear layer through a linear decoder to perform point density prediction; the trajectory similarity learner is used to analyze and compare time series or spatial trajectory data to determine the similarity between different trajectories. The trajectory similarity learner first uses a trajectory-level encoder to obtain the trajectory-level feature representation of the input segment, and then uses a similarity calculation loss function to calculate the similarity between paired segments based on the trajectory-level feature representation.

2. The method for segmenting agricultural machinery operation trajectories based on a self-supervised representation model according to claim 1, characterized in that: The method of pre-training the trajectory representation model based on the original trajectory dataset by self-supervised learning specifically includes: Perform data cleaning on the original trajectory dataset; Perform segment sampling on the original trajectory data set after data cleaning to obtain a sampled segment data set; Based on the sampling segment dataset, the trajectory representation model is pre-trained in a self-supervised manner.

3. The method for segmenting agricultural machinery operation trajectories based on a self-supervised representation model according to claim 2, characterized in that: The segment sampling of the original trajectory data set after data cleaning to obtain the sampled segment data set specifically includes: Based on the algorithm, the similarity between any two trajectories in the original trajectory data set after data cleaning is calculated, and the trajectories are selected in order from low to high similarity to obtain a set of sampled trajectories; Segmenting the trajectories in the sampled trajectory set according to fixed lengths to obtain a segmented segment set; The similarity between any two trajectory segments in the segmented segment set is calculated based on a dynamic time warping algorithm, and trajectory segments are selected in order of similarity from low to high to obtain a sampling segment data set.

4. The method for segmenting agricultural machinery operation trajectories based on a self-supervised representation model according to claim 2, characterized in that: The pre-training of the trajectory representation model based on the sampling segment data set in a self-supervised manner specifically includes: Generating a pseudo label of the sampling segment data set based on the local density of the trajectory points in the sampling segment data set; Based on a plurality of statistical motion parameters, encoding the sampling segment data set to obtain a sampling segment code; Based on the pseudo labels of the sampling segment data set and the sampling segment encoding, the trajectory representation model is pre-trained in a self-supervised manner.

5. The method for segmenting agricultural machinery operation trajectories based on a self-supervised representation model according to claim 2, characterized in that: The data cleaning of the original trajectory data set is implemented by one or more of the following methods: Cleaning the resampled trajectory points in the original trajectory data set; Cleaning the trajectory points belonging to the stationary trajectory in the original trajectory data set; Cleaning the trajectory points of static drift of the original trajectory data set; The trajectory points with abnormal longitude and latitude in the original trajectory data set are cleaned.

6. A field road segmentation device for agricultural machinery operation trajectory based on a self-supervised representation model, characterized in that: include: An acquisition module is used to input the operation trajectory data of the agricultural machinery into the trained trajectory segmentation model to obtain the field road segmentation result corresponding to the operation trajectory data; The trajectory segmentation model consists of a point-level encoder based on point density classification, a trajectory-level encoder based on trajectory similarity learning, and a bidirectional long short-term memory network layer; The trajectory segmentation model is trained according to the following method: Based on the original trajectory data set, the trajectory representation model is pre-trained by self-supervised learning; the trajectory representation model consists of a point density classifier, a trajectory similarity learner and a joint model training module; Further training the trajectory segmentation model based on the network weights of the pre-trained trajectory representation model, the labeled trajectory dataset, and the cross entropy loss function; The point-level encoder is composed of a convolutional neural network and a Transformer model; the trajectory-level encoder is composed of a skip-gram model and a long short-term memory layer; The point density classifier is used to predict whether the neighborhood of a point has a high point density. For points with matrix input features, the point density classifier extracts new feature representations through a point-level encoder, and then uses a linear layer through a linear decoder to perform point density prediction; the trajectory similarity learner is used to analyze and compare time series or spatial trajectory data to determine the similarity between different trajectories. The trajectory similarity learner first uses a trajectory-level encoder to obtain the trajectory-level feature representation of the input segment, and then uses a similarity calculation loss function to calculate the similarity between paired segments based on the trajectory-level feature representation.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for segmenting the field road of the agricultural machinery operation trajectory based on the self-supervised representation model as described in any one of claims 1 to 5 is implemented.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for segmenting the field road of the agricultural machinery operation trajectory based on the self-supervised representation model as described in any one of claims 1 to 5 is implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for segmenting the field road of the agricultural machinery operation trajectory based on the self-supervised representation model as described in any one of claims 1 to 5 is implemented.

Images